PRIVILEGED AND CONFIDENTIAL — ATTORNEY WORK PRODUCT
This memorandum is protected by the attorney-client privilege and the work-product doctrine. It is prepared in anticipation of litigation and for the purpose of providing legal advice. It should not be disclosed to any person outside of the attorney-client relationship without prior authorization from Calderwood & Harkness LLP.
CALDERWOOD & HARKNESS LLP 555 Mission Street, Suite 3100 San Francisco, California 94105 Telephone: (415) 892-4000 Facsimile: (415) 892-4001
MEMORANDUM
TO: Catherine Bloom, Responsible Partner, intellectual Property
FROM: Peter Vassiliev, Lead Associate, intellectual Property Wendy Schultz, Working Associate, intellectual Property
DATE: August 11, 2025
RE: Preliminary Claim Construction Analysis — Genome Dx Therapeutics Inc. v. Helios Precision Diagnostics, Inc., Case No. 3:25-cv-01847-JWL (N.D. Cal.)
SUBJECT: Analysis of Key Disputed Claim Terms Across U.S. Patent Nos. 10,438,271; 11,204,893; and 11,891,647; Proposed Constructions and Anticipated Defense Positions Following August 4, 2025 Exchange of Proposed Claim Construction Terms
I. EXECUTIVE SUMMARY
This memorandum provides an internal analysis of the claim construction issues arising in the above-captioned patent infringement action, prepared in advance of the Markman hearing currently scheduled for September 18, 2026, before the Honorable Judge Jessica W. Langston in the Northern District of California. The purpose of this analysis is to guide our litigation strategy, inform expert preparation, and identify areas requiring further factual development through discovery.
On August 4, 2025, the parties exchanged their respective proposed claim construction terms and proposed constructions pursuant to the Scheduling Order issued by Judge Langston at the Initial Case Management Conference held on June 18, 2025, consistent with the requirements of the Northern District of California Patent Local Rules. This exchange has now crystallized the claim construction disputes that will form the core of the Markman proceedings.
Based on our review of the exchanged proposals, the intrinsic record for each of the three asserted patents, and the relevant case law, we have identified five key disputed claim terms that will likely control both the infringement and validity analyses in this case:
1. "trained machine-learning classifier" — '271 Patent, Claim 1
2. "random-forest classifier" — '893 Patent, Claim 1
3. "panel of at least 22 genes selected from a predefined gene set" — '271 Patent, Claim 1
4. "risk category" — '271 Patent, Claim 1; '647 Patent, Claim 1
5. "composite risk metric" — '647 Patent, Claim 1
Our preliminary assessment is that GenomeDx's proposed constructions are well-supported by the intrinsic record — including the claim language, the patent specifications, and the prosecution histories — as well as by ordinary meaning principles as understood by a person of ordinary skill in the art. Helios's anticipated constructions, as revealed through the August 4 exchange, appear designed to narrow claim scope in a manner calculated to exclude the ProstaSignal™ test's gradient-boosted decision tree architecture and 24-gene panel. Specifically, Helios seeks to limit "trained machine-learning classifier" in the '271 Patent to only random-forest and support vector machine classifiers, and to construe "random-forest classifier" in the '893 Patent to exclude gradient-boosted decision trees. These narrowing constructions are, in our view, inconsistent with the patent specifications and prosecution histories, although certain terms present greater risk than others.
As a strategic matter, we note that the claim construction of "random-forest classifier" in the '893 Patent represents the most critical battleground in the claim construction proceedings. If the court construes this term narrowly to exclude gradient-boosted decision trees — a construction that carries moderate-to-high risk — GenomeDx will be forced to rely on the doctrine of equivalents for its '893 Patent infringement theory. Such reliance raises prosecution history estoppel concerns arising from the applicant's distinction of the Nakamura reference during prosecution, which are addressed separately in a forthcoming Prosecution History Estoppel Analysis Memorandum that we are preparing in parallel.
However, even in the event of an adverse construction on the '893 Patent term, GenomeDx retains strong alternative infringement theories under the '271 Patent (which uses the broader "trained machine-learning classifier" term) and the '647 Patent (which uses "a classifier comprising at least 20 gene features" without specifying a random-forest algorithm), provided those terms receive the constructions we propose.
II. BACKGROUND
A. The Patents-in-Suit
GenomeDx has asserted three related patents in this action, all arising from a common research program directed to genomic classifier methods for prostate cancer prognosis. The patents are summarized below.
1. U.S. Patent No. 10,438,271 ("the '271 Patent")
Title: "Methods for Predicting Prostate Cancer Outcomes Using Multi-Gene Expression Classifiers"
Filing Date: June 14, 2016
Issue Date: October 8, 2019
Expiration Date: June 14, 2036
Named Inventors: Dr. Rajan Mehta, Dr. Carolyn Xu, Dr. David Petrossian
Asserted Claims: Claims 1, 3, 5, 7, 12, and 18
The '271 Patent is the parent patent in the asserted patent family and claims methods for predicting prostate cancer outcomes using multi-gene expression classifiers. Key independent Claim 1 recites a method comprising: (a) obtaining a biological sample containing tumor RNA from a prostate cancer patient; (b) measuring expression levels of a panel of at least 22 genes selected from a predefined gene set; (c) applying a trained machine-learning classifier to generate a risk score from the measured expression levels; and (d) classifying the patient into a risk category based on the risk score. The dependent claims add specificity regarding particular gene subsets, classifier training parameters, and risk-category thresholds, but do not further narrow the type of machine-learning classifier employed.
The '271 Patent specification describes range machine-learning approaches as suitable for the claimed classifier, including random forests, support vector machines, and neural networks, among others. The specification discusses these approaches as exemplary embodiments and does not restrict the invention to any single classifier type.
