MEMORANDUM
PRIVILEGED AND CONFIDENTIAL — ATTORNEY WORK PRODUCT
TO: Catherine Bloom, Responsible Partner, intellectual Property
FROM: Wendy Schultz, Working Associate, intellectual Property
CC: Peter Vassiliev, Lead Associate; Alan Ngo, Billing Partner
DATE: February 2026
RE: Prior Art Analysis — Invalidity Contentions of Helios Precision Diagnostics, Inc.; Genome Dx Therapeutics Inc. v. Helios Precision Diagnostics, Inc., Case No. 3:25-cv-01847-JWL (N.D. Cal.)
CLIENT: Genome Dx Therapeutics Inc. ("GenomeDx" or "GDT")
MATTER NO.: 1013-00005
Right-click to update Table of Contents
I. Executive Summary
This memorandum analyzes the four prior art references identified by Helios Precision Diagnostics, Inc. ("Helios") in its invalidity contentions, which were served with its Answer and Counterclaims on May 12, 2025, and further elaborated in Helios's responses to GenomeDx's First Set of Interrogatories (Nos. 1–15), served on September 8, 2025. The purpose of this memorandum is to assess the strength and weaknesses of Helios's invalidity theories under 35 U.S.C. §§ 102 and 103 with respect to each of the three GenomeDx patents asserted in this litigation, to identify distinguishing arguments available to GenomeDx, and to recommend strategies for expert testimony, discovery, and claim construction.
Patents at Issue. GenomeDx asserts three patents in this action:
1. U.S. Patent No. 10,438,271 ("the '271 Patent"), titled "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"), titled "Genomic Classifier Systems and Methods for Post-Surgical Cancer Prognosis," issued December 21, 2021.
3. U.S. Patent No. 11,891,647 ("the '647 Patent"), titled "Improved Methods for Genomic Risk Stratification in Prostate Cancer Using Combined Biomarker and Clinical Variable Models," issued February 6, 2024.
Prior Art References Under Analysis. Helios identifies four prior art references in support of its invalidity theories:
1. Nakamura et al., U.S. Patent No. 9,617,340 ("Nakamura"), issued April 4, 2017.
2. Gagnon, WO 2018/065413 ("Gagnon"), published April 12, 2018.
3. Thornberry et al., "Machine Learning Approaches to Genomic Risk Stratification in Prostate Cancer," Journal of Computational Oncology, Vol. 14, No. 3, pp. 287–301 (March 2015) ("Thornberry").
4. Klein et al., "A 17-Gene Assay to Predict Prostate Cancer Aggressiveness," European Urology, Vol. 66, No. 1, pp. 45–52 (2014) ("Klein").
High-Level Conclusions. After detailed analysis, it is my assessment that Helios's invalidity theories under §§ 102 and 103 face significant weaknesses. None of the four references individually anticipates the asserted claims of any of the three patents-in-suit. The obviousness combinations advanced by Helios require substantial hindsight reconstruction, the combination of multiple references addressing disparate aspects of the claimed inventions, and the impermissible assumption that a person of ordinary skill in the art ("POSITA") would have been motivated to select specific elements from each reference and assemble them into the precise claimed methods without the benefit of the inventors' disclosures.
The closest call arises with respect to Claims 1 and 3 of the '271 Patent under § 102 based on Nakamura. Nakamura is the most relevant single reference and was identified by the Patent Office during prosecution of the related '893 Patent. However, distinguishing arguments are available — most critically, differences in gene panel size and composition, the nature of the classifier, and the specific post-radical prostatectomy clinical context — and these arguments proved successful during prosecution of the '893 Patent, where a § 102 rejection based on Nakamura was overcome.
The '893 Patent benefits from a strong prosecution history record in which the Examiner expressly accepted the distinctions drawn over Nakamura, including the "random-forest classifier" limitation and the requirement that the classifier be "trained on a cohort of at least 500 patient samples."
The '647 Patent is best positioned to withstand invalidity attack. Its claims require a combined genomic-clinical variable model — the integration of a genomic risk score with at least two clinical covariates to generate a composite risk metric — which is not disclosed in any of the prior art references, individually or in combination. Moreover, the § 103 rejection of the '647 Patent during prosecution, which was based on the combination of Nakamura and Gagnon, was overcome through a Rule 1.132 declaration from co-inventor Dr. Rajan Mehta establishing unexpected improvements in predictive accuracy from the combined genomic-clinical model relative to genomic-only models.
