1013-00005 · Genome Dx Therapeutics Inc · Validity

prosecution-history-estoppel-memo.docx DOCUMENT Litigation

35 / 35 ← ✕

PRIVILEGED & CONFIDENTIAL — ATTORNEY WORK PRODUCT

MEMORANDUM

TO: Catherine Bloom, Partner, intellectual Property Practice Group, Calderwood & Harkness LLP

FROM: Wendy Schultz, Associate, intellectual Property Practice Group, Calderwood & Harkness LLP

DATE: February 19, 2026

RE: Prosecution History Estoppel Analysis — Impact on Claim Scope and Doctrine of Equivalents; Genome Dx Therapeutics Inc. v. Helios Precision Diagnostics, Inc., Case No. 3:25-cv-01847-JWL (N.D. Cal.)

MATTER NO.: 1013-00005

PRIVILEGED & CONFIDENTIAL — ATTORNEY-CLIENT PRIVILEGE / ATTORNEY WORK PRODUCT — PREPARED IN ANTICIPATION OF LITIGATION. This memorandum was prepared by counsel for Genome Dx Therapeutics Inc. in anticipation of litigation and is protected from disclosure by the attorney-client privilege and the work product doctrine. Distribution of this memorandum, in whole or in part, to any person outside the attorney-client relationship may result in waiver of applicable privileges.

Right-click to update Table of Contents

I. Executive Summary

This memorandum assesses whether amendments and arguments made during the prosecution of the three patents-in-suit — U.S. Patent Nos. 10,438,271 (the "'271 Patent"), 11,204,893 (the "'893 Patent"), and 11,891,647 (the "'647 Patent") — create prosecution history estoppel ("PHE") that would limit our client Genome Dx Therapeutics Inc.'s ("GenomeDx") ability to assert infringement under the doctrine of equivalents ("DOE") against Helios Precision Diagnostics, Inc.'s ("Helios") ProstaSignal™ Genomic Classifier product. This analysis is undertaken in connection with the pending litigation captioned above, currently before the Honorable James W. Larkin in the United States District Court for the Northern District of California.

After a detailed review of the complete prosecution file wrappers for each of the three patents-in-suit, the relevant case law, and the technical characteristics of the accused ProstaSignal™ product, we reach the following principal conclusions:

PHE poses a moderate-to-high risk to GenomeDx's DOE theories with respect to the '893 Patent's "random-forest classifier" limitation. This is the most critical DOE argument in the case, because Helios's ProstaSignal™ product uses a gradient-boosted decision tree classifier rather than a random-forest classifier as recited in the '893 Patent claims. During prosecution of the '893 Patent, GenomeDx's patent counsel made affirmative arguments distinguishing the claimed "random-forest classifier" from the support vector machine ("SVM") classifier disclosed in the Nakamura prior art reference. Helios will almost certainly invoke these arguments to contend that GenomeDx disclaimed all non-random-forest machine-learning classifiers, including gradient-boosted decision trees. We assess this argument-based estoppel theory as a genuine litigation risk, although GenomeDx has viable rebuttal arguments — particularly the "tangential relation" prong of the Festo framework.

PHE poses a lower risk for the '271 Patent's amended limitations. The amendments made to the '271 Patent during prosecution were driven by a Section 101 patent-eligibility rejection and added specific laboratory-technique limitations (FFPE sample extraction and reference-gene normalization). However, Helios's ProstaSignal™ product literally satisfies these amended limitations, meaning GenomeDx does not need to invoke DOE for these claim elements. The critical "trained machine-learning classifier" term in the '271 Patent was never amended during prosecution and remains broad enough to encompass gradient-boosted decision tree classifiers on a literal-infringement theory.

PHE risk is minimal for the '647 Patent. No narrowing amendments were made during prosecution of the '647 Patent; the applicant overcame the examiner's obviousness rejection through an evidentiary showing (the Mehta declaration) rather than through claim narrowing. Moreover, GenomeDx alleges literal infringement of the '647 Patent's combined biomarker/clinical variable claims, eliminating the need to rely on DOE.

Strategic Recommendation. GenomeDx's strongest infringement theories are: (1) literal infringement under the '271 Patent, where the broad "trained machine-learning classifier" language was never narrowed during prosecution and comfortably encompasses Helios's gradient-boosted decision tree; and (2) literal infringement under the '647 Patent, where the combined biomarker/clinical variable claims are directly practiced by ProstaSignal™. The DOE theory for the '893 Patent requires careful claim construction positioning and affirmative rebuttal of the Festo presumption. We recommend that GenomeDx lead with the '271 and '647 Patent literal-infringement theories while maintaining the '893 Patent DOE theory as supplemental support, fortified by the tangential-relation rebuttal argument and strong expert testimony from Dr. Franklin Ogawa.

This memorandum should be read in conjunction with the Preliminary Claim Construction Memorandum (dated August 2025) and the Pre-Filing Investigation Memorandum, which together contain the full literal-infringement and invalidity analyses for all asserted claims.

II. Legal Framework — Prosecution History Estoppel

Before turning to the patent-specific prosecution histories, it is useful to set forth the legal framework governing prosecution history estoppel and its interaction with the doctrine of equivalents. This framework will inform the analysis in Sections III through V below.

A. Doctrine of Equivalents — General Principles

The doctrine of equivalents is a judicially created doctrine that allows a patentee to establish infringement even where the accused product or method does not literally satisfy every limitation of a patent claim. Under the DOE, a patent claim may cover an accused device or method that does not literally fall within the claim's scope if each limitation of the claim is met by a corresponding element in the accused product that is "equivalent" to the claimed element. The Supreme Court has long recognized that the DOE serves an essential purpose in preventing a copyist from evading patent protection by making only insubstantial changes to a patented invention. See Graver Tank & Mfg. Co. v. Linde Air Products Co., 339 U.S. 605, 607 (1950).

The equivalence inquiry may be conducted under either of two tests. Under the "function-way-result" test, a claim limitation is met by an equivalent element if that element performs substantially the same function, in substantially the same way, to achieve substantially the same result as the claimed limitation. Id. at 608. Alternatively, a court may apply the "insubstantial differences" test, which asks whether the differences between the claimed element and the accused element are insubstantial. Warner-Jenkinson Co. v. Harborline Enterprises Davis Chemical Co., 520 U.S. 17, 40 (1997). Either formulation is acceptable, and courts in the Federal Circuit regularly apply both. The choice of test can be outcome-determinative in cases involving complex technical subject matter, such as the machine-learning classifier technologies at issue here.

The DOE is applied on a limitation-by-limitation basis — that is, equivalence is assessed for each claim element individually, not for the claim as a whole. Warner-Jenkinson, 520 U.S. at 29. This element-by-element approach means that the PHE analysis must also proceed on a limitation-by-limitation basis, examining whether amendments or arguments during prosecution affect the availability of DOE for each particular claim element.

B. Prosecution History Estoppel — The *Festo* Framework

The principal limitation on the doctrine of equivalents is prosecution history estoppel. PHE provides that a patentee who has narrowed a claim during patent prosecution — either by formal amendment or by argument — may be precluded from later asserting that the surrendered claim scope is recaptured through the DOE.

The Supreme Court established the modern framework for amendment-based PHE in Festo Corp. v. Fenmore & Co., 535 U.S. 722 (2002). Under Festo, when a patentee makes a narrowing amendment to a claim during prosecution for reasons related to patentability, a presumption arises that the patentee has surrendered the territory between the original claim limitation and the amended claim limitation. Id. at 740. This presumption applies regardless of whether the amendment was made to overcome prior art under Sections 102 or 103 of the Patent Act, or to address other statutory requirements such as patent eligibility under Section 101 or the written description and enablement requirements of Section 112. Id. at 736-37. The Court reasoned that any narrowing amendment made to secure patent issuance represents a concession by the patentee that the broader claim language was not warranted, and competitors are entitled to rely on that concession in determining the scope of the patent right.

The Festo presumption significantly expanded the reach of PHE beyond its earlier articulation. Before Festo, the Court in Warner-Jenkinson Co. v. Harborline Enterprises Davis Chemical Co., 520 U.S. 17 (1997), held that an unexplained amendment to a patent claim creates a presumption that the amendment was made for a reason related to patentability. Under Warner-Jenkinson, if the reason for the amendment is unknown, the patentee bears the burden of establishing that the amendment was not related to patentability. Id. at 33. If the patentee cannot carry this burden, the amendment is presumed to be a patentability-driven narrowing, and PHE attaches. Together, Warner-Jenkinson and Festo create a comprehensive framework under which virtually any narrowing amendment made during prosecution may trigger PHE, regardless of whether the reason for the amendment is known or unknown, and regardless of the specific statutory basis for the amendment.

C. Rebutting the *Festo* Presumption

The Festo presumption is not irrebuttable. The Supreme Court identified three avenues by which a patentee may overcome the presumption of surrender:

(1) Unforeseeability. The patentee may show that the alleged equivalent was unforeseeable at the time of the amendment. If the equivalent technology or embodiment did not exist or was not known in the art at the time the narrowing amendment was made, the patentee could not reasonably be expected to have drafted claims covering it. Festo, 535 U.S. at 740-41. This exception recognizes that patent prosecution occurs against a backdrop of existing technology, and patentees should not be penalized for failing to anticipate future developments.

(2) Tangential Relation. The patentee may show that the rationale underlying the amendment bears no more than a tangential relation to the equivalent in question. Id. at 740. This exception applies where the amendment was made for reasons unrelated to the particular feature that distinguishes the accused equivalent from the claimed limitation. For example, if a claim was narrowed to address a written-description deficiency in one aspect of the claim, and the equivalent at issue relates to a different aspect of the claim, the tangential-relation exception may apply.

(3) Other Reason. The patentee may show that there is some other reason suggesting that the patentee could not reasonably be expected to have described the insubstantial substitute in question. Id. This is a catch-all exception that provides flexibility in unusual circumstances.

The burden of establishing any of these three grounds rests on the patentee. Festo, 535 U.S. at 741. In the context of our case, GenomeDx would bear the burden of rebutting the Festo presumption if it arises with respect to any of the patents-in-suit.

D. Argument-Based Estoppel

Prosecution history estoppel is not limited to formal claim amendments. A distinct but related doctrine — argument-based estoppel — provides that statements made by the patent applicant to distinguish prior art or to characterize the invention during prosecution can also limit claim scope under the DOE, even where no formal amendment has been made. See Pharmacia & Upjohn Co. v. Mylan Pharms., Inc., 170 F.3d 1373, 1377 (Fed. Cir. 1999); Schwarz Pharma, Inc. v. Paddock Labs., Inc., 504 F.3d 1371, 1375 (Fed. Cir. 2007).

Argument-based estoppel arises when the applicant makes clear and unmistakable statements to the examiner that limit the scope of the claims. The Federal Circuit has held that an applicant's arguments during prosecution regarding the meaning or scope of claim terms can serve as a disclaimer of claim scope, precluding the patentee from later asserting a broader interpretation. See Omega Eng'g, Inc. v. Raytek Corp., 334 F.3d 1314, 1323-26 (Fed. Cir. 2003). The estoppel extends to any claim scope that is inconsistent with the arguments made during prosecution. Importantly, the applicant's statements must be "clear and unmistakable" to give rise to argument-based estoppel; ambiguous or equivocal statements may not suffice. Id. at 1325-26.

Argument-based estoppel is particularly relevant to the '893 Patent prosecution, as discussed in detail in Section IV below, where GenomeDx's patent counsel made arguments to the examiner specifically distinguishing the claimed "random-forest classifier" from the SVM classifier disclosed in the Nakamura prior art reference.

E. Voluntary vs. Required Amendments

A further doctrinal nuance concerns the distinction between voluntary amendments and amendments required by an examiner's rejection. Under Festo and its progeny, amendments that are not made in direct response to an examiner rejection — i.e., voluntary narrowing amendments — can also give rise to PHE. See Honeywell Int'l Inc. v. Hamilton Sundstrand Corp., 370 F.3d 1131, 1143 (Fed. Cir. 2004) (en banc). The key inquiry is whether the amendment narrowed the claim scope, not whether it was compelled by the examiner.

