CONFIDENTIAL — ATTORNEY WORK PRODUCT — DRAFT FOR REVIEW AND CIRCULATION TO RIDGECREST DURNEVAL LLP
UNITED STATES DISTRICT COURT NORTHERN DISTRICT OF CALIFORNIA
GENOME DX THERAPEUTICS INC.,
Plaintiff,
v.
HELIOS PRECISION DIAGNOSTICS, INC.,
Defendant.
Case No. 3:25-cv-01847-JWL
Hon. Jessica W. Langston
JOINT CLAIM CONSTRUCTION AND PREHEARING STATEMENT
Filed Pursuant to Patent Local Rule 4-3
Markman Hearing: September 18, 2026
DRAFT — Prepared by Calderwood & Harkness LLP for review and circulation to Ridgecrest Durneval LLP. Working draft as of February 2026. Not yet filed with the Court.
Counsel for Plaintiff Genome Dx Therapeutics Inc.:
Catherine Bloom (SBN 198742) Peter Vassiliev (SBN 312547) CALDERWOOD & HARKNESS LLP 555 Mission Street, Suite 3100 San Francisco, California 94105 Telephone: (415) 555-8200 Facsimile: (415) 555-8201 cbloom@calderwoodharkness.com pvassiliev@calderwoodharkness.com
Counsel for Defendant Helios Precision Diagnostics, Inc.:
Marcus Pellegrini (pro hac vice) Danielle Forrest (pro hac vice) RIDGECREST DURNEVAL LLP One Federal Street, Suite 3400 Boston, Massachusetts 02110 Telephone: (617) 555-4400 Facsimile: (617) 555-4401 — and — 350 South Grand Avenue, Suite 2800 Los Angeles, California 90071 Telephone: (213) 555-7700 mpellegrini@ridgecrestkirkpatrick.com dforrest@ridgecrestkirkpatrick.com
Right-click to update Table of Contents
I. INTRODUCTION
Plaintiff Genome Dx Therapeutics Inc. ("GenomeDx") and Defendant Helios Precision Diagnostics, Inc. ("Helios") (collectively, the "Parties") respectfully submit this Joint Claim Construction and Prehearing Statement ("Joint Statement") pursuant to Patent Local Rule 4-3 of the Northern District of California, in advance of the Markman hearing currently scheduled for September 18, 2026 before the Honorable Jessica W. Langston.
This action involves three United States patents owned by GenomeDx, each directed to methods and systems for predicting clinical outcomes in prostate cancer patients through gene-expression-based classifiers:
(1) U.S. Patent No. 10,438,271 (the "'271 Patent"), entitled "Methods for Predicting Prostate Cancer Outcomes Using Multi-Gene Expression Classifiers," issued October 8, 2019;
(2) U.S. Patent No. 11,204,893 (the "'893 Patent"), entitled "Genomic Classifier Systems and Methods for Post-Surgical Cancer Prognosis," issued December 21, 2021; and
(3) U.S. Patent No. 11,891,647 (the "'647 Patent"), entitled "Improved Methods for Genomic Risk Stratification in Prostate Cancer Using Combined Biomarker and Clinical Variable Models," issued February 6, 2024.
The Parties have engaged in extensive meet-and-confer efforts regarding the construction of claim terms in the patents-in-suit. The Parties initially exchanged proposed claim terms and preliminary constructions on August 4, 2025, in accordance with Patent Local Rule 4-1. Subsequent exchanges and telephonic conferences occurred on September 19, 2025; October 31, 2025; December 12, 2025; and January 22, 2026. Through these discussions, the Parties have identified six (6) claim terms for which the Parties have reached agreement on construction and seven (7) claim terms that remain in dispute and that require resolution by the Court at the Markman hearing.
The Court's Scheduling Order, issued June 18, 2025, establishes the timeline for claim construction proceedings in this action, including the September 18, 2026 Markman hearing date. In addition, the Parties note that the Court's Order dated January 13, 2026, denying Helios's Motion to Dismiss Counts II and III of GenomeDx's First Amended Complaint under 35 U.S.C. § 101 (Dkt. No. 78), held that Helios's patent-eligibility challenges were premature absent a full claim construction. The Court reasoned that the scope and meaning of the asserted claims could not be adequately assessed for purposes of § 101 analysis until the disputed claim terms had been construed. That ruling underscores the importance of the instant proceeding to the resolution of the fundamental issues in this litigation, including the scope of the asserted claims, the viability of GenomeDx's infringement theories, and Helios's potential renewed § 101 defenses following claim construction.