2. U.S. Patent No. 11,204,893 ("the '893 Patent")
Title: "Genomic Classifier Systems and Methods for Post-Surgical Cancer Prognosis"
Filing Date: March 22, 2019 (continuation of the '271 Patent family)
Issue Date: December 21, 2021
Expiration Date: June 14, 2036
Named Inventors: Dr. Rajan Mehta, Dr. Carolyn Xu
Asserted Claims: Claims 1, 2, 9, 14, and 20
The '893 Patent is a continuation of the '271 Patent and claims methods specifically directed to post-surgical cancer prognosis using a random-forest classifier. Key independent Claim 1 recites a method comprising: (a) normalizing gene-expression data from a formalin-fixed paraffin-embedded (FFPE) prostate tissue sample using a reference gene panel; (b) inputting normalized expression values into a random-forest classifier trained on a cohort of at least 500 patient samples; and (c) outputting a numerical prognostic score correlated with metastasis risk within 5 years post-surgery.
The '893 Patent reflects a deliberate narrowing from the broader '271 Patent claims by specifying the random-forest classifier methodology and a minimum training cohort size. This narrowing is significant for claim differentiation purposes, as discussed further below. The '893 Patent specification describes random forests as "ensemble methods that construct multiple decision trees using random subsets of features and training samples" and provides detailed descriptions of the training process, feature-selection methodology, and output aggregation.
3. U.S. Patent No. 11,891,647 ("the '647 Patent")
Title: "Improved Methods for Genomic Risk Stratification in Prostate Cancer Using Combined Biomarker and Clinical Variable Models"
Filing Date: September 9, 2021
Issue Date: February 6, 2024
Expiration Date: September 9, 2041
Named Inventors: Dr. Rajan Mehta, Dr. Kevin Ostrowski, Dr. Priya Shah
Asserted Claims: Claims 1, 4, 6, 11, and 15
The '647 Patent is the most recently issued patent in the family and claims improved methods for genomic risk stratification that integrate genomic data with clinical variables. Key independent Claim 1 recites a method comprising: (a) performing RNA extraction and expression profiling on a prostate tumor specimen; (b) computing a genomic risk score using a classifier comprising at least 20 gene features; (c) combining the genomic risk score with at least two clinical covariates selected from the group consisting of Gleason score, PSA level, surgical margin status, and pathological stage; and (d) generating a composite risk metric and transmitting it to a healthcare provider.
The '647 Patent is notable in two respects relevant to claim construction. First, it uses the broader term "a classifier comprising at least 20 gene features" rather than specifying a random-forest algorithm, thereby avoiding the narrow construction risk that affects the '893 Patent. Second, it introduces the "composite risk metric" concept, which integrates genomic and clinical data into a unified prognostic assessment. A terminal disclaimer was filed over the '271 and '893 Patents during prosecution of the '647 Patent.
B. The Accused Product
Helios Precision Diagnostics, Inc. ("Helios") commercially launched its ProstaSignal™ Genomic Classifier on October 7, 2024. ProstaSignal™ is a 24-gene expression-based prognostic test designed for prostate cancer patients following radical prostatectomy. The test is performed at Helios's CLIA-certified laboratory located at 9830 Mesa Rim Road, Suite 140, San Diego, California 92121.
The key technical characteristics of the ProstaSignal™ test, as understood from publicly available information and Helios's marketing materials (subject to confirmation through discovery), are as follows:
• Gene Panel: ProstaSignal™ measures the expression levels of 24 genes. Of these 24 genes, 19 overlap with the predefined gene set identified in Table 1 of the '271 Patent specification. The remaining 5 genes do not appear to correspond directly to genes in Table 1, although further analysis is required to determine whether any of these genes are alternative designations, splice variants, or functional equivalents of genes in the predefined set.
• Classifier Architecture: ProstaSignal™ employs a gradient-boosted decision tree classifier, as opposed to the random-forest classifier specifically claimed in the '893 Patent. Both random forests and gradient-boosted decision trees are ensemble-of-trees methods — that is, they construct multiple decision trees and aggregate their outputs to produce a final prediction — but they differ in their training methodology. Random forests use bootstrap aggregation (bagging) with random feature subsets at each split, training each tree independently, whereas gradient-boosted decision trees train trees sequentially, with each successive tree correcting the errors of the preceding trees.
• Output: ProstaSignal™ generates a "ProstaScore" — a numerical value ranging from 0 to 100 — and classifies patients into Low, Intermediate, or High risk categories based on predefined thresholds.
• Clinical Integration: ProstaSignal™ combines the ProstaScore with Gleason grade group and pre-operative PSA level to produce an integrated clinical report for the treating physician. The extent to which this integration involves a mathematical combination of the genomic and clinical data (as opposed to mere juxtaposition of these data points in a report) is a factual question that must be resolved through discovery.
C. Procedural Posture
GenomeDx filed its Complaint for patent infringement on March 28, 2025, asserting infringement of the '271, '893, and '647 Patents. Helios filed its Answer and Counterclaims on May 12, 2025, asserting counterclaims for non-infringement, invalidity under 35 U.S.C. §§ 101, 102, and 103, and unenforceability. GenomeDx filed its Answer to Helios's Counterclaims on May 30, 2025, denying all counterclaims.
The Initial Case Management Conference was held before Judge Langston on June 18, 2025. At the conference, Judge Langston issued a comprehensive Scheduling Order establishing the case timeline, including a Markman hearing date of September 18, 2026. The Scheduling Order incorporates the Northern District of California Patent Local Rules governing claim construction procedures.
On August 4, 2025, the parties exchanged their respective proposed claim construction terms and proposed constructions in accordance with the Patent Local Rules and the Scheduling Order. This exchange forms the basis for the analysis in this memorandum.