That said, the Federal Circuit has recognized an important distinction between "narrowing" amendments and "clarifying" amendments. A clarifying amendment — one that does not substantively change the scope of the claim but merely clarifies an ambiguity or addresses a formality issue — does not trigger PHE. See Severin Holdings v. Quillon Corporation, 289 F.3d 801, 813 (Fed. Cir. 2002). Determining whether a given amendment is narrowing or clarifying requires careful analysis of the prosecution record, including the examiner's stated reasons for any rejection and the applicant's stated reasons for the amendment.

This distinction is relevant to our analysis of the '271 Patent, where the amendments were made to overcome a Section 101 rejection. GenomeDx's counsel framed the amendments as adding "specific, concrete" technical steps to transform the claim from an abstract idea to a patent-eligible method — an argument that could be characterized as either narrowing (adding new limitations) or clarifying (specifying what was inherent in the original method). We analyze this question in detail in Section III.D below.

III. Prosecution History of the '271 Patent (U.S. Patent No. 10,438,271)

A. Overview. The '271 Patent is titled "Methods for Predicting Prostate Cancer Outcomes Using Multi-Gene Expression Classifiers." It was filed on June 14, 2016 and issued on October 8, 2019, with an expiration date of June 14, 2036. The named inventors are Dr. Rajan Mehta, Dr. Carolyn Xu, and Dr. David Petrossian, all of whom were employed by GenomeDx at the time of filing. The '271 Patent is the parent application in the family that includes the '893 and '647 Patents.

GenomeDx asserts Claims 1, 3, 5, 7, 12, and 18 of the '271 Patent in this litigation. Claim 1 is the sole independent claim among the asserted claims and recites a method for generating a prognostic risk score for prostate cancer recurrence using a trained machine-learning classifier applied to normalized gene-expression data obtained from a biological sample. The dependent claims add further limitations relating to specific gene panels, specific normalization techniques, output thresholds, and reporting formats.

The prosecution history of the '271 Patent is relatively straightforward, involving two office actions before issuance. Importantly for this analysis, no prior-art rejections under Sections 102 or 103 were sustained during prosecution, and the sole substantive rejection was a Section 101 patent-eligibility rejection.

B. First Office Action — Section 101 Rejection

The Patent and Trademark Office ("PTO") issued a first office action on November 7, 2017, rejecting all then-pending claims of the '271 Patent under 35 U.S.C. § 101 as directed to patent-ineligible subject matter. The examiner determined that the claims were directed to abstract ideas and natural phenomena — specifically, the natural correlation between gene expression levels and cancer prognosis, and the mathematical algorithms used to analyze those expression levels.

The examiner cited the Supreme Court's decisions in Mayo Collaborative Services v. Prometheus Laboratories, Inc., 566 U.S. 66 (2012) and Association for Molecular Pathology v. Myriad Genetics, Inc., 569 U.S. 576 (2013), as well as the PTO's then-current Guidance on patent eligibility for diagnostic and personalized-medicine claims. Under the Alice/Mayo two-step framework, the examiner found at Step 1 that the claims were directed to an abstract idea (applying a mathematical algorithm to data) and a natural phenomenon (the relationship between gene expression and cancer recurrence), and at Step 2 that the claims did not recite an "inventive concept" sufficient to transform them into patent-eligible subject matter, because the recited steps were described at a high level of generality.

Original Claim 1 as filed broadly recited:

A method for prognosing prostate cancer recurrence, comprising:

(a) obtaining a biological sample containing tumor RNA from a patient who has undergone prostatectomy;

(b) measuring expression levels of a panel of genes in said biological sample;

(c) applying a trained machine-learning classifier to the measured expression levels to generate a risk score indicative of the probability of cancer recurrence; and

(d) classifying the patient into a risk category based on the risk score.

The examiner found this claim to be drawn at too high a level of generality, noting that steps (a) and (b) amounted to conventional data-gathering steps and that steps (c) and (d) amounted to applying a mathematical algorithm to the gathered data.

It is noteworthy that no prior-art rejections under Sections 102 or 103 were sustained at any point during prosecution of the '271 Patent. The examiner identified no references that anticipated or rendered obvious the claimed method. This fact simplifies the PHE analysis because the only estoppel concern arises from the Section 101-driven amendments rather than from any prior-art-related narrowing.

C. Applicant's Response — Amendment to Overcome Section 101

On May 14, 2018, GenomeDx's prosecuting counsel filed a response to the first office action that included amendments to Claim 1 and supporting arguments regarding patent eligibility. The key amendments to Claim 1 were as follows:

(a) The step of "obtaining a biological sample containing tumor RNA from a patient who has undergone prostatectomy" was amended to read: "obtaining a biological sample comprising tumor RNA extracted from a formalin-fixed paraffin-embedded (FFPE) tissue sample from a patient who has undergone prostatectomy."

(b) The step of "measuring expression levels of a panel of genes in said biological sample" was amended to read: "measuring expression levels of a panel of at least 17 genes in said biological sample, wherein measuring expression levels comprises performing reference-gene normalization using at least three housekeeping genes to account for variability in RNA quality and quantity."

(c) The step of "applying a trained machine-learning classifier to the measured expression levels to generate a risk score" was not amended — the term "trained machine-learning classifier" was retained without modification.

In the accompanying remarks, GenomeDx's counsel argued that the amended claims recite a "specific, concrete diagnostic improvement that transforms the claim from an abstract idea to a patent-eligible method rooted in a particular technological process." Counsel emphasized that the use of FFPE tissue samples and reference-gene normalization are specific laboratory techniques that address real-world technical challenges in molecular diagnostics — specifically, the degradation of RNA in archived tissue samples and the variability in RNA quality across samples. Counsel argued that these steps are not mere "data-gathering" but represent meaningful technical choices that improve the accuracy and reliability of the diagnostic method.

The examiner was persuaded by these amendments and arguments. The Section 101 rejection was withdrawn in a second office action dated September 21, 2018, which was a Notice of Allowance with minor formality corrections. The '271 Patent issued on October 8, 2019.

D. PHE Analysis — '271 Patent

We now analyze the prosecution history estoppel implications of the amendments made during prosecution of the '271 Patent.

Amendment Identified. Two narrowing amendments to Claim 1 are identified: (1) the addition of the requirement that the biological sample comprises tumor RNA "extracted from a formalin-fixed paraffin-embedded (FFPE) tissue sample"; and (2) the addition of the requirement that measuring expression levels "comprises performing reference-gene normalization using at least three housekeeping genes to account for variability in RNA quality and quantity."

Nature of Amendment. Both amendments are narrowing in nature — they added specific technical limitations that were not present in the original claims. The original Claim 1 broadly recited "obtaining a biological sample containing tumor RNA," which would have encompassed fresh tissue, frozen tissue, blood-based liquid biopsy samples, and other sources in addition to FFPE tissue. The amended claim is limited to FFPE samples. Similarly, the original claim did not specify any particular normalization methodology; the amended claim requires reference-gene normalization using at least three housekeeping genes.

Statutory Basis. The amendments were made to overcome the examiner's Section 101 patent-eligibility rejection, not to distinguish prior art under Sections 102 or 103. Under Festo, however, amendments made to satisfy any requirement of the Patent Act — including Section 101 — create the presumption of prosecution history estoppel. Festo, 535 U.S. at 736-37. Therefore, the Festo presumption applies to these amendments.

Scope of Surrender. The territory surrendered by these amendments is: (1) the territory between the original broad limitation of "obtaining a biological sample containing tumor RNA" and the amended limitation of "obtaining ... from a formalin-fixed paraffin-embedded (FFPE) tissue sample" — i.e., non-FFPE biological samples such as fresh tissue, frozen tissue, and liquid biopsy samples; and (2) the territory between the original broad limitation of "measuring expression levels" (without specifying a normalization methodology) and the amended limitation of "measuring expression levels ... using reference-gene normalization" — i.e., normalization methods other than reference-gene normalization using at least three housekeeping genes, or no normalization at all.

Impact on DOE Against Helios. The practical impact of this PHE on GenomeDx's case against Helios is limited. Our investigation of Helios's ProstaSignal™ product confirms that ProstaSignal™ does use FFPE tissue samples as its input material and does employ reference-gene normalization using housekeeping genes (specifically, ProstaSignal™ uses five housekeeping genes for normalization, exceeding the claim's minimum of three). Therefore, GenomeDx is asserting literal infringement of these amended limitations — not DOE. Because GenomeDx does not need to extend these claim limitations via DOE to reach Helios's product, the PHE arising from the Section 101 amendments has no practical effect on the infringement case.

Key Risk Scenario. There is one scenario in which the PHE from the '271 Patent prosecution could become relevant. If Helios argues at claim construction that "reference-gene normalization" as used in the '271 Patent has a specific technical meaning that is narrower than GenomeDx contends — for example, if Helios argues that the term requires a particular mathematical normalization algorithm (such as the delta-delta-Ct method) and excludes Helios's particular normalization methodology (which we understand uses a regression-based approach) — then GenomeDx might need to invoke DOE to argue that Helios's normalization method is equivalent to the claimed "reference-gene normalization." In that scenario, the PHE from the Section 101 amendment could potentially foreclose DOE for this limitation. However, GenomeDx's primary position, supported by the claim construction analysis in the Preliminary Claim Construction Memorandum, is that "reference-gene normalization" is a broad term that encompasses any normalization method using housekeeping reference genes, without limitation to a particular mathematical approach. We consider GenomeDx's literal-infringement position on this limitation to be strong.

"Trained Machine-Learning Classifier" — No PHE. Critically, the term "trained machine-learning classifier" was present in the original Claim 1 as filed and was not amended at any point during prosecution of the '271 Patent. It was carried through to the issued claim without modification. As a result, no prosecution history estoppel of any kind — neither amendment-based nor argument-based — attaches to this term for the '271 Patent. This is significant because GenomeDx's primary theory of literal infringement under the '271 Patent rests on the contention that "trained machine-learning classifier" encompasses both random-forest classifiers and gradient-boosted decision tree classifiers, the latter being the type used by Helios's ProstaSignal™ product. The absence of any PHE on this term preserves GenomeDx's full DOE flexibility should it prove necessary, though our expectation is that literal infringement will be established on this element.

Conclusion for the '271 Patent. PHE risk is LOW for the DOE theories relevant to this case. The narrowing amendments were directed to sample-preparation and normalization limitations that Helios's product literally satisfies, and the critical "trained machine-learning classifier" limitation was never amended during prosecution.

IV. Prosecution History of the '893 Patent (U.S. Patent No. 11,204,893)

A. Overview. The '893 Patent is titled "Genomic Classifier Systems and Methods for Post-Surgical Cancer Prognosis." It was filed on March 22, 2019 as a continuation of the '271 Patent, sharing the same priority chain and the same ultimate expiration date of June 14, 2036. The named inventors are Dr. Rajan Mehta and Dr. Carolyn Xu. The '893 Patent issued on December 21, 2021.

GenomeDx asserts Claims 1, 2, 9, 14, and 20 of the '893 Patent. Claim 1 is the independent claim and specifically recites "inputting normalized expression values into a random-forest classifier trained on a cohort of at least 500 patient samples" — a significantly narrower limitation than the '271 Patent's "trained machine-learning classifier" language. This narrower claim language is the source of the most significant PHE risk in the case.

The prosecution history of the '893 Patent involved one office action before issuance. Unlike the '271 Patent, the '893 Patent faced a prior-art anticipation rejection under Section 102, which the applicant overcame through argument-based distinction rather than claim amendment.

B. Office Action — Section 102 Rejection Based on Nakamura

The PTO issued an office action on January 14, 2021, rejecting all then-pending claims of the '893 Patent under 35 U.S.C. § 102(a)(1) as anticipated by Nakamura et al., U.S. Patent No. 9,617,340 (titled "Gene Expression Profiling for Cancer Prognosis," issued April 4, 2017) (hereinafter "Nakamura"). Nakamura disclosed a method of gene expression profiling using machine-learning classifiers to generate cancer prognosis scores for urological cancers. The examiner found that Nakamura's disclosure of a machine-learning classifier applied to gene-expression data to generate a cancer prognosis score anticipated the originally-filed claims of the '893 Patent.

Specifically, the examiner identified the following correspondences between Nakamura and the claims of the '893 Patent: (1) Nakamura disclosed obtaining gene expression data from tumor tissue; (2) Nakamura disclosed normalizing gene expression values; (3) Nakamura disclosed inputting normalized expression values into a machine-learning classifier to generate a prognostic score; and (4) Nakamura disclosed classifying the patient based on the output score. The examiner concluded that these correspondences were sufficient to establish anticipation under Section 102.