This Joint Statement sets forth: (1) a description of the patents-in-suit and their prosecution histories; (2) a description of the accused instrumentality; (3) the Parties' agreed claim constructions; (4) the Parties' disputed claim constructions, together with each Party's proposed construction, supporting evidence, and the significance of each dispute; (5) identification of extrinsic evidence the Parties anticipate relying upon at the Markman hearing; (6) a proposed briefing schedule; and (7) additional matters relevant to claim construction.
II. THE PATENTS-IN-SUIT
A. 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, and Dr. David Petrossian
Asserted Claims: Claims 1, 3, 5, 7, 12, and 18
The '271 Patent is directed to methods for predicting the likelihood of clinical recurrence of prostate cancer following radical prostatectomy. The disclosed methods employ multi-gene expression classifiers applied to tumor RNA extracted from biological samples to generate prognostic risk assessments. The patent describes a process by which RNA is obtained from tumor tissue, the expression levels of a defined panel of genes are measured, and a trained machine-learning classifier is applied to those expression levels to generate a risk score that is used to classify patients into prognostic risk categories. The '271 Patent specification describes this technology as providing a clinically actionable tool for identifying patients at elevated risk of disease recurrence, thereby enabling clinicians to make more informed decisions regarding adjuvant therapy following surgical intervention.
Representative Independent Claim (Claim 1):
1\. A method for prognosing the likelihood of clinical recurrence following radical prostatectomy, the method comprising:
(a) obtaining a biological sample containing tumor RNA from a patient who has undergone radical prostatectomy;
(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 the measured expression levels to generate a risk score; and
(d) classifying the patient into a risk category based on the risk score.
Prosecution History Notes: The '271 Patent issued after two office actions during prosecution before the United States Patent and Trademark Office. In the first office action, the Examiner rejected all pending claims under 35 U.S.C. § 101, applying the two-step framework established by the Supreme Court in Mayo Collaborative Services v. Prometheus Laboratories, Inc., 566 U.S. 66 (2012), and Alice Corp. v. CLS Bank International, 573 U.S. 208 (2014). The Examiner determined that the claims were directed to the abstract idea of correlating gene-expression data with clinical outcomes. In response, the applicant amended the claims to recite specific technical steps, including RNA extraction from formalin-fixed paraffin-embedded ("FFPE") samples and reference-gene normalization procedures. The applicant argued in a response dated March 8, 2019 that the claims, as amended, were not directed to an abstract idea but rather to a specific, concrete diagnostic improvement comprising particular physical and computational steps that transform tumor tissue into a clinically actionable risk classification. The Examiner was persuaded by these arguments and withdrew the § 101 rejection in the second office action. No prior-art rejections under 35 U.S.C. §§ 102 or 103 were sustained during prosecution. The '271 Patent issued without any narrowing amendments to the classifier-related claim limitations.
B. 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 (same priority chain as the '271 Patent)
Named Inventors: Dr. Rajan Mehta and Dr. Carolyn Xu
Asserted Claims: Claims 1, 2, 9, 14, and 20
The '893 Patent is a continuation application within the same patent family as the '271 Patent and claims priority to the same earliest effective filing date of June 14, 2016. The '893 Patent is directed to systems and methods for post-surgical cancer prognosis using normalized gene-expression data obtained from FFPE prostate tissue samples. A central feature of the '893 Patent claims is the recitation of a "random-forest classifier" trained on a specific minimum patient cohort of at least 500 patient samples. The specification describes in detail the random-forest algorithm implementation, including the number of trees used, the feature-selection methodology, the bootstrap sampling procedure, and the aggregation of tree-level predictions to produce a final prognostic score. The '893 Patent further specifies that the resulting numerical prognostic score is correlated with metastasis risk within a five-year post-surgical window.
Representative Independent Claim (Claim 1):
1\. A method comprising:
(a) normalizing gene-expression data from a formalin-fixed paraffin-embedded prostate tissue sample using a reference gene panel;
(b) inputting the 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 of radical prostatectomy.
Prosecution History Notes: The '893 Patent issued after one office action. The Examiner initially rejected claims under 35 U.S.C. § 102(a)(1) based on Nakamura et al., U.S. Patent No. 9,617,340 (issued April 4, 2017). In its response dated August 16, 2021, the applicant distinguished the claims over Nakamura by arguing that Nakamura did not disclose the specific random-forest classifier training methodology recited in the claims, and in particular did not disclose a random-forest classifier trained on a cohort of at least 500 patient samples. The applicant specifically argued that "Nakamura teaches a logistic regression model trained on a smaller patient cohort and does not disclose or suggest the use of a random-forest classifier trained on a minimum of 500 patient samples as required by the present claims." The applicant further argued that the random-forest training methodology yields distinct and superior prognostic performance compared to the logistic regression approach of Nakamura. The Examiner found these arguments persuasive and withdrew the § 102 rejection. No other rejections were raised, and the patent issued. The prosecution history of the '893 Patent, and in particular the applicant's arguments distinguishing Nakamura on the basis of the random-forest classifier training methodology, is a significant point of contention in the Parties' claim construction dispute regarding the term "random-forest classifier," as discussed in Section V.B below.