On September 15, 2025, Helios filed a Motion to Dismiss Counts II and III of GenomeDx's Complaint under 35 U.S.C. § 101, arguing that certain claims of the '893 and '647 Patents are directed to patent-ineligible subject matter under the Mayo/Alice framework. That motion is pending at the time of this memorandum and is not addressed in detail here, although we note that the outcome of the motion could affect the claim construction landscape. GenomeDx believes the motion should be denied as premature before claim construction.
Defendant Helios is represented by Ridgecrest Durneval LLP. The defense litigation team is led by Marcus Pellegrini (lead partner), with Danielle Forrest (senior associate) and Owen Chu (mid-level associate) serving as the primary working attorneys on the matter.
III. LEGAL FRAMEWORK FOR CLAIM CONSTRUCTION
Claim construction is a question of law for the court. Markman v. Westview Instruments, Inc., 517 U.S. 370, 372 (1996). The words of a patent claim are generally given their ordinary and customary meaning as understood by a person of ordinary skill in the art ("POSITA") at the time of the invention, in light of the specification and prosecution history. Phillips v. AWH Corp., 415 F.3d 1303, 1312–13 (Fed. Cir. 2005) (en banc). The Phillips framework establishes a well-defined hierarchy of evidence for claim construction purposes.
Intrinsic Evidence. The Federal Circuit has consistently held that intrinsic evidence is the most significant source of the legally operative meaning of disputed claim terms. Intrinsic evidence comprises three categories, in descending order of importance:
1. Claim Language. The claims themselves define the scope of the patent right. The analysis begins with the words of the claims, read in the context of the entire claim and in relation to other claims in the same patent. Id. at 1314. The doctrine of claim differentiation creates a rebuttable presumption that different claims have different scope, and that the use of different terms across related claims reflects deliberate drafting choices. Comark Commc'ns, Inc. v. Harris Corp., 156 F.3d 1182, 1187 (Fed. Cir. 1998).
2. Specification. The specification, including the written description and figures, is "the single best guide to the meaning of a disputed term." Phillips, 415 F.3d at 1315 (quoting Ridgeway & Co. v. Conceptronic, Inc., 90 F.3d 1576, 1582 (Fed. Cir. 1996)). The specification may reveal a special definition given to a claim term or may clarify the meaning of a term in the context of the invention. However, the Federal Circuit has repeatedly cautioned against importing limitations from the specification into the claims. Id. at 1323. The fact that the specification describes a preferred embodiment does not limit the claims to that embodiment. Teleflex, Inc. v. Ficosa N. Am. Corp., 299 F.3d 1313, 1327 (Fed. Cir. 2002).
3. Prosecution History. The prosecution history provides evidence of how the patent holder and the Patent Office understood the patent during the examination process. Phillips, 415 F.3d at 1317. Statements made during prosecution, including claim amendments and arguments distinguishing prior art, can limit the scope of the claims through prosecution history disclaimer or prosecution history estoppel. The prosecution history is relevant not only to the patent being construed but also to related patents in the same family. Kestrel Partners LLC v. Multi-Tech Sys., Inc., 357 F.3d 1340, 1349 (Fed. Cir. 2004).
Extrinsic Evidence. Extrinsic evidence, including dictionaries, treatises, technical reference works, and expert testimony, may be used to assist the court in understanding the underlying technology and the meaning of terms in the relevant art. Phillips, 415 F.3d at 1318. However, extrinsic evidence may not be used to contradict or override the meaning of claim terms established by the intrinsic evidence. Id. at 1319. Expert testimony is particularly useful in establishing the understanding of a POSITA and in explaining complex technology to the court.
Means-Plus-Function Construction. Under 35 U.S.C. § 112(f), a claim limitation expressed in "means for" language is construed to cover the corresponding structure described in the specification and equivalents thereof. A rebuttable presumption arises that a claim term that does not use "means for" language is not a means-plus-function limitation. Ostrander Holdings, 792 F.3d 1339, 1348 (Fed. Cir. 2015) (en banc in relevant part). None of the five disputed claim terms in this case uses "means for" language, and we do not anticipate that Helios will invoke § 112(f) with respect to any of these terms. We note, however, that if Helios were to argue that any of the classifier-related terms are means-plus-function limitations, the analysis would change significantly, as construction would be limited to the structures disclosed in the specification and their equivalents.
Definition of POSITA. For purposes of construing the claims of the '271, '893, and '647 Patents, we propose that a person of ordinary skill in the art would be a person holding at least a Ph.D. in computational biology, bioinformatics, genomics, or a closely related field, with at least two to three years of experience in developing or applying machine-learning classifiers for gene-expression-based cancer diagnostics. Alternatively, a POSITA could be a person holding an M.D./Ph.D. with equivalent research experience in molecular oncology and computational methods. This definition reflects the interdisciplinary nature of the patented technology, which spans molecular biology, genomics, and machine learning. We anticipate that Helios may propose a different POSITA definition — potentially one that emphasizes greater specialization in machine-learning algorithms — and we should be prepared to address any such dispute.
IV. DISPUTED CLAIM TERM ANALYSIS
A. "Trained Machine-Learning Classifier" ('271 Patent, Claim 1)
1. Claim Language in Context
The term "trained machine-learning classifier" appears in the context of the following limitation of Claim 1 of the '271 Patent:
"applying a trained machine-learning classifier to generate a risk score"
This term appears in all asserted claims of the '271 Patent — Claims 1, 3, 5, 7, 12, and 18. Claim 1 is the key independent claim; the dependent claims add specificity regarding particular gene subsets, classifier training parameters, and risk-category thresholds, but critically, none of the dependent claims narrows the type of machine-learning classifier employed. For example, dependent Claim 3 adds a requirement that the classifier be "trained on a dataset comprising expression profiles from at least 200 patient samples," and dependent Claim 7 recites specific risk-category thresholds, but neither specifies a particular classifier algorithm.