C. Applicant's Response — Argument and Claim Differentiation

On July 19, 2021, GenomeDx's prosecuting counsel filed a response to the office action. Notably, counsel did not amend the independent claim language. Instead, counsel distinguished Nakamura through argument, relying on claim limitations already present in the as-filed claims of the '893 Patent. The key arguments are set forth below, and verbatim excerpts are provided in Appendix B.

First, counsel emphasized that Claim 1 of the '893 Patent recites "inputting normalized expression values into a random-forest classifier trained on a cohort of at least 500 patient samples" (emphasis added by counsel in the response). Counsel argued that Nakamura's classifier was a support vector machine ("SVM"), not a random-forest classifier, and that these are fundamentally different machine-learning methodologies. Counsel stated:

"Nakamura's classifier is an SVM, which is a single-model discriminative classifier that operates by finding an optimal separating hyperplane in high-dimensional feature space. In contrast, applicant's claimed random-forest classifier is an ensemble learning method that constructs a multitude of independently trained decision trees at training time and outputs the class that is the mode of the classes predicted by the individual trees. The random-forest methodology provides distinct technical advantages not present in Nakamura's SVM approach, including but not limited to: (i) ensemble learning through bootstrap aggregating (bagging) of multiple independently constructed decision trees; (ii) out-of-bag error estimation, which provides an unbiased estimate of generalization error without the need for a separate validation set; and (iii) feature importance ranking derived from the contribution of each gene to the ensemble's predictive accuracy. These are structural and methodological differences, not merely implementation details."

Second, counsel argued that Nakamura did not disclose the minimum cohort size of 500 patient samples for training the classifier, which is a specific technical requirement of the '893 Patent claims. Counsel pointed to the specification of the '893 Patent, which explains that training on a cohort of at least 500 samples improves the statistical power and generalizability of the random-forest model.

The examiner was persuaded by these arguments. In a subsequent communication dated September 8, 2021, the examiner withdrew the Section 102 rejection and allowed the claims. The '893 Patent issued on December 21, 2021.

D. PHE Analysis — '893 Patent (Critical Analysis)

The '893 Patent prosecution history presents the most significant PHE risk in this litigation. We analyze this risk in detail below, addressing amendment-based estoppel, argument-based estoppel, the DOE analysis for the relevant limitation, potential Festo rebuttal arguments, and our overall conclusions.

i. Amendment-Based Estoppel. Strictly speaking, amendment-based estoppel under Festo does not apply to the "random-forest classifier" term in the '893 Patent. The claim language reciting "a random-forest classifier" was present in the originally-filed claims of the '893 Patent as filed on March 22, 2019. No formal amendment to Claim 1 was made at any point during prosecution — counsel distinguished Nakamura through argument alone. Because there was no narrowing amendment to the claim language, the Festo presumption of surrender does not technically arise with respect to the "random-forest classifier" limitation.

No other narrowing amendments were identified in the prosecution file wrapper of the '893 Patent. The only amendments made during prosecution were minor formatting corrections that did not change claim scope.

However, we note a potential argument that Helios could raise: because the '893 Patent is a continuation of the '271 Patent, and because the '271 Patent claims the broader term "trained machine-learning classifier" while the '893 Patent claims the narrower term "random-forest classifier," Helios could argue that GenomeDx's deliberate choice to claim only "random-forest classifier" in the continuation application constitutes a constructive narrowing — analogous to a formal amendment — that triggers amendment-based estoppel. The Federal Circuit has not squarely addressed whether the choice of narrower claim language in a continuation application, without any prosecution-driven amendment, gives rise to Festo-style amendment-based estoppel. The safer analytical approach is to treat this as a potential issue but to focus our analysis on argument-based estoppel, which presents the more immediate risk as discussed below.

ii. Argument-Based Estoppel — The Primary Risk. Argument-based estoppel is the most significant PHE concern for the '893 Patent. As detailed in Section IV.C above, GenomeDx's prosecuting counsel made affirmative arguments to the examiner specifically distinguishing the claimed "random-forest classifier" from Nakamura's SVM classifier. These arguments emphasized the specific technical characteristics of random-forest classifiers that differentiate them from SVMs — namely, ensemble learning through bagging, out-of-bag error estimation, and feature importance ranking.

Under argument-based estoppel principles, GenomeDx may be estopped from asserting that Helios's gradient-boosted decision tree classifier is "equivalent" to the claimed "random-forest classifier" if the prosecution arguments are read as disclaiming non-random-forest machine-learning methods generally. The critical question is the breadth of the disclaimer: did GenomeDx disclaim all non-random-forest classifiers, or did it disclaim only SVMs specifically?

Favorable Reading for GenomeDx. A fair reading of the prosecution arguments supports the position that GenomeDx's disclaimer was narrow and directed specifically to SVMs. The applicant's arguments were made in response to a Section 102 rejection based on Nakamura, which disclosed an SVM classifier. The arguments were specifically tailored to distinguish Nakamura's SVM — the applicant did not make sweeping statements disclaiming all machine-learning classifiers other than random forests. The applicant did not state, for example, that "the claimed random-forest classifier excludes all other types of machine-learning classifiers" or that "only random-forest classifiers are within the scope of the claims." Rather, the applicant argued that "Nakamura's classifier is an SVM," which is "fundamentally different" from a random-forest classifier, and then described the specific attributes that distinguish random forests from SVMs.

Importantly, the technical distinctions emphasized by the applicant — ensemble learning, out-of-bag error estimation, and feature importance ranking — are characteristics that distinguish random forests from SVMs specifically. Gradient-boosted decision trees share several of these characteristics with random forests: both are ensemble methods that use decision trees as base learners, both produce feature importance rankings, and both are used extensively in genomics for the same types of predictive tasks. The prosecution arguments were targeted at the structural differences between random forests and SVMs, not at the differences between random forests and other tree-based ensemble methods.

Unfavorable Reading (Helios's Likely Position). Helios will likely argue for a broader reading of the prosecution disclaimer. Helios's counsel at Ridgecrest Durneval LLP (lead partner Marcus Pellegrini) will contend that GenomeDx's prosecution arguments emphasized specific attributes of random-forest classifiers — ensemble learning through bagging, out-of-bag error estimation, and feature importance ranking — as the basis for patentability, thereby limiting the claims to classifiers that possess these exact attributes. Helios will argue that gradient-boosted decision trees, while tree-based, lack several of these specific attributes: gradient boosting uses sequential training rather than bagging; gradient-boosted trees do not employ out-of-bag error estimation in the same manner as random forests (because each tree is trained on the full dataset with a loss-gradient weighting, not on a bootstrap sample); and the mechanism for deriving feature importance differs between the two methods. Helios will contend that by identifying these specific attributes as the distinguishing features of the claimed invention, GenomeDx effectively disclaimed any classifier that does not possess them.

Assessment. This is a genuinely contested question that does not admit of easy resolution on the current record. The breadth of the argument-based disclaimer will likely be a focal point of briefing at the Markman hearing, currently scheduled for September 18, 2026. Both sides can marshal reasonable arguments, and the outcome may depend on how the court interprets the specificity of GenomeDx's prosecution statements and whether the court views gradient-boosted decision trees as falling within or outside the scope of those statements.

We assess the argument-based estoppel theory as Helios's strongest defense against the DOE claim under the '893 Patent. However, we note that argument-based estoppel requires a showing of "clear and unmistakable" disclaimer, and GenomeDx can credibly argue that its prosecution statements were specifically directed to distinguishing SVMs, not to disclaiming all non-random-forest classifiers. The question of whether the argument-based estoppel extends to gradient-boosted decision trees — which are much more closely related to random forests than SVMs are — presents a genuinely close issue.

iii. DOE Analysis for Random Forest vs. Gradient-Boosted Decision Tree. Assuming that the "random-forest classifier" limitation is construed to exclude gradient-boosted decision trees under a literal-infringement theory, we analyze the viability of asserting DOE for this limitation.

Function-Way-Result Test:

Function: Both random forests and gradient-boosted decision trees function as ensemble tree-based classifiers that take gene-expression input data and generate a predictive risk score correlated with clinical outcomes. In the specific context of the claimed invention, both methods serve the function of transforming normalized gene-expression values into a numerical prognostic score for cancer recurrence. This is substantially the same function.

Way: This is the weakest element of the function-way-result analysis for GenomeDx. Random forests use bagging (bootstrap aggregating) to train an ensemble of independently constructed decision trees, with the final prediction determined by majority voting (for classification) or averaging (for regression) across all trees. Gradient-boosted decision trees use sequential training of shallow decision trees where each successive tree is trained to correct the errors (residuals) of the prior ensemble, with the final prediction determined by a weighted sum of all trees' outputs. These are meaningfully different ensemble-construction methodologies. Random forests emphasize variance reduction through averaging of independently trained high-variance models; gradient boosting emphasizes bias reduction through iterative refinement. Helios will argue that these differences are substantial, not insubstantial.

However, GenomeDx can argue that the "way" should be assessed at the appropriate level of abstraction — which, in the context of the claimed invention, is the use of an ensemble of decision trees to produce a gene-expression-based prognostic score. At this level, both methods construct ensembles of decision trees and aggregate their predictions to produce a single output score. The differences in how the ensemble is constructed (bagging vs. boosting) are implementation details that do not change the fundamental "way" in which the classifier operates — by aggregating multiple tree-based predictors.

Result: Both methods produce a numerical prognostic score correlated with metastasis risk, which is then used to classify the patient into a risk category. The output format, dynamic range, and clinical utility of the scores are substantially the same. This is substantially the same result.

Insubstantial Differences Test (Alternative). Under this alternative formulation, GenomeDx can argue that the differences between gradient-boosted decision trees and random forests are insubstantial in the context of the claimed invention. Both are ensemble tree-based methods that are widely recognized in the machine-learning and genomics literature as closely related variants. Multiple published papers and textbooks discuss them in conjunction as "ensemble tree methods" or "tree-based ensemble approaches." Both were well-known at the time of filing and are routinely used for the same types of gene-expression classification tasks. GenomeDx can point to published comparative studies showing that random forests and gradient-boosted decision trees achieve comparable performance on gene-expression datasets, with neither consistently outperforming the other.

Assessment. The DOE argument for the '893 Patent is arguable but carries meaningful risk, particularly given the prosecution arguments discussed above. The "way" element of the function-way-result test is the weakest link, and Helios will focus its defense on the algorithmic differences between bagging and boosting. GenomeDx's best arguments on DOE are: (1) the insubstantial-differences test, which focuses on the overall technical relationship between the two methods rather than parsing the algorithmic details; (2) expert testimony from Dr. Ogawa establishing that a person of ordinary skill in the art would view gradient-boosted decision trees and random forests as closely related ensemble tree methods; and (3) published literature treating the two methods as interchangeable alternatives for gene-expression classification tasks.

iv. *Festo* Rebuttal Arguments (If Amendment-Based Estoppel Were Applied). Even if a court were to find amendment-based estoppel applicable — for example, by treating the continuation application's deliberate choice of "random-forest classifier" as a constructive narrowing relative to the parent '271 Patent's broader "trained machine-learning classifier" language — GenomeDx has three potential rebuttal arguments under Festo:

(1) Unforeseeability. Gradient-boosted decision tree methods, including the widely-used XGBoost framework (published by Chen and Guestrin in 2016), were well-known in the field at the time the '893 Patent was filed on March 22, 2019. The LightGBM and CatBoost frameworks were also publicly available before the filing date. Accordingly, the gradient-boosted decision tree equivalent was clearly foreseeable at the time of filing. This rebuttal argument is weak and we do not recommend relying on it.

(2) Tangential Relation. The prosecution arguments were directed specifically to distinguishing Nakamura's SVM classifier. The rationale underlying the distinction — that SVMs are single-model discriminative classifiers operating via hyperplane separation, fundamentally different from ensemble tree methods — bears only a tangential relation to gradient-boosted decision trees, which are ensemble tree-based methods sharing the same general architecture as random forests. The reasons for distinguishing an SVM from a random forest (single model vs. ensemble; hyperplane vs. decision trees) do not logically extend to distinguishing a gradient-boosted tree ensemble from a random-forest ensemble, because gradient-boosted trees share the very characteristics that the applicant identified as distinguishing features — ensemble architecture and decision-tree-based prediction. This rebuttal argument is the strongest of the three and we recommend emphasizing it.