C. 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, and Dr. Priya Shah
Asserted Claims: Claims 1, 4, 6, 11, and 15
The '647 Patent is directed to improved methods for generating integrated prognostic reports that combine genomic risk scores derived from gene-expression classifiers with clinical covariates — such as Gleason score, PSA level, surgical margin status, and pathological stage — to produce composite risk metrics. The specification describes this integrated approach as an advancement over genomic-only models, providing enhanced prognostic accuracy by synthesizing molecular and clinical data into a unified assessment. The '647 Patent discloses specific mathematical formulas for combining genomic scores with clinical variables, as well as methods for validating the resulting composite metrics against clinical outcomes in independent patient cohorts.
Representative Independent Claim (Claim 1):
1\. A method for generating an integrated prognostic report, the method comprising:
(a) performing RNA extraction and expression profiling on a prostate tumor specimen obtained from a patient who has undergone radical prostatectomy;
(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 to produce a composite risk metric; and
(d) generating the integrated prognostic report comprising the composite risk metric and transmitting the integrated prognostic report to a healthcare provider.
Prosecution History Notes: The '647 Patent issued after three office actions, reflecting a more extensive prosecution than its predecessors. The Examiner initially rejected the claims under 35 U.S.C. § 103 based on a combination of Nakamura et al. (U.S. Patent No. 9,617,340) and Gagnon (WO 2018/065413, published April 12, 2018). The Examiner determined that Nakamura disclosed gene-expression-based prognostic scoring and that Gagnon disclosed the combination of genomic scores with clinical covariates. The applicant overcame this rejection by submitting a declaration under 37 C.F.R. § 1.132 from named inventor Dr. Rajan Mehta, which presented experimental data demonstrating that the combination of genomic risk scores with clinical covariates as specifically claimed produced unexpected and statistically significant improvements in prognostic accuracy compared to genomic-only models, including improvements in area-under-the-curve (AUC) metrics and net reclassification indices. Dr. Mehta's declaration established that the claimed combined genomic-clinical model achieved an AUC of 0.88 for predicting five-year metastasis, compared to an AUC of 0.76 for the genomic-only model and 0.71 for a clinical-only model. The Examiner found the evidence of unexpected results persuasive and withdrew the § 103 rejection. A terminal disclaimer was filed over both the '271 Patent and the '893 Patent to overcome a non-statutory double-patenting rejection, thereby linking the enforceability of the '647 Patent to the '271 and '893 Patents.
III. THE ACCUSED PRODUCT
The accused instrumentality in this action is Helios's "ProstaSignal™ Genomic Classifier" (the "ProstaSignal™" test or the "Accused Product"). ProstaSignal™ is a 24-gene expression-based prognostic test designed for prostate cancer patients following radical prostatectomy. Helios commercially launched ProstaSignal™ on October 7, 2024, and has marketed the test to urologists and oncologists throughout the United States.
ProstaSignal™ operates by extracting RNA from FFPE prostate tissue specimens, measuring expression levels of a panel of 24 genes, and applying a computational classifier to generate a prognostic score. The classifier employed by ProstaSignal™ is a gradient-boosted decision tree model, which produces a numerical output designated the "ProstaScore," ranging from 0 to 100. Based on the ProstaScore, patients are classified into one of three risk categories: Low risk, Intermediate risk, or High risk. These risk categories are intended to assist physicians in making treatment decisions regarding adjuvant therapy, including radiation therapy and hormonal therapy, following radical prostatectomy.
In addition to the ProstaScore, the ProstaSignal™ test generates an integrated report that combines the ProstaScore with Gleason grade group and pre-operative PSA level. The integrated report presents these values alongside one another and provides a narrative risk summary for the treating physician. The report is delivered to ordering physicians electronically through Helios's proprietary web portal and, in some cases, via integration with electronic health record systems.
Of the 24 genes included in the ProstaSignal™ panel, 19 genes overlap with GenomeDx's predefined gene set as identified in the specification of the '271 Patent (Table 1). The remaining 5 genes in the ProstaSignal™ panel are not part of GenomeDx's predefined gene set. This overlap is a central factual issue relevant to several of the disputed claim terms discussed in Section V below.
ProstaSignal™ is performed exclusively at Helios's CLIA-certified clinical laboratory located at 9830 Mesa Rim Road, Suite 140, San Diego, California 92121. Helios obtained a Multianalyte Assays with Algorithmic Analyses ("MAAA") code — specifically, code 0239M — from the American Medical Association's CPT Editorial Panel in September 2024 for billing and reimbursement purposes in connection with the ProstaSignal™ test.