The use of the broad genus term "machine-learning classifier" — as opposed to a narrower species term such as "random-forest classifier" or "support vector machine classifier" — is a deliberate drafting choice that carries significant claim construction implications, particularly when viewed in contrast to the '893 Patent's use of the narrower "random-forest classifier" term.
2. GenomeDx's Proposed Construction
GenomeDx proposes the following construction for "trained machine-learning classifier":
"A computational model that has been trained on labeled data using any supervised or semi-supervised machine-learning algorithm — including but not limited to random forests, gradient-boosted decision trees, support vector machines, neural networks, and logistic regression — to generate a numerical output (risk score) from gene-expression input data."
This proposed construction is supported by the following intrinsic evidence:
Specification Support. The '271 Patent specification describes the classifier in generic terms and expressly identifies multiple machine-learning approaches as suitable embodiments. At column 8, lines 15–42, the specification states:
"The trained machine-learning classifier may be implemented using any of range supervised or semi-supervised machine-learning algorithms known in the art. In preferred embodiments, the classifier is a random forest classifier, a support vector machine, or a neural network. However, persons of ordinary skill in the art will recognize that other machine-learning methods, including but not limited to logistic regression, k-nearest neighbor, gradient boosting machines, and Bayesian classifiers, may also be employed within the scope of the present invention."
This passage is significant because it (a) expressly identifies the classifier as a genus encompassing multiple species, (b) identifies random forests, support vector machines, and neural networks as "preferred" but not exclusive embodiments, and (c) explicitly states that "other machine-learning methods" may be employed, with gradient boosting machines specifically enumerated.
Prosecution History. During prosecution of the '271 Patent, the Examiner issued a rejection under 35 U.S.C. § 101 based on the Mayo/Alice framework, arguing that the claimed method was directed to an abstract idea (categorizing gene-expression data). The applicant overcame this rejection by amending the claims to recite specific technical steps — RNA extraction from FFPE samples and reference-gene normalization — that transformed the abstract concept into a patent-eligible application. Notably, the applicant did not amend the "trained machine-learning classifier" term or narrow it to any specific algorithm type. The absence of any narrowing amendment to this term during prosecution provides strong support for a broad construction.
Critical Application. This construction is of paramount importance because it directly encompasses Helios's gradient-boosted decision tree classifier used in ProstaSignal™. If the court adopts GenomeDx's proposed construction, the '271 Patent provides a direct path to literal infringement on the classifier limitation.
3. Helios's Anticipated Proposed Construction
Based on the August 4, 2025 exchange, Helios proposes the following construction for "trained machine-learning classifier":
"A random-forest or support vector machine classifier that has been trained on a specific labeled training dataset as described in the specification."
We anticipate that Helios will advance the following arguments in support of this narrower construction:
First, Helios will likely argue that the specification's "preferred embodiment" language should be treated as limiting. Helios will point to the specification's detailed description of random-forest and support vector machine classifiers — which spans several columns and includes specific algorithmic parameters — and contrast this with the specification's briefer and more general references to other machine-learning methods. Helios will argue that the detailed disclosure of random forests and SVMs, combined with the absence of any detailed description of gradient-boosted decision trees, reflects the patent holder's intent to limit the claim to these two methods.
Second, Helios may invoke the '893 Patent's use of the narrower "random-forest classifier" term as evidence that the '271 Patent's "trained machine-learning classifier" should be interpreted in the context of the patent family's overall focus on random forests. Under this theory, the broader language in the '271 Patent was merely a drafting convenience, and the patent family as a whole is directed to random-forest-based classifiers.
Third, Helios will likely argue that the specification's statement that "other machine-learning methods ... may also be employed" is aspirational or enabling language rather than a claim scope statement, and that the claim term should be construed in light of what the patent actually teaches, not what it merely suggests.
4. Analysis and Recommendation
GenomeDx's proposed construction is strongly supported by the intrinsic record and should prevail under controlling Federal Circuit precedent.
The '271 Patent specification expressly lists multiple classifier types as examples of the "trained machine-learning classifier" genus. The passage at column 8, lines 15–42, is unambiguous: it identifies random forests, SVMs, and neural networks as "preferred embodiments" and then states that "other machine-learning methods, including but not limited to ... gradient boosting machines ... may also be employed within the scope of the present invention." This language — particularly the phrase "within the scope of the present invention" — is a clear indication that the patent holder intended the claim term to encompass classifier types beyond the preferred embodiments.
The Federal Circuit has repeatedly held that a patentee's use of exemplary or preferred embodiments does not limit claim scope to those embodiments. In Teleflex, Inc. v. Ficosa North America Corp., 299 F.3d 1313, 1327 (Fed. Cir. 2002), the court held that "the claims are not to be limited to the preferred embodiment where the claim language is broader than the preferred embodiment." Similarly, in Thorner v. Sony Computer Entm't Am. LLC, 669 F.3d 1362, 1366 (Fed. Cir. 2012), the court cautioned that "[t]he patentee is entitled to the full scope of his claims, and we will not limit him to his preferred embodiment." The specification here does precisely what Teleflex and Thorner prohibit: it identifies preferred embodiments while also describing the genus more broadly.
Moreover, the presence of the narrower "random-forest classifier" term in the '893 Patent — a continuation of the '271 Patent — actually supports GenomeDx's argument that the '271 Patent's use of the broader "trained machine-learning classifier" term was intentional and meaningful. Under the doctrine of claim differentiation, there is a presumption that different claims — and especially claims across related patents — have different scope. See Comark Commc'ns, 156 F.3d at 1187. If "trained machine-learning classifier" in the '271 Patent meant the same thing as "random-forest classifier" in the '893 Patent, there would have been no reason for the patentee to use different terms. The deliberate choice to use a broader genus term in the '271 Patent and a narrower species term in the '893 Patent reflects a purposeful claim-drafting strategy that should be given effect.