(3) Other Reason. GenomeDx could argue that it could not reasonably have been expected to describe every tree-based ensemble variant in its claims, because the field of machine learning was rapidly evolving and new ensemble methods were being developed continuously. However, this argument is undermined by the fact that gradient boosting was a well-established technique by 2019, having been used extensively in data science competitions and published applications for several years prior to the '893 Patent filing. This rebuttal argument is weak and should be used only as a supplementary argument.

Strongest Rebuttal: Tangential Relation. To summarize, if amendment-based estoppel is found to apply, GenomeDx's strongest rebuttal is the tangential-relation prong. The prosecution arguments were specifically aimed at SVMs, and the rationale for distinguishing SVMs (lack of ensemble tree structure, lack of out-of-bag error estimation, fundamentally different model architecture) bears only a tangential relation to gradient-boosted decision trees, which are tree-based ensembles that share far more in common with random forests than SVMs do. A court applying the tangential-relation test should find that the reasons for the applicant's distinction of Nakamura's SVM do not logically compel the exclusion of gradient-boosted decision trees from the scope of equivalents.

v. Conclusion for the '893 Patent. PHE risk is MODERATE TO HIGH for asserting DOE on the "random-forest classifier" limitation of the '893 Patent. Argument-based estoppel is the most significant concern. The applicant's prosecution arguments clearly distinguished the claimed random-forest classifier from Nakamura's SVM on the basis of specific technical attributes, and Helios will argue that these statements disclaim any classifier lacking those exact attributes — including gradient-boosted decision trees. While GenomeDx has a credible response (the arguments were narrowly targeted at SVMs), this is a genuinely contested issue that may be resolved against GenomeDx at the Markman hearing or at trial.

To mitigate this risk, we recommend the following strategic approaches:

(a) GenomeDx should pursue a claim construction of "random-forest classifier" that is broad enough to literally encompass gradient-boosted decision trees — or at least to create sufficient ambiguity that the issue goes to the jury. If the term is construed to include ensemble tree-based classifiers that use decision trees as base learners (regardless of whether the ensemble is constructed via bagging or boosting), the DOE/PHE problem is eliminated entirely. However, this construction faces significant technical and linguistic challenges, because "random forest" has a specific meaning in the machine-learning literature that is generally understood to refer to the bagging-based ensemble method developed by Breiman (2001).

(b) As an alternative, GenomeDx should prepare a robust DOE argument supported by expert testimony from Dr. Ogawa, emphasizing the tangential-relation rebuttal to the Festo presumption and the insubstantial-differences test for equivalence.

(c) GenomeDx should also emphasize the dependent claims of the '893 Patent (Claims 2, 9, 14, and 20), which contain additional limitations that may be met literally by ProstaSignal™. If the independent claim's "random-forest classifier" limitation proves to be an obstacle under DOE/PHE, the dependent claims may provide alternative paths to infringement that reduce the importance of the independent claim's DOE argument. A detailed analysis of the dependent claims' infringement posture is provided in the Preliminary Claim Construction Memorandum.

V. Prosecution History of the '647 Patent (U.S. Patent No. 11,891,647)

A. Overview. The '647 Patent is titled "Improved Methods for Genomic Risk Stratification in Prostate Cancer Using Combined Biomarker and Clinical Variable Models." It was filed on September 9, 2021 and issued on February 6, 2024, with an expiration date of September 9, 2041 (subject to the terminal disclaimer discussed below). The named inventors are Dr. Rajan Mehta, Dr. Kevin Ostrowski, and Dr. Priya Shah.

GenomeDx asserts Claims 1, 4, 6, 11, and 15 of the '647 Patent. Claim 1 is the independent claim and recites a method for genomic risk stratification that combines a genomic risk score derived from at least 20 gene features with at least two clinical covariates selected from the group consisting of Gleason score, PSA level, surgical margin status, and pathological stage. The '647 Patent thus differs from the '271 and '893 Patents in its focus on the combination of genomic and clinical data, rather than on the specific machine-learning classifier used to process genomic data alone.

The prosecution history of the '647 Patent was more complex than the other two patents, involving three office actions before issuance. However, as detailed below, no narrowing amendments were made to the claim language, and the rejections were overcome through evidentiary showing.

B. First and Second Office Actions — Section 103 Rejections

The PTO issued a first office action on April 3, 2022, rejecting all then-pending claims of the '647 Patent under 35 U.S.C. § 103 as obvious over the combination of Nakamura et al. (U.S. Patent No. 9,617,340) and Gagnon (WO 2018/065413, titled "Multi-Gene Prognostic Assays for Urological Cancers," published April 12, 2018). The examiner found that Nakamura disclosed gene-expression-based cancer prognosis using machine-learning classifiers and that Gagnon disclosed methods for combining genomic risk scores with clinical covariates (including Gleason score and PSA level) to generate integrated prognostic models for urological cancers. The examiner concluded that it would have been obvious to a person of ordinary skill in the art to combine the teachings of Nakamura and Gagnon to arrive at the claimed method.

GenomeDx's counsel filed a response on October 5, 2022, arguing that the combination of Nakamura and Gagnon did not render the claims obvious because neither reference, alone or in combination, taught or suggested the specific claimed combination of at least 20 gene features with at least two of the four enumerated clinical covariates, or the specific synergistic improvement in predictive accuracy that results from this combination. The examiner was not fully persuaded and maintained the Section 103 rejection in a second office action dated February 14, 2023, noting that Gagnon's disclosure of combining genomic and clinical data was sufficiently analogous to the claimed method and that the claimed minimum of 20 gene features was within the routine optimization capabilities of a skilled artisan.

C. Applicant's Response — Declaration Under 37 C.F.R. § 1.132

In response to the maintained Section 103 rejection, GenomeDx's counsel filed a supplemental response on August 28, 2023 that included a declaration under 37 C.F.R. § 1.132 from Dr. Rajan Mehta, the lead inventor. The Mehta declaration presented detailed experimental data establishing that the specific combination recited in the claims — at least 20 gene features combined with at least two clinical covariates selected from Gleason score, PSA level, surgical margin status, and pathological stage — produces unexpected results compared to genomic-only models or clinical-only models.

Specifically, the Mehta declaration presented comparative performance data from a validation cohort of 1,247 patients, showing that the combined genomic-clinical model achieved an area under the receiver operating characteristic curve ("AUC") of 0.87, compared to 0.78 for the genomic-only model (using the same 22-gene panel without clinical covariates) and 0.72 for the clinical-only model (using Gleason score, PSA, margin status, and pathological stage without genomic features). The declaration explained that the improvement from 0.78 to 0.87 was statistically significant (p < 0.001) and clinically meaningful, representing a qualitative improvement in the ability to discriminate patients at high risk of metastasis from those at low risk.

In the accompanying remarks, counsel argued that this synergistic improvement — the combined model significantly outperforming the sum of its parts — constituted evidence of unexpected results that rebutted the prima facie case of obviousness. Counsel further argued that neither Nakamura nor Gagnon, alone or in combination, taught or suggested that combining genomic features with the specific enumerated clinical covariates would produce this degree of synergistic improvement.

Critically for the PHE analysis, no narrowing amendments to the independent claim language were made at any point during the prosecution of the '647 Patent. The applicant overcame the Section 103 rejection entirely through the evidentiary showing in the Mehta declaration, without amending the claim language.

D. Terminal Disclaimer

During prosecution, GenomeDx filed a terminal disclaimer over the '271 and '893 Patents. This was required due to common inventorship and overlapping subject matter between the patents in the same family. The terminal disclaimer provides that the '647 Patent will not be enforced after the expiration of the '271 and '893 Patents (June 14, 2036) if ownership does not remain common.

The terminal disclaimer does not create prosecution history estoppel. It addresses only the double-patenting issue and limits the enforceable life of the '647 Patent. Because the '647 Patent would otherwise expire on September 9, 2041, the terminal disclaimer effectively shortens its enforceable term to June 14, 2036 (the expiration date of the '271 and '893 Patents) unless ownership remains common. This is a standard double-patenting resolution and has no bearing on the PHE analysis.

E. Third Office Action — Allowance

Following submission of the Mehta declaration, the terminal disclaimer, and the accompanying arguments, the examiner issued a Notice of Allowance on October 30, 2023. The examiner found that the Mehta declaration's evidence of unexpected synergistic results was persuasive in overcoming the Section 103 rejection. The '647 Patent issued on February 6, 2024.

F. PHE Analysis — '647 Patent

The PHE analysis for the '647 Patent is straightforward and favorable for GenomeDx.

No Narrowing Amendments. Because no narrowing amendments were made to the claim language during prosecution, the Festo presumption of surrender does not arise. The applicant overcame the Section 103 rejection through evidence (the Mehta declaration) rather than by amending the claims. The absence of any narrowing amendment eliminates the primary source of PHE risk.

Argument-Based Estoppel Risk. The applicant's arguments during prosecution emphasized the specific combination of genomic and clinical variables as producing unexpected synergistic results. This could theoretically give rise to argument-based estoppel limiting the scope of the "combined" model limitation. However, any such estoppel would not materially affect GenomeDx's infringement case because Helios's ProstaSignal™ product literally satisfies the combined-model limitation. ProstaSignal™ combines its genomic ProstaScore (derived from a 24-gene panel, exceeding the claimed minimum of 20 gene features) with Gleason grade group and pre-operative PSA level — two of the four enumerated clinical covariates. The claim requires "at least two clinical covariates selected from the group consisting of Gleason score, PSA level, surgical margin status, and pathological stage," and Helios's product uses two of these four covariates. Literal infringement is established without resort to DOE.

No DOE Needed. Because the '647 Patent's key limitations appear to be literally infringed by ProstaSignal™, there is no need to invoke the doctrine of equivalents for any limitation of the asserted claims. The PHE analysis is therefore largely academic for the '647 Patent — even if some form of estoppel were found to exist, it would have no practical impact on the infringement case because DOE is not necessary.

Conclusion for the '647 Patent. PHE risk is MINIMAL. No narrowing amendments were made during prosecution, literal infringement is alleged for all limitations, and DOE is not needed for the infringement theory. The '647 Patent provides GenomeDx with a clean, PHE-free path to infringement.

VI. Cross-Patent Analysis and Strategic Implications

A. Summary Table of PHE Risk by Patent

The following table summarizes the PHE risk assessment for each patent-in-suit:

B. Interplay Between the '271 and '893 Patents

The relationship between the '271 and '893 Patents is a critical strategic consideration that warrants careful analysis. The two patents share the same priority chain (the '893 Patent is a continuation of the '271 Patent) and the same expiration date (June 14, 2036). However, they differ in a crucial respect: the '271 Patent claims the broader term "trained machine-learning classifier" in Claim 1, while the '893 Patent claims the narrower term "random-forest classifier" in Claim 1.

This difference has significant implications for the infringement case. If the court adopts GenomeDx's proposed construction of "trained machine-learning classifier" in the '271 Patent — i.e., that the term encompasses any machine-learning classifier that has been trained on data, including gradient-boosted decision tree classifiers — then the '271 Patent provides a direct path to literal infringement of Helios's ProstaSignal™ product without any need to invoke DOE. In this scenario, the '893 Patent's narrower claims become less critical to the overall infringement case, providing additional leverage and redundancy but not serving as the primary basis for liability.

Conversely, if the court construes "trained machine-learning classifier" in the '271 Patent narrowly — for example, by limiting the term to the specific types of classifiers disclosed in the specification (random forests and logistic regression) — then the '271 Patent's literal-infringement theory may be weakened, and the '893 Patent's DOE argument becomes more important. But paradoxically, this is also the scenario in which the '893 Patent's DOE theory is most vulnerable to PHE, because a narrow construction of "trained machine-learning classifier" in the '271 Patent would suggest that the patent family's claims were deliberately drafted to cover specific classifier types rather than the entire field of machine learning.