IV. AGREED CLAIM CONSTRUCTIONS
The Parties have conferred in good faith and have reached agreement on the construction of the following six (6) claim terms. The Parties submit that no further briefing, argument, or testimony is required with respect to these terms, and respectfully request that the Court adopt the agreed constructions set forth below.
Term 1: "biological sample containing tumor RNA"
Patent/Claim(s): '271 Patent, Claim 1, limitation (a)
Agreed Construction: "A tissue sample obtained from a patient that contains ribonucleic acid (RNA) expressed by tumor cells."
Both Parties agree that this term carries its ordinary and customary meaning as understood by a person of ordinary skill in the art and does not require further construction beyond the agreed definition set forth above. The term encompasses any tissue sample — including, without limitation, surgically excised tissue, biopsy specimens, and FFPE tissue blocks — that contains RNA originating from tumor cells.
Term 2: "expression levels"
Patent/Claim(s): '271 Patent, Claims 1, 3, 5; '893 Patent, Claim 1; '647 Patent, Claims 1, 4
Agreed Construction: "Quantitative measurements of the amount of RNA transcripts for specified genes in a sample."
The Parties agree that this term has its plain and ordinary meaning in the art and encompasses any quantitative measurement methodology for determining the abundance of RNA transcripts, including but not limited to microarray hybridization, reverse-transcription quantitative PCR (RT-qPCR), and RNA sequencing (RNA-seq).
Term 3: "formalin-fixed paraffin-embedded prostate tissue sample"
Patent/Claim(s): '893 Patent, Claim 1
Agreed Construction: "A prostate tissue specimen that has been preserved using formalin fixation and embedded in paraffin wax, as that process is understood by one of ordinary skill in the art of histopathology."
The Parties agree that this is a standard term of art in histopathology and clinical laboratory science. FFPE processing is the predominant tissue preservation method used in clinical pathology laboratories, and the meaning of this term is well-established and not subject to dispute.
Term 4: "reference gene panel"
Patent/Claim(s): '893 Patent, Claim 1
Agreed Construction: "A set of genes used as internal controls for normalization of gene-expression data."
The Parties agree that this term encompasses housekeeping genes and other internal-control genes whose expression levels remain relatively stable across different experimental conditions, tissue types, and patient samples. The reference gene panel serves as a baseline against which the expression levels of the target genes are normalized in order to account for variability in sample quality, RNA extraction efficiency, and assay performance.
Term 5: "Gleason score" / "PSA level" / "surgical margin status" / "pathological stage"
Patent/Claim(s): '647 Patent, Claim 1, limitation (c)
Agreed Construction: Each of these four clinical covariate terms carries its ordinary and customary meaning as understood by a person of ordinary skill in the art of urological oncology:
• Gleason score: A histological grading system for prostate cancer based on the microscopic appearance of tumor tissue, in which a primary and secondary pattern grade are assigned and summed to produce a total score reflecting tumor differentiation and aggressiveness.
• PSA level: The serum concentration of prostate-specific antigen, a protein produced by prostate tissue, typically measured in nanograms per milliliter (ng/mL), which serves as a biomarker for prostate cancer detection and monitoring.
• Surgical margin status: An assessment of whether tumor cells are present at the inked edge of surgically removed tissue (positive margin) or absent from the edge (negative margin), as determined by pathological examination of the resection specimen.
• Pathological stage: The staging classification of tumor extent based on pathological examination of the resected specimen and associated lymph nodes, typically reported according to the American Joint Committee on Cancer (AJCC) TNM staging system.
The Parties agree that these terms are widely used and well-understood in clinical urology and oncology and do not require further construction.
Term 6: "healthcare provider"
Patent/Claim(s): '647 Patent, Claim 1, limitation (d)
Agreed Construction: "A licensed medical professional or institution authorized to provide medical care to a patient."
The Parties agree that this term is not limited to any particular type of medical professional or institution and encompasses physicians, nurse practitioners, physician assistants, hospitals, clinics, and other entities authorized to deliver medical care.
V. DISPUTED CLAIM CONSTRUCTIONS
The Parties have been unable to reach agreement on the construction of the following seven (7) claim terms despite good-faith meet-and-confer efforts. The Parties respectfully request that the Court construe these terms at the Markman hearing scheduled for September 18, 2026. For each disputed term, the Parties set forth their respective proposed constructions, the bases for their positions, the intrinsic and extrinsic evidence supporting their positions, and the significance of the dispute to the issues of infringement and validity in this action.