Helios's proposed limitation of the term to "random-forest or support vector machine" classifiers finds no basis in the claim language itself — the claim says "machine-learning classifier," not "random-forest or support vector machine classifier." Nor does the prosecution history support Helios's narrower construction, as no narrowing amendment was ever made to this term. At most, the specification describes random forests and SVMs as preferred embodiments, but this is precisely the type of limitation-importing that Phillips prohibits.
Risk Assessment: We assess a low-to-moderate risk that the court will narrow this term beyond its ordinary and customary meaning. The intrinsic support for GenomeDx's construction is strong. The primary risk factor is the possibility that the court may view the specification's detailed discussion of random forests and SVMs — compared to the cursory reference to gradient boosting — as evidence that the patentee did not fully contemplate gradient-boosted classifiers as within the scope of the claims. However, we believe this argument is foreclosed by the specification's express enumeration of "gradient boosting machines" as a method that "may also be employed within the scope of the present invention."
B. "Random-Forest Classifier" ('893 Patent, Claim 1)
1. Claim Language in Context
The term "random-forest classifier" appears in the context of the following limitation of Claim 1 of the '893 Patent:
"inputting normalized expression values into a random-forest classifier trained on a cohort of at least 500 patient samples"
This term appears in all asserted claims of the '893 Patent — Claims 1, 2, 9, 14, and 20. The '893 Patent is a continuation of the '271 Patent, and the use of the narrower "random-forest classifier" term — as opposed to the broader "trained machine-learning classifier" used in the parent '271 Patent — reflects a deliberate narrowing of claim scope in the continuation application. This narrowing is significant both for claim differentiation purposes (discussed above) and for prosecution history estoppel purposes (discussed in Section VI below).
2. GenomeDx's Proposed Construction
GenomeDx proposes the following construction for "random-forest classifier":
"A classifier based on an ensemble of decision trees constructed using random feature selection and bootstrap aggregation (bagging), or a functionally equivalent ensemble-of-trees method."
This proposed construction is supported by the following rationale:
Specification Support. The '893 Patent specification describes random forests as "ensemble methods that construct multiple decision trees using random subsets of features and training samples" (col. 6, ll. 28–35). This description focuses on the functional attributes of the random-forest methodology — namely, that it is an ensemble of decision trees, that it employs random feature selection, and that it aggregates the outputs of multiple trees — rather than defining the term by reference to a rigid algorithmic specification. GenomeDx contends that a POSITA would understand the specification's description as emphasizing the functional characteristics shared by all ensemble-of-trees methods, including gradient-boosted decision trees.
Technical Argument. Both random forests and gradient-boosted decision trees are ensemble-of-trees methods. They share the following core functional characteristics: (a) both construct multiple decision trees from training data; (b) both use subsets of features in the tree construction process; and (c) both aggregate the outputs of multiple trees to produce a final prediction. The primary difference between the two methods lies in the training process: random forests train each tree independently on a bootstrapped sample of the training data (bagging), whereas gradient-boosted decision trees train trees sequentially, with each successive tree correcting the errors of the preceding trees (boosting). GenomeDx contends that a POSITA would view this difference as a variation in training methodology rather than a fundamental distinction in the nature of the classifier, and would therefore understand "random-forest classifier" to encompass functionally equivalent ensemble-of-trees methods.
Alternative Argument. Even if the court construes "random-forest classifier" to exclude gradient-boosted decision trees as a matter of literal infringement, GenomeDx asserts that Helios's gradient-boosted classifier infringes under the doctrine of equivalents. Under the function-way-result test, the gradient-boosted classifier performs substantially the same function (generating a prognostic score from gene-expression data using an ensemble of decision trees), in substantially the same way (aggregating outputs from multiple trees constructed on subsets of features), to achieve substantially the same result (a numerical score predictive of metastasis risk). The doctrine-of-equivalents theory is discussed further below and in the forthcoming Prosecution History Estoppel Analysis Memorandum.
3. Helios's Anticipated Proposed Construction
Based on the August 4, 2025 exchange, Helios proposes the following construction for "random-forest classifier":
"A classifier that specifically uses the random forest algorithm — i.e., an ensemble of decision trees trained using bootstrap aggregation (bagging) with random feature subsets at each split — and excludes other tree-based ensemble methods such as gradient boosting."
We anticipate that Helios will advance the following arguments in support of this construction:
First, Helios will argue that random forests and gradient-boosted decision trees are distinct, well-recognized algorithms in the machine-learning literature, with different training mechanisms, different theoretical foundations, and different performance characteristics. Helios will cite machine-learning textbooks and treatises establishing that a POSITA would understand "random forest" to refer specifically to the Breiman (2001) algorithm employing bagging and random feature selection, and would not confuse this term with gradient boosting, which traces to Friedman (2001) and employs a fundamentally different sequential-correction training approach.
Second, Helios will point to the '893 Patent specification itself, arguing that the specification's description of random forests as using "random subsets of features and training samples" (i.e., bootstrap samples) is specific to the bagging methodology and does not describe sequential boosting. Helios will contend that the specification distinguishes random forests from other methods and deliberately chose to claim only random forests, not all ensemble-of-trees methods.
Third, and most significantly, Helios will invoke the prosecution history of the '893 Patent. During prosecution, the Examiner rejected the claims under 35 U.S.C. § 102 based on Nakamura et al. (U.S. Patent No. 9,617,340). The applicant overcame this rejection by distinguishing the claimed invention based on the "random-forest classifier training methodology" and the "minimum cohort size" of at least 500 patient samples. Helios will argue that this distinction constitutes a prosecution history disclaimer — an admission that the claim term "random-forest classifier" is limited to the specific random-forest algorithm and excludes other tree-based ensemble methods. Helios will further argue that this distinction gives rise to prosecution history estoppel that bars GenomeDx from asserting the doctrine of equivalents to encompass gradient-boosted decision trees.