All three patents benefit from strong secondary considerations of non-obviousness, including the commercial success of GenomeDx's Decipher® test ($387.6 million in FY 2024 revenue; approximately 68–74% market share), evidence of copying (19-of-24 gene overlap between Helios's ProstaSignal™ and GenomeDx's patented gene set), licensing activity (the Veridian Molecular Labs license at a 15% running royalty rate), and the failure of others to develop comparable genomic classifiers despite a long-felt need.
Section 101 Invalidity. This memorandum does not address Helios's separate invalidity theory under 35 U.S.C. § 101 (patent eligibility), which is the subject of a separate memorandum. I note, however, that the Court's January 13, 2026 Order denying Helios's Motion to Dismiss under § 101 — in which Judge Lee held that § 101 challenges were premature without the benefit of full claim construction — is a favorable procedural development for GenomeDx. While the Order does not resolve the substantive patent-eligibility issue, it defers the question to a stage of the litigation where GenomeDx will have the benefit of a developed record and expert testimony.
Independent Development. Helios also contends that its scientists independently developed the ProstaSignal™ classifier beginning in 2020, relying on publicly available gene-expression datasets (principally GEO Series GSE46691 and GSE79957 from the NCBI Gene Expression Omnibus). Independent development is not a defense to patent infringement under U.S. law, but it may be relevant to Helios's arguments regarding willfulness, damages, and, tangentially, obviousness. These issues are addressed in Section X of this memorandum.
II. Overview of the Patents-in-Suit and Asserted Claims
This section provides an overview of each patent-in-suit, including its filing and priority dates, named inventors, prosecution history, and the specific claims asserted in this litigation. A detailed understanding of each patent's claims and prosecution history is essential to the prior art analysis, as the scope of the claims defines the universe of prior art that is potentially invalidating, and the prosecution history may both support and constrain the arguments available to GenomeDx.
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, Dr. David Petrossian
Claims Asserted: Claims 1, 3, 5, 7, 12, and 18
The '271 Patent is the parent patent in the GenomeDx patent family at issue. It is directed to methods for predicting prostate cancer outcomes — specifically, the likelihood of clinical recurrence following radical prostatectomy — using multi-gene expression classifiers. The specification describes the development and validation of a genomic classifier that measures the expression levels of a defined gene panel in prostate tumor tissue and applies a trained machine-learning algorithm to generate a numerical risk score, which is then used to classify patients into risk categories (e.g., low, intermediate, or high risk of clinical recurrence).
Key Independent Claim 1 recites:
A method for prognosing the likelihood of clinical recurrence following radical prostatectomy in a prostate cancer patient, the method comprising:
(a) obtaining a biological sample containing tumor RNA from the patient;
(b) measuring expression levels of a panel of at least 22 genes selected from a predefined gene set in the biological sample;
(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.
Dependent Claim 3 adds the limitation that the biological sample is a formalin-fixed, paraffin-embedded (FFPE) tissue sample, and that the expression levels are normalized using a reference gene panel prior to application of the machine-learning classifier.
Dependent Claim 5 further specifies that the machine-learning classifier is selected from the group consisting of a random-forest classifier, a support vector machine, and a gradient-boosted decision tree ensemble.
Dependent Claim 7 specifies that the risk score is a continuous numerical score between 0 and 1.
Dependent Claim 12 specifies that the predefined gene set comprises genes involved in cell cycle progression, stromal response, and androgen receptor signaling pathways.
Dependent Claim 18 is a method claim that further requires generating a written report setting forth the risk score and risk category and transmitting the report to the referring physician.
Prosecution History. The '271 Patent issued after two Office Actions. In the first Office Action, the Examiner rejected all claims under 35 U.S.C. § 101 on the ground that the claims were directed to the abstract idea of correlating gene expression data to cancer prognosis without reciting significantly more than the judicial exception, citing Mayo Collaborative Services v. Prometheus Laboratories, 566 U.S. 66 (2012), and Alice Corp. v. CLS Bank International, 573 U.S. 208 (2014). The Examiner did not raise any prior art rejections under §§ 102 or 103 in either Office Action.