A. Disputed Term 1: "trained machine-learning classifier"
Patent/Claim(s): '271 Patent, Claims 1, 7, 12, and 18
1. Plaintiff GenomeDx's Proposed Construction
Proposed Construction: "A computational model that has been trained on labeled data using any 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 predictive output."
Basis for GenomeDx's Proposed Construction:
GenomeDx submits that the term "trained machine-learning classifier" as used in the '271 Patent is a broad term that encompasses any supervised machine-learning algorithm that has been fitted to labeled training data for the purpose of generating a predictive output. The '271 Patent specification describes the classifier at a high level of generality and expressly references multiple machine-learning approaches as suitable for implementing the claimed methods. Specifically, the specification states:
"The classifier used in the methods of the present invention may be any supervised machine-learning model capable of learning patterns from labeled training data and generating predictive outputs. Suitable classifiers include, but are not limited to, random forests, gradient-boosted decision trees, support vector machines, artificial neural networks, and penalized logistic regression models."
'271 Patent, col. 8, ll. 14–42. This passage demonstrates that the inventors intended the term to be broad and not limited to any single algorithm.
The prosecution history of the '271 Patent further supports GenomeDx's proposed construction. During prosecution, the applicant amended the claims to overcome the Examiner's § 101 rejection by adding specific physical processing steps — namely, RNA extraction from FFPE samples and reference-gene normalization — rather than by narrowing the scope of the classifier limitation. At no point during prosecution did the applicant limit the "trained machine-learning classifier" to any particular algorithm or class of algorithms. The amendments and arguments that overcame the § 101 rejection were directed to the sample-processing limitations, not the classifier type. This history confirms that the breadth of the "trained machine-learning classifier" term was preserved throughout prosecution.
Principles of claim differentiation also support GenomeDx's proposed construction. The '893 Patent, which is a continuation of the '271 Patent and shares the same priority date, specifically claims a "random-forest classifier" in its independent claims. If the broader term "trained machine-learning classifier" in the '271 Patent were construed to mean the same thing as "random-forest classifier" in the '893 Patent, the two patents would have coextensive claim scope, rendering the use of different terminology across the patent family meaningless. Under well-established claim differentiation principles, the use of different terms in related patents within the same family creates a presumption that the terms have different meanings. See Curtiss-Wright Flow Control Corp. v. Velan, Inc., 438 F.3d 1374, 1380–81 (Fed. Cir. 2006); Phillips v. AWH Corp., 415 F.3d 1303, 1314–15 (Fed. Cir. 2005) (en banc).
Intrinsic Evidence Supporting GenomeDx's Proposed Construction:
(a) '271 Patent specification, col. 8, ll. 14–42 (discussing multiple machine-learning approaches as suitable classifiers);
(b) '271 Patent prosecution history, including the applicant's response dated March 8, 2019, demonstrating that the § 101 rejection was overcome by amendments to sample-processing steps, not by narrowing the classifier type;
(c) Claim differentiation between the '271 Patent ("trained machine-learning classifier") and the '893 Patent ("random-forest classifier"), supporting a broader construction of the '271 Patent term; and
(d) Dependent claims of the '271 Patent, which add further limitations to the classifier without specifying a particular algorithm type, confirming that the independent claim term is broad.
Extrinsic Evidence Supporting GenomeDx's Proposed Construction:
(a) Anticipated testimony of Dr. Franklin Ogawa, Professor of Computational Oncology at Stanford University, regarding the understanding of a person of ordinary skill in the art as to the meaning of "trained machine-learning classifier," including that the term would be understood to encompass any supervised learning algorithm;
(b) Technical publications and textbooks on machine learning in genomics, to be identified in subsequent briefing; and
(c) Standard technical dictionaries defining "machine-learning classifier" in broad terms.
2. Defendant Helios's Proposed Construction
Proposed Construction: "A random-forest or ensemble tree-based classifier that has been trained on labeled genomic data from patient samples to generate a predictive output."
Basis for Helios's Proposed Construction:
Helios contends that the specification of the '271 Patent, while mentioning multiple machine-learning approaches in a general introductory passage, predominantly describes and exemplifies random-forest classifiers and closely related ensemble tree-based methods throughout the remainder of the disclosure. Every working example in the specification employs a random-forest model. The specification devotes extensive discussion to the construction, training, and validation of random-forest classifiers, including the number of trees, the feature-selection criteria, the bootstrap aggregation procedure, and the out-of-bag error estimation methodology. No working example uses a support vector machine, a neural network, or a logistic regression model. Under Federal Circuit precedent, while claims are not limited to the preferred embodiments disclosed in the specification, the specification remains the best guide to claim meaning, and a patentee may not use claim construction to recapture scope that is unsupported by the disclosure. See Phillips, 415 F.3d at 1316; Tronzo v. Biomet, Inc., 156 F.3d 1154, 1159 (Fed. Cir. 1998).