The practical implication is that GenomeDx's claim construction strategy for the '271 Patent is the single most important determinant of the overall litigation posture. If GenomeDx prevails on a broad construction of "trained machine-learning classifier," the PHE risk for the '893 Patent becomes a secondary concern. If GenomeDx does not prevail on this construction, the case becomes more dependent on the '893 Patent's DOE theory and the '647 Patent's lite

PRIVILEGED & CONFIDENTIAL — ATTORNEY WORK PRODUCT
MEMORANDUM
TO: Catherine Bloom, Partner, intellectual Property Practice Group, Calderwood & Harkness LLP
FROM: Wendy Schultz, Associate, intellectual Property Practice Group, Calderwood & Harkness LLP
DATE: February 19, 2026
RE: Prosecution History Estoppel Analysis — Impact on Claim Scope and Doctrine of Equivalents; Genome Dx Therapeutics Inc. v. Helios Precision Diagnostics, Inc., Case No. 3:25-cv-01847-JWL (N.D. Cal.)
MATTER NO.: 1013-00005
PRIVILEGED & CONFIDENTIAL — ATTORNEY-CLIENT PRIVILEGE / ATTORNEY WORK PRODUCT — PREPARED IN ANTICIPATION OF LITIGATION. This memorandum was prepared by counsel for Genome Dx Therapeutics Inc. in anticipation of litigation and is protected from disclosure by the attorney-client privilege and the work product doctrine. Distribution of this memorandum, in whole or in part, to any person outside the attorney-client relationship may result in waiver of applicable privileges.
Right-click to update Table of Contents
I. Executive Summary
This memorandum assesses whether amendments and arguments made during the prosecution of the three patents-in-suit — U.S. Patent Nos. 10,438,271 (the "'271 Patent"), 11,204,893 (the "'893 Patent"), and 11,891,647 (the "'647 Patent") — create prosecution history estoppel ("PHE") that would limit our client Genome Dx Therapeutics Inc.'s ("GenomeDx") ability to assert infringement under the doctrine of equivalents ("DOE") against Helios Precision Diagnostics, Inc.'s ("Helios") ProstaSignal™ Genomic Classifier product. This analysis is undertaken in connection with the pending litigation captioned above, currently before the Honorable James W. Larkin in the United States District Court for the Northern District of California.
After a detailed review of the complete prosecution file wrappers for each of the three patents-in-suit, the relevant case law, and the technical characteristics of the accused ProstaSignal™ product, we reach the following principal conclusions:
PHE poses a moderate-to-high risk to GenomeDx's DOE theories with respect to the '893 Patent's "random-forest classifier" limitation. This is the most critical DOE argument in the case, because Helios's ProstaSignal™ product uses a gradient-boosted decision tree classifier rather than a random-forest classifier as recited in the '893 Patent claims. During prosecution of the '893 Patent, GenomeDx's patent counsel made affirmative arguments distinguishing the claimed "random-forest classifier" from the support vector machine ("SVM") classifier disclosed in the Nakamura prior art reference. Helios will almost certainly invoke these arguments to contend that GenomeDx disclaimed all non-random-forest machine-learning classifiers, including gradient-boosted decision trees. We assess this argument-based estoppel theory as a genuine litigation risk, although GenomeDx has viable rebuttal arguments — particularly the "tangential relation" prong of the Festo framework.
PHE poses a lower risk for the '271 Patent's amended limitations. The amendments made to the '271 Patent during prosecution were driven by a Section 101 patent-eligibility rejection and added specific laboratory-technique limitations (FFPE sample extraction and reference-gene normalization). However, Helios's ProstaSignal™ product literally satisfies these amended limitations, meaning GenomeDx does not need to invoke DOE for these claim elements. The critical "trained machine-learning classifier" term in the '271 Patent was never amended during prosecution and remains broad enough to encompass gradient-boosted decision tree classifiers on a literal-infringement theory.
PHE risk is minimal for the '647 Patent. No narrowing amendments were made during prosecution of the '647 Patent; the applicant overcame the examiner's obviousness rejection through an evidentiary showing (the Mehta declaration) rather than through claim narrowing. Moreover, GenomeDx alleges literal infringement of the '647 Patent's combined biomarker/clinical variable claims, eliminating the need to rely on DOE.
Strategic Recommendation. GenomeDx's strongest infringement theories are: (1) literal infringement under the '271 Patent, where the broad "trained machine-learning classifier" language was never narrowed during prosecution and comfortably encompasses Helios's gradient-boosted decision tree; and (2) literal infringement under the '647 Patent, where the combined biomarker/clinical variable claims are directly practiced by ProstaSignal™. The DOE theory for the '893 Patent requires careful claim construction positioning and affirmative rebuttal of the Festo presumption. We recommend that GenomeDx lead with the '271 and '647 Patent literal-infringement theories while maintaining the '893 Patent DOE theory as supplemental support, fortified by the tangential-relation rebuttal argument and strong expert testimony from Dr. Franklin Ogawa.
This memorandum should be read in conjunction with the Preliminary Claim Construction Memorandum (dated August 2025) and the Pre-Filing Investigation Memorandum, which together contain the full literal-infringement and invalidity analyses for all asserted claims.
II. Legal Framework — Prosecution History Estoppel
Before turning to the patent-specific prosecution histories, it is useful to set forth the legal framework governing prosecution history estoppel and its interaction with the doctrine of equivalents. This framework will inform the analysis in Sections III through V below.
A. Doctrine of Equivalents — General Principles
The doctrine of equivalents is a judicially created doctrine that allows a patentee to establish infringement even where the accused product or method does not literally satisfy every limitation of a patent claim. Under the DOE, a patent claim may cover an accused device or method that does not literally fall within the claim's scope if each limitation of the claim is met by a corresponding element in the accused product that is "equivalent" to the claimed element. The Supreme Court has long recognized that the DOE serves an essential purpose in preventing a copyist from evading patent protection by making only insubstantial changes to a patented invention. See Graver Tank & Mfg. Co. v. Linde Air Products Co., 339 U.S. 605, 607 (1950).
The equivalence inquiry may be conducted under either of two tests. Under the "function-way-result" test, a claim limitation is met by an equivalent element if that element performs substantially the same function, in substantially the same way, to achieve substantially the same result as the claimed limitation. Id. at 608. Alternatively, a court may apply the "insubstantial differences" test, which asks whether the differences between the claimed element and the accused element are insubstantial. Warner-Jenkinson Co. v. Harborline Enterprises Davis Chemical Co., 520 U.S. 17, 40 (1997). Either formulation is acceptable, and courts in the Federal Circuit regularly apply both. The choice of test can be outcome-determinative in cases involving complex technical subject matter, such as the machine-learning classifier technologies at issue here.
The DOE is applied on a limitation-by-limitation basis — that is, equivalence is assessed for each claim element individually, not for the claim as a whole. Warner-Jenkinson, 520 U.S. at 29. This element-by-element approach means that the PHE analysis must also proceed on a limitation-by-limitation basis, examining whether amendments or arguments during prosecution affect the availability of DOE for each particular claim element.
B. Prosecution History Estoppel — The *Festo* Framework
The principal limitation on the doctrine of equivalents is prosecution history estoppel. PHE provides that a patentee who has narrowed a claim during patent prosecution — either by formal amendment or by argument — may be precluded from later asserting that the surrendered claim scope is recaptured through the DOE.
The Supreme Court established the modern framework for amendment-based PHE in Festo Corp. v. Fenmore & Co., 535 U.S. 722 (2002). Under Festo, when a patentee makes a narrowing amendment to a claim during prosecution for reasons related to patentability, a presumption arises that the patentee has surrendered the territory between the original claim limitation and the amended claim limitation. Id. at 740. This presumption applies regardless of whether the amendment was made to overcome prior art under Sections 102 or 103 of the Patent Act, or to address other statutory requirements such as patent eligibility under Section 101 or the written description and enablement requirements of Section 112. Id. at 736-37. The Court reasoned that any narrowing amendment made to secure patent issuance represents a concession by the patentee that the broader claim language was not warranted, and competitors are entitled to rely on that concession in determining the scope of the patent right.
The Festo presumption significantly expanded the reach of PHE beyond its earlier articulation. Before Festo, the Court in Warner-Jenkinson Co. v. Harborline Enterprises Davis Chemical Co., 520 U.S. 17 (1997), held that an unexplained amendment to a patent claim creates a presumption that the amendment was made for a reason related to patentability. Under Warner-Jenkinson, if the reason for the amendment is unknown, the patentee bears the burden of establishing that the amendment was not related to patentability. Id. at 33. If the patentee cannot carry this burden, the amendment is presumed to be a patentability-driven narrowing, and PHE attaches. Together, Warner-Jenkinson and Festo create a comprehensive framework under which virtually any narrowing amendment made during prosecution may trigger PHE, regardless of whether the reason for the amendment is known or unknown, and regardless of the specific statutory basis for the amendment.
C. Rebutting the *Festo* Presumption
The Festo presumption is not irrebuttable. The Supreme Court identified three avenues by which a patentee may overcome the presumption of surrender:
(1) Unforeseeability. The patentee may show that the alleged equivalent was unforeseeable at the time of the amendment. If the equivalent technology or embodiment did not exist or was not known in the art at the time the narrowing amendment was made, the patentee could not reasonably be expected to have drafted claims covering it. Festo, 535 U.S. at 740-41. This exception recognizes that patent prosecution occurs against a backdrop of existing technology, and patentees should not be penalized for failing to anticipate future developments.
(2) Tangential Relation. The patentee may show that the rationale underlying the amendment bears no more than a tangential relation to the equivalent in question. Id. at 740. This exception applies where the amendment was made for reasons unrelated to the particular feature that distinguishes the accused equivalent from the claimed limitation. For example, if a claim was narrowed to address a written-description deficiency in one aspect of the claim, and the equivalent at issue relates to a different aspect of the claim, the tangential-relation exception may apply.
(3) Other Reason. The patentee may show that there is some other reason suggesting that the patentee could not reasonably be expected to have described the insubstantial substitute in question. Id. This is a catch-all exception that provides flexibility in unusual circumstances.
The burden of establishing any of these three grounds rests on the patentee. Festo, 535 U.S. at 741. In the context of our case, GenomeDx would bear the burden of rebutting the Festo presumption if it arises with respect to any of the patents-in-suit.
D. Argument-Based Estoppel
Prosecution history estoppel is not limited to formal claim amendments. A distinct but related doctrine — argument-based estoppel — provides that statements made by the patent applicant to distinguish prior art or to characterize the invention during prosecution can also limit claim scope under the DOE, even where no formal amendment has been made. See Pharmacia & Upjohn Co. v. Mylan Pharms., Inc., 170 F.3d 1373, 1377 (Fed. Cir. 1999); Schwarz Pharma, Inc. v. Paddock Labs., Inc., 504 F.3d 1371, 1375 (Fed. Cir. 2007).
Argument-based estoppel arises when the applicant makes clear and unmistakable statements to the examiner that limit the scope of the claims. The Federal Circuit has held that an applicant's arguments during prosecution regarding the meaning or scope of claim terms can serve as a disclaimer of claim scope, precluding the patentee from later asserting a broader interpretation. See Omega Eng'g, Inc. v. Raytek Corp., 334 F.3d 1314, 1323-26 (Fed. Cir. 2003). The estoppel extends to any claim scope that is inconsistent with the arguments made during prosecution. Importantly, the applicant's statements must be "clear and unmistakable" to give rise to argument-based estoppel; ambiguous or equivocal statements may not suffice. Id. at 1325-26.
Argument-based estoppel is particularly relevant to the '893 Patent prosecution, as discussed in detail in Section IV below, where GenomeDx's patent counsel made arguments to the examiner specifically distinguishing the claimed "random-forest classifier" from the SVM classifier disclosed in the Nakamura prior art reference.
E. Voluntary vs. Required Amendments
A further doctrinal nuance concerns the distinction between voluntary amendments and amendments required by an examiner's rejection. Under Festo and its progeny, amendments that are not made in direct response to an examiner rejection — i.e., voluntary narrowing amendments — can also give rise to PHE. See Honeywell Int'l Inc. v. Hamilton Sundstrand Corp., 370 F.3d 1131, 1143 (Fed. Cir. 2004) (en banc). The key inquiry is whether the amendment narrowed the claim scope, not whether it was compelled by the examiner.
That said, the Federal Circuit has recognized an important distinction between "narrowing" amendments and "clarifying" amendments. A clarifying amendment — one that does not substantively change the scope of the claim but merely clarifies an ambiguity or addresses a formality issue — does not trigger PHE. See Severin Holdings v. Quillon Corporation, 289 F.3d 801, 813 (Fed. Cir. 2002). Determining whether a given amendment is narrowing or clarifying requires careful analysis of the prosecution record, including the examiner's stated reasons for any rejection and the applicant's stated reasons for the amendment.