In response to the § 101 rejection, GenomeDx's prosecution counsel amended the claims to add specificity regarding the technical steps performed on the biological sample, including the requirement that tumor RNA be obtained from the sample and that expression levels be measured for a panel of at least 22 genes from a predefined gene set. Prosecution counsel argued that the amended claims recite a specific, practical application of genomic data analysis tied to concrete laboratory steps — specifically, obtaining a biological sample, extracting and measuring RNA expression, and applying a defined computational method — rather than an abstract concept. The Examiner found these amendments persuasive and withdrew the § 101 rejection. No prior art rejections were sustained at any point during prosecution.
It is significant that the '271 Patent uses the broader term "trained machine-learning classifier" in its independent claims, rather than limiting the classifier to a specific type such as a random forest. This broad language has implications for both the infringement and invalidity analyses. On the infringement side, the broader term encompasses ProstaSignal™'s ensemble gradient-boosted classifier. On the invalidity side, the broader term lowers the bar for prior art references to meet this limitation — any reference disclosing a machine-learning classifier applied to gene expression data for cancer prognosis could potentially satisfy the element, depending on how the term is construed.
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
Priority Date: June 14, 2016 (through continuity with the '271 Patent)
Expiration Date: June 14, 2036 (same priority chain as '271 Patent)
Named Inventors: Dr. Rajan Mehta, Dr. Carolyn Xu
Claims Asserted: Claims 1, 2, 9, 14, and 20
The '893 Patent is a continuation of the '271 Patent and shares a common specification with the parent. However, the '893 Patent's claims are narrower in certain material respects, reflecting amendments made during prosecution to overcome a prior art rejection based on Nakamura.
Key Independent Claim 1 recites:
A method for predicting post-surgical cancer prognosis comprising:
(a) normalizing gene-expression data obtained from a formalin-fixed, paraffin-embedded (FFPE) 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.
Dependent Claim 2 adds that the reference gene panel comprises at least three housekeeping genes.
Dependent Claim 9 specifies that the random-forest classifier comprises at least 100 decision trees.
Dependent Claim 14 specifies that the method further comprises comparing the numerical prognostic score to a predetermined threshold to classify the patient as having high or low metastasis risk.
Dependent Claim 20 specifies that the gene-expression data comprises expression levels of at least 22 genes selected from a predefined gene set.
Prosecution History. The '893 Patent issued after one Office Action. During prosecution, the Examiner rejected the original claims under 35 U.S.C. § 102(a)(1) as anticipated by Nakamura et al. (U.S. Patent No. 9,617,340). This rejection is of central importance to the invalidity analysis because it reflects the Patent Office's determination that Nakamura was the closest prior art, and the basis on which GenomeDx overcame the rejection establishes the distinctions that GenomeDx drew — and the Examiner accepted — between the claimed invention and Nakamura's disclosure.
GenomeDx's prosecution counsel overcame the § 102 rejection by amending the claims to specifically recite (i) a "random-forest classifier" rather than a generic machine-learning classifier, and (ii) training on "a cohort of at least 500 patient samples." Prosecution counsel argued that Nakamura does not disclose either of these features. Specifically, counsel argued that Nakamura describes a penalized regression model (elastic net regularization), not a random-forest classifier, and that Nakamura's training cohort comprised only 235 patient samples — well below the 500-sample minimum required by the amended claims. The Examiner found these distinctions persuasive and allowed the claims.
Critical Distinguishing Features: The '893 Patent's claims contain three principal limitations that differentiate them from the prior art: (1) the specific requirement of a "random-forest classifier" (as opposed to any machine-learning classifier or statistical method); (2) the requirement that the classifier be "trained on a cohort of at least 500 patient samples"; and (3) the requirement of normalization using a "reference gene panel" applied to data from FFPE tissue. These limitations are central to the invalidity analysis and are likely to be the subject of claim construction disputes.
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, Dr. Priya Shah
Claims Asserted: Claims 1, 4, 6, 11, and 15
The '647 Patent represents the next generation of GenomeDx's patented technology, directed to methods that integrate genomic risk scores with clinical covariates to produce a more accurate composite risk assessment. Unlike the '271 and '893 Patents, which focus on genomic-only classification, the '647 Patent claims a combined genomic-clinical model that incorporates established clinical parameters — such as Gleason score, PSA level, surgical margin status, and pathological stage — alongside the genomic risk score to generate a composite risk metric.