Helios further contends that the patent family as a whole — including the '893 Patent continuation — confirms that the inventors understood the classifier to be a random-forest or ensemble tree-based model. The '893 Patent's narrower recitation of "random-forest classifier" reflects a refinement of the same technology, not a fundamentally different algorithm. Reading the '271 Patent term "trained machine-learning classifier" as encompassing any conceivable supervised learning algorithm would render the term unbounded and divorced from the specification's disclosure.
Intrinsic Evidence Supporting Helios's Proposed Construction:
(a) '271 Patent specification working examples (all employing random-forest models);
(b) '893 Patent Claim 1, specifying "random-forest classifier" as a refinement of the same technology;
(c) '893 Patent prosecution history, in which the applicant distinguished Nakamura based on the random-forest classifier training methodology, confirming the centrality of the random-forest algorithm to the patent family; and
(d) The overall disclosure of the '271 Patent, which focuses extensively on ensemble tree-based methods.
Extrinsic Evidence Supporting Helios's Proposed Construction:
Helios intends to identify its extrinsic evidence in its opening claim construction brief. Helios anticipates presenting expert testimony from a defense technical expert (to be identified) regarding the understanding of a person of ordinary skill in the art and the distinctions between different classes of machine-learning algorithms.
3. Significance of Dispute
The construction of "trained machine-learning classifier" is of critical importance to the infringement analysis for the '271 Patent. If the Court adopts GenomeDx's proposed construction, Helios's gradient-boosted decision tree classifier (used in ProstaSignal™) would fall within the literal scope of the "trained machine-learning classifier" limitation, and GenomeDx could establish literal infringement of the '271 Patent claims (subject to satisfaction of all other claim limitations). If the Court adopts Helios's proposed construction, which limits the term to random-forest or ensemble tree-based classifiers, the question becomes whether a gradient-boosted decision tree is an "ensemble tree-based classifier" within the meaning of the construction — gradient boosting does employ ensembles of decision trees, but through a sequential boosting methodology rather than the parallel bagging methodology characteristic of random forests. Depending on how narrowly Helios's construction is interpreted, GenomeDx may be required to rely on the doctrine of equivalents to establish infringement of the '271 Patent claims with respect to the classifier limitation.
This dispute affects all six asserted claims of the '271 Patent: independent Claims 1 and 12, and dependent Claims 3, 5, 7, and 18.
B. Disputed Term 2: "random-forest classifier"
Patent/Claim(s): '893 Patent, Claims 1, 2, 9, 14, and 20
1. Plaintiff GenomeDx's Proposed Construction
Proposed Construction: "A classifier using an ensemble of decision trees in which multiple decision trees are constructed during training and the output is determined by aggregating the predictions of the individual trees, including variants and functional equivalents of the random-forest algorithm such as gradient-boosted tree ensembles."
Basis for GenomeDx's Proposed Construction:
GenomeDx contends that the term "random-forest classifier" as used in the '893 Patent encompasses the family of ensemble tree-based classification methods, not merely the specific random-forest algorithm as originally described by Breiman (2001). The '893 Patent specification describes the random-forest approach as one implementation of ensemble tree classification methodology and includes passages discussing the broader category of ensemble methods. Specifically, the specification states:
"Ensemble methods, including bagging-based approaches such as random forests and boosting-based approaches such as gradient boosting, leverage the predictive power of multiple decision trees to achieve superior classification performance compared to individual tree models."
'893 Patent, col. 6, ll. 32–38. This passage demonstrates that the inventors understood random forests and gradient-boosted tree models as related approaches within the broader family of ensemble tree methods.
GenomeDx further contends that a person of ordinary skill in the art as of the priority date would have understood the term "random-forest classifier" to encompass closely related ensemble tree methods, because the defining characteristic of the claim is the use of multiple decision trees whose predictions are aggregated — a feature shared by both random forests and gradient-boosted tree ensembles. While the specific mechanisms for constructing and aggregating trees differ between random forests (parallel/bagging) and gradient boosting (sequential/boosting), both methods share the fundamental architecture of ensemble tree classification.
Intrinsic Evidence Supporting GenomeDx's Proposed Construction:
(a) '893 Patent specification, col. 6, ll. 32–38 (discussing bagging and boosting as related ensemble techniques);
(b) '893 Patent specification's description of the random-forest approach as an implementation of ensemble tree classification; and
(c) The claim language itself, which recites "random-forest classifier" without incorporating the specific procedural details of the Breiman (2001) algorithm (e.g., bootstrap sampling, majority voting).