4. Analysis and Recommendation
This is the most critical and highest-risk term for GenomeDx in the claim construction proceedings.
Literal Infringement Analysis. The term "random-forest classifier" has a recognized meaning in the art. The random forest algorithm, as described by Breiman (2001), is a specific ensemble method that combines bootstrap aggregation (bagging) with random feature selection at each tree split. Gradient-boosted decision trees, as described by Friedman (2001), employ a fundamentally different training methodology — sequential boosting rather than bagging. A POSITA would likely recognize these as distinct algorithms, notwithstanding the fact that both are ensemble-of-trees methods.
GenomeDx's proposed construction — which stretches "random-forest classifier" to encompass "a functionally equivalent ensemble-of-trees method" — faces significant headwinds. The Federal Circuit has generally held that claim terms with a recognized meaning in the art should be given that recognized meaning unless the specification clearly indicates a different intent. See, e.g., Enzo Biochem, Inc. v. Corara Corp., 780 F.3d 1149, 1155 (Fed. Cir. 2015). If the machine-learning literature uniformly treats random forests and gradient-boosted trees as distinct methods — which we believe it does — the court may be reluctant to construe "random-forest classifier" to encompass both.
Prosecution History Concern. The distinction drawn over Nakamura during '893 prosecution is a significant concern. The applicant's argument that the claimed invention was distinguishable based on the "random-forest classifier training methodology" could be read as an admission that the claim term is limited to the specific random-forest algorithm, complete with its bagging-based training methodology. If the court reads the Nakamura distinction in this way, it would support Helios's narrow construction and potentially bar the doctrine of equivalents under prosecution history estoppel.
However, there is an important counter-argument. A careful reading of the prosecution history reveals that the Nakamura distinction was based on two grounds: (1) the training methodology (bootstrap aggregation with random feature selection) and (2) the minimum cohort size (at least 500 samples). The applicant's argument focused on the fact that Nakamura's method did not employ the specific random-forest training methodology combined with the claimed minimum cohort size. The argument did not draw a distinction between random forests and gradient-boosted decision trees per se. Rather, it distinguished Nakamura's approach — which used a different predictive modeling framework altogether — on the basis of the training methodology and cohort requirements. This provides some room to argue that the prosecution history does not constitute a disclaimer of all tree-based ensemble methods beyond random forests, but rather a disclaimer of methods that do not employ ensemble-of-trees approaches at all.
Nonetheless, we must acknowledge that this is a close question, and the court could reasonably read the prosecution history either way.
Recommendation. In light of the foregoing analysis, we recommend the following multi-pronged approach:
(a) Advocate for GenomeDx's broader construction in the Markman briefing, emphasizing the specification's functional description of random forests and the common characteristics shared by random forests and gradient-boosted decision trees. Prepare expert testimony from Dr. Franklin Ogawa to support the argument that a POSITA would understand "random-forest classifier" to encompass closely related ensemble-of-trees methods.
(b) Prepare a fallback position for the Markman briefing, arguing in the alternative that even if the term is construed narrowly, the narrow construction does not preclude infringement under the doctrine of equivalents. Identify the function-way-result framework as the appropriate test.
(c) Develop a robust doctrine-of-equivalents argument for the '893 Patent in parallel with the claim construction effort. Under the function-way-result test: the gradient-boosted classifier performs substantially the same function (generating a prognostic score from gene-expression data using an ensemble of decision trees); operates in substantially the same way (aggregating outputs from multiple trees constructed on subsets of features); and achieves substantially the same result (a numerical score predictive of metastasis risk within five years).
(d) Complete the separate Prosecution History Estoppel Analysis Memorandum to fully assess the availability of the doctrine of equivalents in light of the Nakamura distinction. This analysis should address whether the narrowing during prosecution was directed to the specific bagging-vs.-boosting distinction or was more limited in scope.
(e) Emphasize the alternative infringement theories available under the '271 Patent and '647 Patent. Even if the '893 Patent claim construction is adverse, the '271 Patent's broader "trained machine-learning classifier" term provides a direct literal infringement theory, and the '647 Patent's "classifier comprising at least 20 gene features" — which does not specify any particular algorithm — provides an additional alternative.
Risk Assessment: We assess a moderate-to-high risk that the court will adopt Helios's narrower construction of this term. The recognized distinction between random forests and gradient-boosted decision trees in the machine-learning literature, combined with the prosecution history concerns, creates meaningful exposure on this term.
C. "Panel of at Least 22 Genes Selected from a Predefined Gene Set" ('271 Patent, Claim 1)
1. Claim Language in Context
The term "panel of at least 22 genes selected from a predefined gene set" appears in the context of the following limitation of Claim 1 of the '271 Patent:
"measuring expression levels of a panel of at least 22 genes selected from a predefined gene set"
This term appears in all asserted claims of the '271 Patent — Claims 1, 3, 5, 7, 12, and 18. The specification of the '271 Patent defines the "predefined gene set" by reference to a table (Table 1) listing specific genes identified as being associated with prostate cancer prognosis. Table 1 contains a list of genes with their corresponding gene identifiers and functional annotations, and the specification states that "the predefined gene set comprises genes selected based on their demonstrated association with prostate cancer progression and metastatic potential."
2. GenomeDx's Proposed Construction
GenomeDx proposes the following construction:
"Measuring the expression levels of 22 or more genes chosen from the gene set identified in the specification (Table 1 of the '271 Patent), which may be a subset of a larger gene panel measured by the accused method."