This distinction is relevant to our analysis of the '271 Patent, where the amendments were made to overcome a Section 101 rejection. GenomeDx's counsel framed the amendments as adding "specific, concrete" technical steps to transform the claim from an abstract idea to a patent-eligible method — an argument that could be characterized as either narrowing (adding new limitations) or clarifying (specifying what was inherent in the original method). We analyze this question in detail in Section III.D below.
III. Prosecution History of the '271 Patent (U.S. Patent No. 10,438,271)
A. Overview. The '271 Patent is titled "Methods for Predicting Prostate Cancer Outcomes Using Multi-Gene Expression Classifiers." It was filed on June 14, 2016 and issued on October 8, 2019, with an expiration date of June 14, 2036. The named inventors are Dr. Rajan Mehta, Dr. Carolyn Xu, and Dr. David Petrossian, all of whom were employed by GenomeDx at the time of filing. The '271 Patent is the parent application in the family that includes the '893 and '647 Patents.
GenomeDx asserts Claims 1, 3, 5, 7, 12, and 18 of the '271 Patent in this litigation. Claim 1 is the sole independent claim among the asserted claims and recites a method for generating a prognostic risk score for prostate cancer recurrence using a trained machine-learning classifier applied to normalized gene-expression data obtained from a biological sample. The dependent claims add further limitations relating to specific gene panels, specific normalization techniques, output thresholds, and reporting formats.
The prosecution history of the '271 Patent is relatively straightforward, involving two office actions before issuance. Importantly for this analysis, no prior-art rejections under Sections 102 or 103 were sustained during prosecution, and the sole substantive rejection was a Section 101 patent-eligibility rejection.
B. First Office Action — Section 101 Rejection
The Patent and Trademark Office ("PTO") issued a first office action on November 7, 2017, rejecting all then-pending claims of the '271 Patent under 35 U.S.C. § 101 as directed to patent-ineligible subject matter. The examiner determined that the claims were directed to abstract ideas and natural phenomena — specifically, the natural correlation between gene expression levels and cancer prognosis, and the mathematical algorithms used to analyze those expression levels.
The examiner cited the Supreme Court's decisions in Mayo Collaborative Services v. Prometheus Laboratories, Inc., 566 U.S. 66 (2012) and Association for Molecular Pathology v. Myriad Genetics, Inc., 569 U.S. 576 (2013), as well as the PTO's then-current Guidance on patent eligibility for diagnostic and personalized-medicine claims. Under the Alice/Mayo two-step framework, the examiner found at Step 1 that the claims were directed to an abstract idea (applying a mathematical algorithm to data) and a natural phenomenon (the relationship between gene expression and cancer recurrence), and at Step 2 that the claims did not recite an "inventive concept" sufficient to transform them into patent-eligible subject matter, because the recited steps were described at a high level of generality.
Original Claim 1 as filed broadly recited:
A method for prognosing prostate cancer recurrence, comprising:
(a) obtaining a biological sample containing tumor RNA from a patient who has undergone prostatectomy;
(b) measuring expression levels of a panel of genes in said biological sample;
(c) applying a trained machine-learning classifier to the measured expression levels to generate a risk score indicative of the probability of cancer recurrence; and
(d) classifying the patient into a risk category based on the risk score.
The examiner found this claim to be drawn at too high a level of generality, noting that steps (a) and (b) amounted to conventional data-gathering steps and that steps (c) and (d) amounted to applying a mathematical algorithm to the gathered data.
It is noteworthy that no prior-art rejections under Sections 102 or 103 were sustained at any point during prosecution of the '271 Patent. The examiner identified no references that anticipated or rendered obvious the claimed method. This fact simplifies the PHE analysis because the only estoppel concern arises from the Section 101-driven amendments rather than from any prior-art-related narrowing.
C. Applicant's Response — Amendment to Overcome Section 101
On May 14, 2018, GenomeDx's prosecuting counsel filed a response to the first office action that included amendments to Claim 1 and supporting arguments regarding patent eligibility. The key amendments to Claim 1 were as follows:
(a) The step of "obtaining a biological sample containing tumor RNA from a patient who has undergone prostatectomy" was amended to read: "obtaining a biological sample comprising tumor RNA extracted from a formalin-fixed paraffin-embedded (FFPE) tissue sample from a patient who has undergone prostatectomy."
(b) The step of "measuring expression levels of a panel of genes in said biological sample" was amended to read: "measuring expression levels of a panel of at least 17 genes in said biological sample, wherein measuring expression levels comprises performing reference-gene normalization using at least three housekeeping genes to account for variability in RNA quality and quantity."
(c) The step of "applying a trained machine-learning classifier to the measured expression levels to generate a risk score" was not amended — the term "trained machine-learning classifier" was retained without modification.
In the accompanying remarks, GenomeDx's counsel argued that the amended claims recite a "specific, concrete diagnostic improvement that transforms the claim from an abstract idea to a patent-eligible method rooted in a particular technological process." Counsel emphasized that the use of FFPE tissue samples and reference-gene normalization are specific laboratory techniques that address real-world technical challenges in molecular diagnostics — specifically, the degradation of RNA in archived tissue samples and the variability in RNA quality across samples. Counsel argued that these steps are not mere "data-gathering" but represent meaningful technical choices that improve the accuracy and reliability of the diagnostic method.
The examiner was persuaded by these amendments and arguments. The Section 101 rejection was withdrawn in a second office action dated September 21, 2018, which was a Notice of Allowance with minor formality corrections. The '271 Patent issued on October 8, 2019.
D. PHE Analysis — '271 Patent
We now analyze the prosecution history estoppel implications of the amendments made during prosecution of the '271 Patent.
Amendment Identified. Two narrowing amendments to Claim 1 are identified: (1) the addition of the requirement that the biological sample comprises tumor RNA "extracted from a formalin-fixed paraffin-embedded (FFPE) tissue sample"; and (2) the addition of the requirement that measuring expression levels "comprises performing reference-gene normalization using at least three housekeeping genes to account for variability in RNA quality and quantity."
Nature of Amendment. Both amendments are narrowing in nature — they added specific technical limitations that were not present in the original claims. The original Claim 1 broadly recited "obtaining a biological sample containing tumor RNA," which would have encompassed fresh tissue, frozen tissue, blood-based liquid biopsy samples, and other sources in addition to FFPE tissue. The amended claim is limited to FFPE samples. Similarly, the original claim did not specify any particular normalization methodology; the amended claim requires reference-gene normalization using at least three housekeeping genes.
Statutory Basis. The amendments were made to overcome the examiner's Section 101 patent-eligibility rejection, not to distinguish prior art under Sections 102 or 103. Under Festo, however, amendments made to satisfy any requirement of the Patent Act — including Section 101 — create the presumption of prosecution history estoppel. Festo, 535 U.S. at 736-37. Therefore, the Festo presumption applies to these amendments.
Scope of Surrender. The territory surrendered by these amendments is: (1) the territory between the original broad limitation of "obtaining a biological sample containing tumor RNA" and the amended limitation of "obtaining ... from a formalin-fixed paraffin-embedded (FFPE) tissue sample" — i.e., non-FFPE biological samples such as fresh tissue, frozen tissue, and liquid biopsy samples; and (2) the territory between the original broad limitation of "measuring expression levels" (without specifying a normalization methodology) and the amended limitation of "measuring expression levels ... using reference-gene normalization" — i.e., normalization methods other than reference-gene normalization using at least three housekeeping genes, or no normalization at all.
Impact on DOE Against Helios. The practical impact of this PHE on GenomeDx's case against Helios is limited. Our investigation of Helios's ProstaSignal™ product confirms that ProstaSignal™ does use FFPE tissue samples as its input material and does employ reference-gene normalization using housekeeping genes (specifically, ProstaSignal™ uses five housekeeping genes for normalization, exceeding the claim's minimum of three). Therefore, GenomeDx is asserting literal infringement of these amended limitations — not DOE. Because GenomeDx does not need to extend these claim limitations via DOE to reach Helios's product, the PHE arising from the Section 101 amendments has no practical effect on the infringement case.
Key Risk Scenario. There is one scenario in which the PHE from the '271 Patent prosecution could become relevant. If Helios argues at claim construction that "reference-gene normalization" as used in the '271 Patent has a specific technical meaning that is narrower than GenomeDx contends — for example, if Helios argues that the term requires a particular mathematical normalization algorithm (such as the delta-delta-Ct method) and excludes Helios's particular normalization methodology (which we understand uses a regression-based approach) — then GenomeDx might need to invoke DOE to argue that Helios's normalization method is equivalent to the claimed "reference-gene normalization." In that scenario, the PHE from the Section 101 amendment could potentially foreclose DOE for this limitation. However, GenomeDx's primary position, supported by the claim construction analysis in the Preliminary Claim Construction Memorandum, is that "reference-gene normalization" is a broad term that encompasses any normalization method using housekeeping reference genes, without limitation to a particular mathematical approach. We consider GenomeDx's literal-infringement position on this limitation to be strong.
"Trained Machine-Learning Classifier" — No PHE. Critically, the term "trained machine-learning classifier" was present in the original Claim 1 as filed and was not amended at any point during prosecution of the '271 Patent. It was carried through to the issued claim without modification. As a result, no prosecution history estoppel of any kind — neither amendment-based nor argument-based — attaches to this term for the '271 Patent. This is significant because GenomeDx's primary theory of literal infringement under the '271 Patent rests on the contention that "trained machine-learning classifier" encompasses both random-forest classifiers and gradient-boosted decision tree classifiers, the latter being the type used by Helios's ProstaSignal™ product. The absence of any PHE on this term preserves GenomeDx's full DOE flexibility should it prove necessary, though our expectation is that literal infringement will be established on this element.
Conclusion for the '271 Patent. PHE risk is LOW for the DOE theories relevant to this case. The narrowing amendments were directed to sample-preparation and normalization limitations that Helios's product literally satisfies, and the critical "trained machine-learning classifier" limitation was never amended during prosecution.
IV. Prosecution History of the '893 Patent (U.S. Patent No. 11,204,893)
A. Overview. The '893 Patent is titled "Genomic Classifier Systems and Methods for Post-Surgical Cancer Prognosis." It was filed on March 22, 2019 as a continuation of the '271 Patent, sharing the same priority chain and the same ultimate expiration date of June 14, 2036. The named inventors are Dr. Rajan Mehta and Dr. Carolyn Xu. The '893 Patent issued on December 21, 2021.
GenomeDx asserts Claims 1, 2, 9, 14, and 20 of the '893 Patent. Claim 1 is the independent claim and specifically recites "inputting normalized expression values into a random-forest classifier trained on a cohort of at least 500 patient samples" — a significantly narrower limitation than the '271 Patent's "trained machine-learning classifier" language. This narrower claim language is the source of the most significant PHE risk in the case.
The prosecution history of the '893 Patent involved one office action before issuance. Unlike the '271 Patent, the '893 Patent faced a prior-art anticipation rejection under Section 102, which the applicant overcame through argument-based distinction rather than claim amendment.
B. Office Action — Section 102 Rejection Based on Nakamura
The PTO issued an office action on January 14, 2021, rejecting all then-pending claims of the '893 Patent under 35 U.S.C. § 102(a)(1) as anticipated by Nakamura et al., U.S. Patent No. 9,617,340 (titled "Gene Expression Profiling for Cancer Prognosis," issued April 4, 2017) (hereinafter "Nakamura"). Nakamura disclosed a method of gene expression profiling using machine-learning classifiers to generate cancer prognosis scores for urological cancers. The examiner found that Nakamura's disclosure of a machine-learning classifier applied to gene-expression data to generate a cancer prognosis score anticipated the originally-filed claims of the '893 Patent.
Specifically, the examiner identified the following correspondences between Nakamura and the claims of the '893 Patent: (1) Nakamura disclosed obtaining gene expression data from tumor tissue; (2) Nakamura disclosed normalizing gene expression values; (3) Nakamura disclosed inputting normalized expression values into a machine-learning classifier to generate a prognostic score; and (4) Nakamura disclosed classifying the patient based on the output score. The examiner concluded that these correspondences were sufficient to establish anticipation under Section 102.
C. Applicant's Response — Argument and Claim Differentiation
On July 19, 2021, GenomeDx's prosecuting counsel filed a response to the office action. Notably, counsel did not amend the independent claim language. Instead, counsel distinguished Nakamura through argument, relying on claim limitations already present in the as-filed claims of the '893 Patent. The key arguments are set forth below, and verbatim excerpts are provided in Appendix B.