Extrinsic Evidence Supporting GenomeDx's Proposed Construction:
(a) Machine learning textbooks discussing the relationship between random forests and gradient-boosted trees as related ensemble methods;
(b) Anticipated testimony of Dr. Franklin Ogawa regarding the understanding of a person of ordinary skill in the art; and
(c) Technical literature demonstrating that random forests and gradient-boosted tree models are frequently discussed and compared as related approaches in genomic classification applications.
2. Defendant Helios's Proposed Construction
Proposed Construction: "A classifier that uses the specific random-forest algorithm as described by Breiman (2001), in which multiple decision trees are independently trained on bootstrap samples of the training data and predictions are aggregated by majority vote or averaging, and which does not encompass gradient-boosted decision trees or other sequential ensemble methods."
Basis for Helios's Proposed Construction:
Helios contends that "random-forest classifier" is a well-defined term of art in the field of machine learning that refers specifically to the algorithm described in Leo Breiman's seminal 2001 paper, "Random Forests," published in Machine Learning, Vol. 45, pp. 5–32. The random-forest algorithm has three defining characteristics: (1) multiple decision trees are trained independently and in parallel; (2) each tree is trained on a bootstrap sample (i.e., a random sample drawn with replacement) of the training data; and (3) predictions from the individual trees are aggregated by majority vote (for classification) or averaging (for regression). This algorithm is fundamentally and materially different from gradient-boosted decision trees, which employ a sequential training procedure in which each successive tree is constructed to correct the errors of the preceding trees. The gradient-boosted approach, as described in Friedman's foundational 2001 paper, "Greedy Function Approximation: A Gradient Boosting Machine," operates through an entirely different mathematical framework.
Helios further contends that the prosecution history of the '893 Patent creates prosecution history estoppel that precludes GenomeDx from expanding the scope of "random-forest classifier" to encompass gradient-boosted decision trees. During prosecution, the applicant overcame the Examiner's § 102 rejection based on Nakamura by specifically arguing that the claims were distinguishable because Nakamura did not disclose a "random-forest classifier" and instead taught a logistic regression model. The applicant's arguments emphasized the specific training methodology of the random-forest algorithm — including bootstrap sampling and ensemble aggregation — as the basis for distinguishing the prior art. Having narrowed the claim scope by relying on the specific characteristics of the random-forest algorithm to overcome a prior-art rejection, GenomeDx cannot now expand the term through claim construction to recapture methods that were surrendered during prosecution. See Spectrum Int'l, Inc. v. Sterilite Corp., 164 F.3d 1372, 1378 (Fed. Cir. 1998) (prosecution history estoppel prevents a patentee from recapturing through claim construction scope that was relinquished during prosecution).
Intrinsic Evidence Supporting Helios's Proposed Construction:
(a) '893 Patent prosecution history — specifically, the applicant's response dated August 16, 2021, in which the applicant distinguished Nakamura by arguing the specific random-forest classifier training methodology and minimum cohort size requirement;
(b) The Examiner's withdrawal of the § 102 rejection based on the applicant's arguments regarding the random-forest classifier — confirming that the claim scope was defined by reference to the specific random-forest algorithm; and
(c) The '893 Patent specification's detailed description of the random-forest algorithm, including the number of trees, bootstrap sampling procedure, and majority-vote aggregation — all features specific to random forests and not shared by gradient-boosted models.
Extrinsic Evidence Supporting Helios's Proposed Construction:
(a) Breiman, "Random Forests," Machine Learning, Vol. 45, pp. 5–32 (2001) — the foundational reference defining the random-forest algorithm;
(b) Friedman, "Greedy Function Approximation: A Gradient Boosting Machine," Annals of Statistics, Vol. 29, No. 5, pp. 1189–1232 (2001) — defining gradient boosting as a fundamentally different algorithm;
(c) Machine learning reference texts distinguishing random forests from gradient-boosted decision trees as distinct algorithmic approaches; and
(d) Anticipated testimony of a defense technical expert (to be identified) regarding the understanding of a person of ordinary skill in the art.
3. Significance of Dispute
This is the central claim construction dispute for the '893 Patent. The resolution of this dispute will likely determine whether GenomeDx can establish literal infringement of the '893 Patent claims or must instead rely on the doctrine of equivalents.
If the Court adopts GenomeDx's proposed construction, which defines "random-forest classifier" broadly to include gradient-boosted tree ensembles, then Helios's ProstaSignal™ gradient-boosted decision tree classifier may literally satisfy the "random-forest classifier" limitation, and GenomeDx may be able to establish literal infringement of the '893 Patent claims (subject to satisfaction of all other claim limitations).