The key aspect of GenomeDx's proposed construction is that the claim requires measurement of "at least 22 genes" that are "selected from" the predefined gene set — it does not require that the measured panel consist of exactly 22 genes, nor does it prohibit the measurement of additional genes not found in the predefined set. A panel of 30 genes, of which 22 or more come from the predefined set, would satisfy this limitation.
Application to ProstaSignal™. As noted above, Helios's ProstaSignal™ test measures 24 genes. Of these 24 genes, 19 overlap with the predefined gene set identified in Table 1. If this overlap number is accurate, then literal infringement of this limitation is not met, because only 19 — not the required "at least 22" — of the ProstaSignal™ genes come from the predefined gene set.
However, several avenues of investigation remain open. The 5 non-overlapping ProstaSignal™ genes must be carefully analyzed to determine whether any of them are, in fact, genes from the predefined gene set that are designated by alternative names, represent splice variants, or are functional equivalents of genes listed in Table 1. Additionally, the specification must be reviewed to determine whether the "predefined gene set" is defined solely by Table 1 or whether other portions of the specification — including dependent claims listing additional genes, or descriptive passages identifying genes associated with prostate cancer prognosis — broaden the set beyond Table 1.
3. Helios's Anticipated Proposed Construction
Based on the August 4, 2025 exchange, Helios proposes the following construction:
"Measuring expression levels of a panel consisting of at least 22 genes, all of which must be selected from the specific gene set defined in Table 1 of the '271 Patent specification."
Helios's construction largely aligns with GenomeDx's on the basic requirement — at least 22 genes from the predefined set — but adds two noteworthy features. First, Helios uses the word "all," emphasizing that every one of the 22 required genes must come from Table 1, with no room for functional equivalents or genes identified elsewhere in the specification. Second, Helios specifically limits the "predefined gene set" to Table 1, foreclosing the possibility that the set could be defined more broadly by reference to other specification passages.
Helios will argue that ProstaSignal™ measures only 19 genes from the Table 1 gene set, falling short of the "at least 22" threshold, and therefore does not literally infringe this limitation. Helios may further argue that the specification's Table 1 is the exclusive and exhaustive definition of "predefined gene set," and that no additional genes — regardless of functional similarity — can be read into the set.
4. Analysis and Recommendation
This term presents a significant vulnerability for the '271 Patent infringement theory if the gene overlap is confirmed at only 19 of 24.
At the construction level, the parties' proposed constructions are not dramatically different — both require at least 22 genes from the predefined gene set. The real dispute is factual: whether 22 or more ProstaSignal™ genes actually come from the predefined set. The construction dispute focuses on the secondary question of how the "predefined gene set" is defined (Table 1 exclusively, or more broadly).
We recommend the following immediate action items:
(a) Conduct a detailed gene-by-gene comparison of all 24 ProstaSignal™ genes against all genes listed in Table 1 of the '271 Patent specification, as well as any other gene lists appearing in the specification (including genes identified in dependent claims, figures, and descriptive text). This analysis should include cross-referencing gene identifiers using standard databases (NCBI Gene, HGNC) to identify potential alternative nomenclature.
(b) Determine whether any of the 5 non-overlapping ProstaSignal™ genes are alternative names, splice variants, or functional equivalents of genes in the predefined set. Engage Dr. Franklin Ogawa or another genomics expert to perform this analysis.
(c) Review the specification for any broader definition of "predefined gene set" beyond Table 1. The specification may define the gene set by functional criteria (e.g., genes associated with prostate cancer progression) in addition to the specific enumeration in Table 1. If a broader definition exists, GenomeDx may argue that the predefined gene set includes genes beyond those listed in Table 1, potentially bringing additional ProstaSignal™ genes within scope.
If, after this analysis, the overlap remains at 19 genes, literal infringement of this limitation under the '271 Patent is likely not met. In that scenario, GenomeDx would need to rely on the doctrine of equivalents, arguing that the 19-gene subset performs substantially the same function, in substantially the same way, to achieve substantially the same result as a panel of at least 22 genes from the predefined set. This is a viable but not risk-free argument.
Importantly, we note that this vulnerability is specific to the '271 Patent. The '647 Patent uses the broader term "a classifier comprising at least 20 gene features" without requiring that the gene features be selected from a specific predefined gene set. ProstaSignal™ uses 24 genes, clearly satisfying the "at least 20" threshold. Similarly, the '893 Patent Claim 1 does not recite this exact limitation; it refers to "normalized expression values" without specifying a minimum gene count from a predefined set. Therefore, even if the '271 Patent faces difficulty on this term, the '893 and '647 Patents provide alternative infringement theories that avoid this issue.
Risk Assessment: We assess a moderate-to-high risk for the '271 Patent specifically, depending on the outcome of the gene-overlap analysis. This risk is lower for the '893 and '647 Patents, which do not include this specific limitation.
D. "Risk Category" ('271 Patent, Claim 1; '647 Patent, Claim 1)
1. Claim Language in Context
The term "risk category" appears in the following context in Claim 1 of the '271 Patent:
"classifying the patient into a risk category based on the score"
The '647 Patent uses similar language in the context of generating a report that includes a risk classification for the patient. The specifications of both patents describe exemplary risk categories as "low risk," "intermediate risk," and "high risk," but do not limit the number of categories or prescribe specific labels. The specifications indicate that risk categories are assigned based on predefined thresholds applied to the generated risk score, but do not impose any requirements regarding how those thresholds are determined or validated.
2. GenomeDx's Proposed Construction
GenomeDx proposes the following construction:
"A discrete classification grouping (e.g., low, intermediate, high, or other prognostic grouping) assigned to a patient based on the generated risk score, using predefined thresholds."
GenomeDx contends that this term should be given its plain and ordinary meaning as understood by a POSITA. No special construction is required beyond its ordinary meaning. The term "risk category" is a straightforward concept in clinical diagnostics and prognostic testing — it simply refers to a discrete grouping or classification that captures the patient's risk level.