First, counsel emphasized that Claim 1 of the '893 Patent recites "inputting normalized expression values into a random-forest classifier trained on a cohort of at least 500 patient samples" (emphasis added by counsel in the response). Counsel argued that Nakamura's classifier was a support vector machine ("SVM"), not a random-forest classifier, and that these are fundamentally different machine-learning methodologies. Counsel stated:
"Nakamura's classifier is an SVM, which is a single-model discriminative classifier that operates by finding an optimal separating hyperplane in high-dimensional feature space. In contrast, applicant's claimed random-forest classifier is an ensemble learning method that constructs a multitude of independently trained decision trees at training time and outputs the class that is the mode of the classes predicted by the individual trees. The random-forest methodology provides distinct technical advantages not present in Nakamura's SVM approach, including but not limited to: (i) ensemble learning through bootstrap aggregating (bagging) of multiple independently constructed decision trees; (ii) out-of-bag error estimation, which provides an unbiased estimate of generalization error without the need for a separate validation set; and (iii) feature importance ranking derived from the contribution of each gene to the ensemble's predictive accuracy. These are structural and methodological differences, not merely implementation details."
Second, counsel argued that Nakamura did not disclose the minimum cohort size of 500 patient samples for training the classifier, which is a specific technical requirement of the '893 Patent claims. Counsel pointed to the specification of the '893 Patent, which explains that training on a cohort of at least 500 samples improves the statistical power and generalizability of the random-forest model.
The examiner was persuaded by these arguments. In a subsequent communication dated September 8, 2021, the examiner withdrew the Section 102 rejection and allowed the claims. The '893 Patent issued on December 21, 2021.
D. PHE Analysis — '893 Patent (Critical Analysis)
The '893 Patent prosecution history presents the most significant PHE risk in this litigation. We analyze this risk in detail below, addressing amendment-based estoppel, argument-based estoppel, the DOE analysis for the relevant limitation, potential Festo rebuttal arguments, and our overall conclusions.
i. Amendment-Based Estoppel. Strictly speaking, amendment-based estoppel under Festo does not apply to the "random-forest classifier" term in the '893 Patent. The claim language reciting "a random-forest classifier" was present in the originally-filed claims of the '893 Patent as filed on March 22, 2019. No formal amendment to Claim 1 was made at any point during prosecution — counsel distinguished Nakamura through argument alone. Because there was no narrowing amendment to the claim language, the Festo presumption of surrender does not technically arise with respect to the "random-forest classifier" limitation.
No other narrowing amendments were identified in the prosecution file wrapper of the '893 Patent. The only amendments made during prosecution were minor formatting corrections that did not change claim scope.
However, we note a potential argument that Helios could raise: because the '893 Patent is a continuation of the '271 Patent, and because the '271 Patent claims the broader term "trained machine-learning classifier" while the '893 Patent claims the narrower term "random-forest classifier," Helios could argue that GenomeDx's deliberate choice to claim only "random-forest classifier" in the continuation application constitutes a constructive narrowing — analogous to a formal amendment — that triggers amendment-based estoppel. The Federal Circuit has not squarely addressed whether the choice of narrower claim language in a continuation application, without any prosecution-driven amendment, gives rise to Festo-style amendment-based estoppel. The safer analytical approach is to treat this as a potential issue but to focus our analysis on argument-based estoppel, which presents the more immediate risk as discussed below.
ii. Argument-Based Estoppel — The Primary Risk. Argument-based estoppel is the most significant PHE concern for the '893 Patent. As detailed in Section IV.C above, GenomeDx's prosecuting counsel made affirmative arguments to the examiner specifically distinguishing the claimed "random-forest classifier" from Nakamura's SVM classifier. These arguments emphasized the specific technical characteristics of random-forest classifiers that differentiate them from SVMs — namely, ensemble learning through bagging, out-of-bag error estimation, and feature importance ranking.
Under argument-based estoppel principles, GenomeDx may be estopped from asserting that Helios's gradient-boosted decision tree classifier is "equivalent" to the claimed "random-forest classifier" if the prosecution arguments are read as disclaiming non-random-forest machine-learning methods generally. The critical question is the breadth of the disclaimer: did GenomeDx disclaim all non-random-forest classifiers, or did it disclaim only SVMs specifically?
Favorable Reading for GenomeDx. A fair reading of the prosecution arguments supports the position that GenomeDx's disclaimer was narrow and directed specifically to SVMs. The applicant's arguments were made in response to a Section 102 rejection based on Nakamura, which disclosed an SVM classifier. The arguments were specifically tailored to distinguish Nakamura's SVM — the applicant did not make sweeping statements disclaiming all machine-learning classifiers other than random forests. The applicant did not state, for example, that "the claimed random-forest classifier excludes all other types of machine-learning classifiers" or that "only random-forest classifiers are within the scope of the claims." Rather, the applicant argued that "Nakamura's classifier is an SVM," which is "fundamentally different" from a random-forest classifier, and then described the specific attributes that distinguish random forests from SVMs.
Importantly, the technical distinctions emphasized by the applicant — ensemble learning, out-of-bag error estimation, and feature importance ranking — are characteristics that distinguish random forests from SVMs specifically. Gradient-boosted decision trees share several of these characteristics with random forests: both are ensemble methods that use decision trees as base learners, both produce feature importance rankings, and both are used extensively in genomics for the same types of predictive tasks. The prosecution arguments were targeted at the structural differences between random forests and SVMs, not at the differences between random forests and other tree-based ensemble methods.
Unfavorable Reading (Helios's Likely Position). Helios will likely argue for a broader reading of the prosecution disclaimer. Helios's counsel at Ridgecrest Durneval LLP (lead partner Marcus Pellegrini) will contend that GenomeDx's prosecution arguments emphasized specific attributes of random-forest classifiers — ensemble learning through bagging, out-of-bag error estimation, and feature importance ranking — as the basis for patentability, thereby limiting the claims to classifiers that possess these exact attributes. Helios will argue that gradient-boosted decision trees, while tree-based, lack several of these specific attributes: gradient boosting uses sequential training rather than bagging; gradient-boosted trees do not employ out-of-bag error estimation in the same manner as random forests (because each tree is trained on the full dataset with a loss-gradient weighting, not on a bootstrap sample); and the mechanism for deriving feature importance differs between the two methods. Helios will contend that by identifying these specific attributes as the distinguishing features of the claimed invention, GenomeDx effectively disclaimed any classifier that does not possess them.
Assessment. This is a genuinely contested question that does not admit of easy resolution on the current record. The breadth of the argument-based disclaimer will likely be a focal point of briefing at the Markman hearing, currently scheduled for September 18, 2026. Both sides can marshal reasonable arguments, and the outcome may depend on how the court interprets the specificity of GenomeDx's prosecution statements and whether the court views gradient-boosted decision trees as falling within or outside the scope of those statements.
We assess the argument-based estoppel theory as Helios's strongest defense against the DOE claim under the '893 Patent. However, we note that argument-based estoppel requires a showing of "clear and unmistakable" disclaimer, and GenomeDx can credibly argue that its prosecution statements were specifically directed to distinguishing SVMs, not to disclaiming all non-random-forest classifiers. The question of whether the argument-based estoppel extends to gradient-boosted decision trees — which are much more closely related to random forests than SVMs are — presents a genuinely close issue.
iii. DOE Analysis for Random Forest vs. Gradient-Boosted Decision Tree. Assuming that the "random-forest classifier" limitation is construed to exclude gradient-boosted decision trees under a literal-infringement theory, we analyze the viability of asserting DOE for this limitation.
Function-Way-Result Test:
Function: Both random forests and gradient-boosted decision trees function as ensemble tree-based classifiers that take gene-expression input data and generate a predictive risk score correlated with clinical outcomes. In the specific context of the claimed invention, both methods serve the function of transforming normalized gene-expression values into a numerical prognostic score for cancer recurrence. This is substantially the same function.
Way: This is the weakest element of the function-way-result analysis for GenomeDx. Random forests use bagging (bootstrap aggregating) to train an ensemble of independently constructed decision trees, with the final prediction determined by majority voting (for classification) or averaging (for regression) across all trees. Gradient-boosted decision trees use sequential training of shallow decision trees where each successive tree is trained to correct the errors (residuals) of the prior ensemble, with the final prediction determined by a weighted sum of all trees' outputs. These are meaningfully different ensemble-construction methodologies. Random forests emphasize variance reduction through averaging of independently trained high-variance models; gradient boosting emphasizes bias reduction through iterative refinement. Helios will argue that these differences are substantial, not insubstantial.
However, GenomeDx can argue that the "way" should be assessed at the appropriate level of abstraction — which, in the context of the claimed invention, is the use of an ensemble of decision trees to produce a gene-expression-based prognostic score. At this level, both methods construct ensembles of decision trees and aggregate their predictions to produce a single output score. The differences in how the ensemble is constructed (bagging vs. boosting) are implementation details that do not change the fundamental "way" in which the classifier operates — by aggregating multiple tree-based predictors.
Result: Both methods produce a numerical prognostic score correlated with metastasis risk, which is then used to classify the patient into a risk category. The output format, dynamic range, and clinical utility of the scores are substantially the same. This is substantially the same result.
Insubstantial Differences Test (Alternative). Under this alternative formulation, GenomeDx can argue that the differences between gradient-boosted decision trees and random forests are insubstantial in the context of the claimed invention. Both are ensemble tree-based methods that are widely recognized in the machine-learning and genomics literature as closely related variants. Multiple published papers and textbooks discuss them in conjunction as "ensemble tree methods" or "tree-based ensemble approaches." Both were well-known at the time of filing and are routinely used for the same types of gene-expression classification tasks. GenomeDx can point to published comparative studies showing that random forests and gradient-boosted decision trees achieve comparable performance on gene-expression datasets, with neither consistently outperforming the other.
Assessment. The DOE argument for the '893 Patent is arguable but carries meaningful risk, particularly given the prosecution arguments discussed above. The "way" element of the function-way-result test is the weakest link, and Helios will focus its defense on the algorithmic differences between bagging and boosting. GenomeDx's best arguments on DOE are: (1) the insubstantial-differences test, which focuses on the overall technical relationship between the two methods rather than parsing the algorithmic details; (2) expert testimony from Dr. Ogawa establishing that a person of ordinary skill in the art would view gradient-boosted decision trees and random forests as closely related ensemble tree methods; and (3) published literature treating the two methods as interchangeable alternatives for gene-expression classification tasks.
iv. *Festo* Rebuttal Arguments (If Amendment-Based Estoppel Were Applied). Even if a court were to find amendment-based estoppel applicable — for example, by treating the continuation application's deliberate choice of "random-forest classifier" as a constructive narrowing relative to the parent '271 Patent's broader "trained machine-learning classifier" language — GenomeDx has three potential rebuttal arguments under Festo:
(1) Unforeseeability. Gradient-boosted decision tree methods, including the widely-used XGBoost framework (published by Chen and Guestrin in 2016), were well-known in the field at the time the '893 Patent was filed on March 22, 2019. The LightGBM and CatBoost frameworks were also publicly available before the filing date. Accordingly, the gradient-boosted decision tree equivalent was clearly foreseeable at the time of filing. This rebuttal argument is weak and we do not recommend relying on it.
(2) Tangential Relation. The prosecution arguments were directed specifically to distinguishing Nakamura's SVM classifier. The rationale underlying the distinction — that SVMs are single-model discriminative classifiers operating via hyperplane separation, fundamentally different from ensemble tree methods — bears only a tangential relation to gradient-boosted decision trees, which are ensemble tree-based methods sharing the same general architecture as random forests. The reasons for distinguishing an SVM from a random forest (single model vs. ensemble; hyperplane vs. decision trees) do not logically extend to distinguishing a gradient-boosted tree ensemble from a random-forest ensemble, because gradient-boosted trees share the very characteristics that the applicant identified as distinguishing features — ensemble architecture and decision-tree-based prediction. This rebuttal argument is the strongest of the three and we recommend emphasizing it.