If the Court adopts Helios's proposed construction, which limits "random-forest classifier" to the specific Breiman (2001) algorithm and expressly excludes gradient-boosted decision trees, then GenomeDx cannot establish literal infringement of the '893 Patent claims with respect to the classifier limitation. GenomeDx would be required to proceed under the doctrine of equivalents, which is significantly complicated by the prosecution history estoppel analysis. During prosecution of the '893 Patent, the applicant distinguished Nakamura by arguing the specific random-forest training methodology. Helios will argue that this creates prosecution history estoppel under Festo Corp. v. Fenmore & Co., 535 U.S. 722 (2002), precluding GenomeDx from asserting that gradient-boosted decision trees are equivalent to random-forest classifiers. GenomeDx anticipates arguing that the distinction drawn during prosecution was based on the training methodology (bootstrap sampling and minimum cohort size) and the contrast with logistic regression (an entirely different model class), not on the exclusion of related ensemble tree methods such as gradient boosting. GenomeDx will contend that the Festo presumption is rebutted because gradient-boosted decision trees were not foreseeable as a substitute at the time of the narrowing amendment, or alternatively that the rationale for the amendment bore no more than a tangential relationship to the equivalent in question.
The resolution of this dispute affects all five asserted claims of the '893 Patent: independent Claims 1 and 14, and dependent Claims 2, 9, and 20.
C. Disputed Term 3: "panel of at least 22 genes selected from a predefined gene set"
Patent/Claim(s): '271 Patent, Claim 1 (also implicating dependent Claims 3 and 5)
1. Plaintiff GenomeDx's Proposed Construction
Proposed Construction: "A collection of 22 or more genes chosen from a gene set identified and defined in advance, where the gene panel used for expression measurement includes at least 22 genes from the predefined set but may also include additional genes not in the predefined set."
Basis for GenomeDx's Proposed Construction:
GenomeDx contends that the claim language "a panel of at least 22 genes selected from a predefined gene set" establishes a minimum threshold: the panel must include at least 22 genes that are drawn from the predefined gene set disclosed in the specification. Critically, the claim uses the open-ended transitional phrase "comprising" and does not prohibit the panel from additionally including genes that are not part of the predefined gene set. The claim recites what must be included (at least 22 genes from the predefined set) but does not recite what must be excluded.
The '271 Patent specification defines the predefined gene set in Table 1, which lists a collection of genes identified through the inventors' research as having prognostic significance in prostate cancer. The predefined gene set contains substantially more than 22 genes. The claim requires that the panel used for expression measurement include "at least 22" genes chosen from this larger set, but does not require that all genes in the panel be members of the predefined set. A person of ordinary skill in the art would understand that gene-expression panels commonly include genes beyond those of primary predictive interest — for example, reference genes, quality-control genes, or genes of secondary interest — without altering the fundamental character of the panel.
Intrinsic Evidence Supporting GenomeDx's Proposed Construction:
(a) '271 Patent specification, Table 1, listing the predefined gene set (containing more than 22 genes);
(b) Claim language using "at least 22 genes selected from" — the phrase "at least" is open-ended and establishes a minimum, not a maximum;
(c) The use of the open-ended transitional phrase "comprising" in the preamble of Claim 1; and
(d) Dependent Claims 3 and 5, which add further gene-related limitations, confirming that the independent claim sets a minimum threshold rather than a closed requirement.
2. Defendant Helios's Proposed Construction
Proposed Construction: "A collection of exactly or more than 22 genes, all of which must be selected exclusively from the predefined gene set identified in the patent specification, such that the panel does not measure genes outside the predefined set for purposes of the classifier."
Basis for Helios's Proposed Construction:
Helios contends that the phrase "selected from a predefined gene set" imposes a selection constraint: the genes in the panel must be drawn from, and limited to, the predefined gene set. The word "from" in the phrase "selected from a predefined gene set" indicates the source of the genes and, by implication, restricts the panel to genes within that source set. Under Helios's construction, a panel that includes genes outside the predefined set is not a panel "selected from" the predefined set, because some of its genes were selected from outside the set.
Helios further argues that, as a factual matter, this construction has direct implications for infringement. ProstaSignal™ uses a panel of 24 genes, only 19 of which overlap with the predefined gene set in the '271 Patent specification. Because 5 of the 24 ProstaSignal™ genes are drawn from outside the predefined gene set, the ProstaSignal™ panel is not a panel "selected from" the predefined gene set. Moreover, only 19 — not 22 — of the ProstaSignal™ genes come from the predefined set, which means the "at least 22 genes selected from" requirement is numerically unsatisfied regardless of whether the panel may include additional genes.
Intrinsic Evidence Supporting Helios's Proposed Construction:
(a) '271 Patent specification describing the specific predefined gene set and its derivation through the inventors' research; and
(b) Claim language requiring that genes be "selected from" the predefined set — the preposition "from" denotes the source of the selection and limits the panel to genes within the set.