Key Independent Claim 1 recites:
A method for generating an integrated prognostic report for a prostate cancer patient, the method comprising:
(a) extracting RNA from a prostate tumor specimen and performing expression profiling to measure expression levels of a plurality of genes;
(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 generate a composite risk metric; and
(d) transmitting the integrated prognostic report comprising the composite risk metric to a healthcare provider.
Dependent Claim 4 specifies that the classifier is a random-forest classifier or gradient-boosted ensemble.
Dependent Claim 6 specifies that the combining step comprises applying a logistic regression model or Cox proportional hazards model that integrates the genomic risk score with the clinical covariates.
Dependent Claim 11 specifies that the composite risk metric stratifies the patient into one of at least three risk categories.
Dependent Claim 15 specifies that the integrated prognostic report further includes a comparison of the composite risk metric to population-based benchmarks derived from a validation cohort.
Prosecution History. The '647 Patent issued after three Office Actions, reflecting a more contentious prosecution than its predecessors. The most significant rejection during prosecution was a § 103 rejection combining Nakamura and Gagnon, in which the Examiner took the position that it would have been obvious to a POSITA to combine Nakamura's genomic risk stratification methods with Gagnon's multi-gene prognostic assay methodology to arrive at the claimed combined genomic-clinical model.
GenomeDx overcame this rejection through a combination of claim amendments and a Rule 1.132 declaration submitted by co-inventor Dr. Rajan Mehta. Dr. Mehta's declaration presented data establishing that the combined genomic-clinical model achieved a statistically significant improvement in predictive accuracy — as measured by the area under the receiver operating characteristic curve (AUC) — over genomic-only models. Specifically, Dr. Mehta's declaration showed that the combined model achieved an AUC of 0.88 for predicting metastasis within 5 years, compared to an AUC of 0.78 for the genomic-only model and an AUC of 0.72 for clinical variables alone. Dr. Mehta attested that these improvements constituted unexpected results that would not have been predicted by a POSITA based on the prior art, because the prevailing assumption in the field was that genomic classifiers would be most accurate when used independently, without the introduction of potentially confounding clinical variables. The Examiner accepted the declaration and allowed the claims.
In addition, GenomeDx filed a terminal disclaimer over the '271 and '893 Patents to overcome a provisional nonstatutory double-patenting rejection.
Critical Distinguishing Features: The '647 Patent's claims are differentiated from the prior art by three principal features: (1) the combined genomic-clinical model, requiring the integration of a genomic risk score with at least two clinical covariates; (2) the generation of a "composite risk metric" that reflects this integration (as opposed to merely reporting genomic and clinical results separately); and (3) the requirement that the integrated prognostic report be transmitted to a healthcare provider. These features represent a fundamentally different inventive concept from the genomic-only classifiers described in the prior art.
III. Detailed Analysis of Prior Art Reference 1: Nakamura et al. (U.S. Patent No. 9,617,340)
A. Overview of Nakamura
Title: "Gene Expression Profiling for Cancer Prognosis"
Issue Date: April 4, 2017
Named Inventors: Toshihiro Nakamura, Jun Watanabe, Hideki Yamamoto
Assignee: Osaka Genomics Research Institute Co., Ltd.
Nakamura is the most significant prior art reference identified by Helios and is the focus of Helios's anticipation theories. Nakamura was cited by the Patent Office during prosecution of the '893 Patent and served as the basis for a § 102 rejection that was overcome, as described in Section II.B above.
General Disclosure. Nakamura is directed to methods for gene expression profiling in cancer prognosis. The specification describes a methodology for measuring expression levels of selected genes in tumor tissue samples and using computational analysis to predict clinical outcomes, including cancer recurrence and progression. The patent discloses a gene panel comprising 18 marker genes and 4 reference genes (22 genes total), a method for normalizing expression data using the reference genes, and the application of a penalized regression model (specifically, elastic net regularization) to generate a prognostic index.