Application to ProstaSignal™. ProstaSignal™ classifies patients into three categories — Low, Intermediate, and High risk — based on the ProstaScore. This classification directly meets the "risk category" limitation regardless of how the term is construed.
3. Helios's Anticipated Proposed Construction
Based on the August 4, 2025 exchange, Helios proposes the following construction:
"A clinically validated risk grouping based on thresholds determined from the same training cohort used to develop the classifier, corresponding to distinct clinical outcomes."
Helios's proposed construction imports two significant limitations not found in the claim language. First, it requires that the risk categories be "clinically validated" — a term that Helios does not define but presumably means that the categories have been validated through clinical studies demonstrating their correlation with actual patient outcomes. Second, it requires that the thresholds used to define the categories be "determined from the same training cohort used to develop the classifier."
We anticipate that Helios will advance this construction for both offensive and defensive purposes. Offensively, Helios will argue that ProstaSignal™'s risk thresholds were not derived from the same training cohort used to develop the gradient-boosted classifier, and therefore ProstaSignal™'s risk categories do not meet the "risk category" limitation as Helios construes it. Defensively, Helios may use this narrower construction to support invalidity arguments, contending that the narrower inventive concept does not distinguish the claimed invention from the prior art.
4. Analysis and Recommendation
GenomeDx's proposed construction is well-supported by the intrinsic record and should prevail.
The claim language imposes no "clinical validation" requirement. Claim 1 simply recites "classifying the patient into a risk category based on the score." There is no mention of clinical validation, threshold derivation methodology, or correlation with specific clinical outcomes. Helios's attempt to inject these requirements into the claim term is a textbook example of improperly importing limitations from the specification into the claims, which is prohibited under Phillips. See Phillips, 415 F.3d at 1323 ("We have . . . expressly rejected the contention that if a patent describes only a single embodiment, the claims of the patent must be construed as being limited to that embodiment.").
The specification describes risk categories broadly. While the specification identifies specific risk-category labels and discusses threshold determination using training data, it does not require that thresholds be derived from any particular cohort or that categories be clinically validated. The specification discusses clinical validation as a desirable attribute of the classifier system — not as a definitional requirement of the "risk category" concept. A POSITA would understand "risk category" to mean simply a discrete classification grouping based on the risk score, without additional requirements regarding how the groupings were established or validated.
Furthermore, Helios's proposed construction would effectively read a method-of-validation limitation into a claim term that describes a classification output. The claim does not recite any steps of validating the risk categories; it simply recites the step of classifying the patient into a category. The distinction between the classifier's output (a category) and the method by which the categories were established is significant, and Helios's construction improperly conflates the two.
Risk Assessment: We assess a low risk that the court will adopt Helios's narrower construction. This term is likely to be construed favorably for GenomeDx or given its plain and ordinary meaning.
E. "Composite Risk Metric" ('647 Patent, Claim 1)
1. Claim Language in Context
The term "composite risk metric" appears in the context of the following limitation of Claim 1 of the '647 Patent:
"generating a composite risk metric and transmitting it to a healthcare provider"
This limitation follows the step of "combining the genomic risk score with at least two clinical covariates selected from Gleason score, PSA level, surgical margin status, and pathological stage." The "composite risk metric" is thus the output of the combination step — it is the result of integrating the genomic risk score with the selected clinical covariates. This term is unique to the '647 Patent and does not appear in the '271 or '893 Patents.
2. GenomeDx's Proposed Construction
GenomeDx proposes the following construction:
"A single numerical or categorical output that integrates the genomic risk score with the selected clinical covariates into a unified prognostic assessment."
This construction is supported by the '647 Patent specification, which describes the composite risk metric as an "integrated prognostic output" that provides enhanced predictive value over the genomic score alone. At column 4, lines 10–25, the specification states:
"The composite risk metric provides an integrated prognostic output that combines the genomic risk score with clinical and pathological variables to produce a more comprehensive assessment of the patient's risk of disease progression. The composite risk metric may be expressed as a single numerical value, a categorical classification (e.g., low, intermediate, or high composite risk), or a probability of a clinical outcome within a defined time frame."
This passage is significant because it defines the composite risk metric broadly — as either a numerical value or a categorical classification — and describes it as an "integrated prognostic output" that "combines" genomic and clinical data. The specification does not require any specific mathematical operation for the integration.
Application to ProstaSignal™. ProstaSignal™ combines the ProstaScore with Gleason grade group and pre-operative PSA to generate an integrated clinical report for the treating physician. GenomeDx contends that this report constitutes a "composite risk metric" because it integrates genomic and clinical data into a unified prognostic assessment. However, the extent to which ProstaSignal™ mathematically combines these data inputs — as opposed to merely displaying them side by side in a report — is a factual question that must be resolved through discovery.
3. Helios's Anticipated Proposed Construction
Based on the August 4, 2025 exchange, Helios proposes the following construction:
"A single numerical score derived from a mathematical combination (e.g., weighted sum, regression model) of the genomic risk score and clinical covariates, as opposed to merely presenting the genomic score alongside clinical variables in a report."
Helios's construction adds a critical limitation: it requires that the composite risk metric be derived from a "mathematical combination" of the genomic and clinical data, and explicitly excludes the mere presentation of these data alongside each other in a report. This construction would create a factual battleground: if ProstaSignal™ mathematically combines the ProstaScore with Gleason and PSA data to generate a single integrated score, the limitation is met; if ProstaSignal™ merely displays these data points in separate fields of a report without mathematical integration, the limitation is not met.
Helios will argue that the claim language requires "generating" a composite metric, which implies a computational step rather than a mere formatting step. Helios will further argue that the term "composite" implies that discrete inputs are combined — i.e., mathematically integrated — into a single output, and that merely juxtaposing the inputs in a report does not create a "composite" metric.