(3) Other Reason. GenomeDx could argue that it could not reasonably have been expected to describe every tree-based ensemble variant in its claims, because the field of machine learning was rapidly evolving and new ensemble methods were being developed continuously. However, this argument is undermined by the fact that gradient boosting was a well-established technique by 2019, having been used extensively in data science competitions and published applications for several years prior to the '893 Patent filing. This rebuttal argument is weak and should be used only as a supplementary argument.
Strongest Rebuttal: Tangential Relation. To summarize, if amendment-based estoppel is found to apply, GenomeDx's strongest rebuttal is the tangential-relation prong. The prosecution arguments were specifically aimed at SVMs, and the rationale for distinguishing SVMs (lack of ensemble tree structure, lack of out-of-bag error estimation, fundamentally different model architecture) bears only a tangential relation to gradient-boosted decision trees, which are tree-based ensembles that share far more in common with random forests than SVMs do. A court applying the tangential-relation test should find that the reasons for the applicant's distinction of Nakamura's SVM do not logically compel the exclusion of gradient-boosted decision trees from the scope of equivalents.
v. Conclusion for the '893 Patent. PHE risk is MODERATE TO HIGH for asserting DOE on the "random-forest classifier" limitation of the '893 Patent. Argument-based estoppel is the most significant concern. The applicant's prosecution arguments clearly distinguished the claimed random-forest classifier from Nakamura's SVM on the basis of specific technical attributes, and Helios will argue that these statements disclaim any classifier lacking those exact attributes — including gradient-boosted decision trees. While GenomeDx has a credible response (the arguments were narrowly targeted at SVMs), this is a genuinely contested issue that may be resolved against GenomeDx at the Markman hearing or at trial.
To mitigate this risk, we recommend the following strategic approaches:
(a) GenomeDx should pursue a claim construction of "random-forest classifier" that is broad enough to literally encompass gradient-boosted decision trees — or at least to create sufficient ambiguity that the issue goes to the jury. If the term is construed to include ensemble tree-based classifiers that use decision trees as base learners (regardless of whether the ensemble is constructed via bagging or boosting), the DOE/PHE problem is eliminated entirely. However, this construction faces significant technical and linguistic challenges, because "random forest" has a specific meaning in the machine-learning literature that is generally understood to refer to the bagging-based ensemble method developed by Breiman (2001).
(b) As an alternative, GenomeDx should prepare a robust DOE argument supported by expert testimony from Dr. Ogawa, emphasizing the tangential-relation rebuttal to the Festo presumption and the insubstantial-differences test for equivalence.
(c) GenomeDx should also emphasize the dependent claims of the '893 Patent (Claims 2, 9, 14, and 20), which contain additional limitations that may be met literally by ProstaSignal™. If the independent claim's "random-forest classifier" limitation proves to be an obstacle under DOE/PHE, the dependent claims may provide alternative paths to infringement that reduce the importance of the independent claim's DOE argument. A detailed analysis of the dependent claims' infringement posture is provided in the Preliminary Claim Construction Memorandum.
V. Prosecution History of the '647 Patent (U.S. Patent No. 11,891,647)
A. Overview. The '647 Patent is titled "Improved Methods for Genomic Risk Stratification in Prostate Cancer Using Combined Biomarker and Clinical Variable Models." It was filed on September 9, 2021 and issued on February 6, 2024, with an expiration date of September 9, 2041 (subject to the terminal disclaimer discussed below). The named inventors are Dr. Rajan Mehta, Dr. Kevin Ostrowski, and Dr. Priya Shah.
GenomeDx asserts Claims 1, 4, 6, 11, and 15 of the '647 Patent. Claim 1 is the independent claim and recites a method for genomic risk stratification that combines a genomic risk score derived from at least 20 gene features with at least two clinical covariates selected from the group consisting of Gleason score, PSA level, surgical margin status, and pathological stage. The '647 Patent thus differs from the '271 and '893 Patents in its focus on the combination of genomic and clinical data, rather than on the specific machine-learning classifier used to process genomic data alone.
The prosecution history of the '647 Patent was more complex than the other two patents, involving three office actions before issuance. However, as detailed below, no narrowing amendments were made to the claim language, and the rejections were overcome through evidentiary showing.
B. First and Second Office Actions — Section 103 Rejections
The PTO issued a first office action on April 3, 2022, rejecting all then-pending claims of the '647 Patent under 35 U.S.C. § 103 as obvious over the combination of Nakamura et al. (U.S. Patent No. 9,617,340) and Gagnon (WO 2018/065413, titled "Multi-Gene Prognostic Assays for Urological Cancers," published April 12, 2018). The examiner found that Nakamura disclosed gene-expression-based cancer prognosis using machine-learning classifiers and that Gagnon disclosed methods for combining genomic risk scores with clinical covariates (including Gleason score and PSA level) to generate integrated prognostic models for urological cancers. The examiner concluded that it would have been obvious to a person of ordinary skill in the art to combine the teachings of Nakamura and Gagnon to arrive at the claimed method.
GenomeDx's counsel filed a response on October 5, 2022, arguing that the combination of Nakamura and Gagnon did not render the claims obvious because neither reference, alone or in combination, taught or suggested the specific claimed combination of at least 20 gene features with at least two of the four enumerated clinical covariates, or the specific synergistic improvement in predictive accuracy that results from this combination. The examiner was not fully persuaded and maintained the Section 103 rejection in a second office action dated February 14, 2023, noting that Gagnon's disclosure of combining genomic and clinical data was sufficiently analogous to the claimed method and that the claimed minimum of 20 gene features was within the routine optimization capabilities of a skilled artisan.
C. Applicant's Response — Declaration Under 37 C.F.R. § 1.132
In response to the maintained Section 103 rejection, GenomeDx's counsel filed a supplemental response on August 28, 2023 that included a declaration under 37 C.F.R. § 1.132 from Dr. Rajan Mehta, the lead inventor. The Mehta declaration presented detailed experimental data establishing that the specific combination recited in the claims — at least 20 gene features combined with at least two clinical covariates selected from Gleason score, PSA level, surgical margin status, and pathological stage — produces unexpected results compared to genomic-only models or clinical-only models.
Specifically, the Mehta declaration presented comparative performance data from a validation cohort of 1,247 patients, showing that the combined genomic-clinical model achieved an area under the receiver operating characteristic curve ("AUC") of 0.87, compared to 0.78 for the genomic-only model (using the same 22-gene panel without clinical covariates) and 0.72 for the clinical-only model (using Gleason score, PSA, margin status, and pathological stage without genomic features). The declaration explained that the improvement from 0.78 to 0.87 was statistically significant (p < 0.001) and clinically meaningful, representing a qualitative improvement in the ability to discriminate patients at high risk of metastasis from those at low risk.
In the accompanying remarks, counsel argued that this synergistic improvement — the combined model significantly outperforming the sum of its parts — constituted evidence of unexpected results that rebutted the prima facie case of obviousness. Counsel further argued that neither Nakamura nor Gagnon, alone or in combination, taught or suggested that combining genomic features with the specific enumerated clinical covariates would produce this degree of synergistic improvement.
Critically for the PHE analysis, no narrowing amendments to the independent claim language were made at any point during the prosecution of the '647 Patent. The applicant overcame the Section 103 rejection entirely through the evidentiary showing in the Mehta declaration, without amending the claim language.
D. Terminal Disclaimer
During prosecution, GenomeDx filed a terminal disclaimer over the '271 and '893 Patents. This was required due to common inventorship and overlapping subject matter between the patents in the same family. The terminal disclaimer provides that the '647 Patent will not be enforced after the expiration of the '271 and '893 Patents (June 14, 2036) if ownership does not remain common.
The terminal disclaimer does not create prosecution history estoppel. It addresses only the double-patenting issue and limits the enforceable life of the '647 Patent. Because the '647 Patent would otherwise expire on September 9, 2041, the terminal disclaimer effectively shortens its enforceable term to June 14, 2036 (the expiration date of the '271 and '893 Patents) unless ownership remains common. This is a standard double-patenting resolution and has no bearing on the PHE analysis.
E. Third Office Action — Allowance
Following submission of the Mehta declaration, the terminal disclaimer, and the accompanying arguments, the examiner issued a Notice of Allowance on October 30, 2023. The examiner found that the Mehta declaration's evidence of unexpected synergistic results was persuasive in overcoming the Section 103 rejection. The '647 Patent issued on February 6, 2024.
F. PHE Analysis — '647 Patent
The PHE analysis for the '647 Patent is straightforward and favorable for GenomeDx.
No Narrowing Amendments. Because no narrowing amendments were made to the claim language during prosecution, the Festo presumption of surrender does not arise. The applicant overcame the Section 103 rejection through evidence (the Mehta declaration) rather than by amending the claims. The absence of any narrowing amendment eliminates the primary source of PHE risk.
Argument-Based Estoppel Risk. The applicant's arguments during prosecution emphasized the specific combination of genomic and clinical variables as producing unexpected synergistic results. This could theoretically give rise to argument-based estoppel limiting the scope of the "combined" model limitation. However, any such estoppel would not materially affect GenomeDx's infringement case because Helios's ProstaSignal™ product literally satisfies the combined-model limitation. ProstaSignal™ combines its genomic ProstaScore (derived from a 24-gene panel, exceeding the claimed minimum of 20 gene features) with Gleason grade group and pre-operative PSA level — two of the four enumerated clinical covariates. The claim requires "at least two clinical covariates selected from the group consisting of Gleason score, PSA level, surgical margin status, and pathological stage," and Helios's product uses two of these four covariates. Literal infringement is established without resort to DOE.
No DOE Needed. Because the '647 Patent's key limitations appear to be literally infringed by ProstaSignal™, there is no need to invoke the doctrine of equivalents for any limitation of the asserted claims. The PHE analysis is therefore largely academic for the '647 Patent — even if some form of estoppel were found to exist, it would have no practical impact on the infringement case because DOE is not necessary.
Conclusion for the '647 Patent. PHE risk is MINIMAL. No narrowing amendments were made during prosecution, literal infringement is alleged for all limitations, and DOE is not needed for the infringement theory. The '647 Patent provides GenomeDx with a clean, PHE-free path to infringement.
VI. Cross-Patent Analysis and Strategic Implications
A. Summary Table of PHE Risk by Patent
The following table summarizes the PHE risk assessment for each patent-in-suit:
B. Interplay Between the '271 and '893 Patents
The relationship between the '271 and '893 Patents is a critical strategic consideration that warrants careful analysis. The two patents share the same priority chain (the '893 Patent is a continuation of the '271 Patent) and the same expiration date (June 14, 2036). However, they differ in a crucial respect: the '271 Patent claims the broader term "trained machine-learning classifier" in Claim 1, while the '893 Patent claims the narrower term "random-forest classifier" in Claim 1.
This difference has significant implications for the infringement case. If the court adopts GenomeDx's proposed construction of "trained machine-learning classifier" in the '271 Patent — i.e., that the term encompasses any machine-learning classifier that has been trained on data, including gradient-boosted decision tree classifiers — then the '271 Patent provides a direct path to literal infringement of Helios's ProstaSignal™ product without any need to invoke DOE. In this scenario, the '893 Patent's narrower claims become less critical to the overall infringement case, providing additional leverage and redundancy but not serving as the primary basis for liability.
Conversely, if the court construes "trained machine-learning classifier" in the '271 Patent narrowly — for example, by limiting the term to the specific types of classifiers disclosed in the specification (random forests and logistic regression) — then the '271 Patent's literal-infringement theory may be weakened, and the '893 Patent's DOE argument becomes more important. But paradoxically, this is also the scenario in which the '893 Patent's DOE theory is most vulnerable to PHE, because a narrow construction of "trained machine-learning classifier" in the '271 Patent would suggest that the patent family's claims were deliberately drafted to cover specific classifier types rather than the entire field of machine learning.
The practical implication is that GenomeDx's claim construction strategy for the '271 Patent is the single most important determinant of the overall litigation posture. If GenomeDx prevails on a broad construction of "trained machine-learning classifier," the PHE risk for the '893 Patent becomes a secondary concern. If GenomeDx does not prevail on this construction, the case becomes more dependent on the '893 Patent's DOE theory and the '647 Patent's lite

PRIVILEGED & CONFIDENTIAL — ATTORNEY WORK PRODUCT

MEMORANDUM

MATTER NO.: 1013-00005

Record

Parties

Dates

Amounts