Gene Panel. Nakamura discloses a panel of 18 marker genes selected based on their association with cancer prognosis, along with 4 reference (housekeeping) genes used for normalization purposes. The total count of 22 genes is significant because the '271 Patent's Claim 1 requires "a panel of at least 22 genes selected from a predefined gene set." The question is whether Nakamura's 22 genes (18 marker + 4 reference) satisfy the "panel of at least 22 genes" limitation — a question that depends on claim construction of the term "panel." If "panel" encompasses both marker genes and reference genes, Nakamura may meet the numerical requirement. If "panel" is construed to include only the marker genes whose expression levels are used as inputs to the classifier (excluding housekeeping genes used only for normalization), Nakamura's 18-gene marker panel falls short of the 22-gene threshold.
Classifier. Nakamura describes the use of a penalized regression model, specifically elastic net regularization, to analyze gene expression data and generate a prognostic index. This model is a statistical method that combines L1 (lasso) and L2 (ridge) regularization to perform variable selection and regularization. The question is whether elastic net regularization constitutes a "trained machine-learning classifier" within the meaning of the '271 Patent's claims. This is a claim construction question with significant implications. While elastic net regularization involves a form of supervised learning (the model is trained on labeled data to minimize a loss function), it is a linear statistical method that is traditionally distinguished from the non-linear, ensemble methods (such as random forests, support vector machines, and gradient-boosted trees) that are typically associated with the term "machine learning" in the computational oncology literature.
Sample Preparation. Nakamura describes obtaining tissue samples from surgical specimens and extracting RNA for gene expression analysis. The specification refers to both fresh-frozen and FFPE tissue, though the primary embodiments appear to utilize fresh-frozen tissue. The FFPE-specific methods described in the '271 and '893 Patents — including the particular normalization protocols required for degraded RNA from FFPE samples — are not described with the same specificity in Nakamura.
Clinical Context. Nakamura's specification describes the application of the gene expression profiling methodology to cancer prognosis generally, with specific examples drawn from breast cancer and colorectal cancer. The specification mentions prostate cancer only in passing, in a paragraph listing potential applications. Nakamura does not specifically describe the post-radical prostatectomy clinical context, nor does it describe prognosing the likelihood of clinical recurrence following radical prostatectomy, as required by the '271 Patent's claims.
Training Cohort. Nakamura describes training the elastic net model on a cohort of 235 patient samples (180 breast cancer and 55 colorectal cancer samples). This cohort size is well below the 500-patient minimum required by the '893 Patent's claims.
B. Helios's Anticipation Argument (§ 102) — '271 Patent Claims 1 and 3
Helios contends that Nakamura anticipates Claims 1 and 3 of the '271 Patent under 35 U.S.C. § 102(a)(1). Helios's theory, as set forth in its Invalidity Contentions and supplemented in its interrogatory responses, is that Nakamura discloses each and every element of Claim 1 of the '271 Patent.
Element-by-Element Analysis:
(a) "Obtaining a biological sample containing tumor RNA from the patient."
Nakamura discloses obtaining tissue samples from surgical specimens and extracting RNA for gene expression analysis. Nakamura's specification states: "Total RNA was extracted from tissue specimens using standard extraction protocols." Nakamura's disclosure of RNA extraction from tissue samples appears to satisfy this element. However, Nakamura does not specify that the RNA is "tumor RNA" as opposed to RNA from normal adjacent tissue or a mixed sample; the specification describes tissue specimens generally without specifying tumor microdissection or tumor-enrichment protocols. Whether this distinction is material will depend on claim construction of "tumor RNA" — if the term is given its plain meaning (RNA derived from tumor tissue), there is an argument that Nakamura's general tissue samples do not necessarily satisfy the limitation, as the quality and specificity of the RNA source is a relevant technical distinction.
(b) "Measuring expression levels of a panel of at least 22 genes selected from a predefined gene set in the biological sample."
This is the most contested element. As discussed above, Nakamura discloses a total of 22 genes — 18 marker genes and 4 reference genes. Helios argues that Nakamura's 22 total genes satisfy the "panel of at least 22 genes" limitation. GenomeDx can counter that the 4 reference genes are housekeeping genes used solely for normalization and are not part of the "panel" of genes whose expression levels are measured for prognostic purposes. Under this construction, Nakamura's 18-gene marker panel is insufficient.
Additionally, the claim requires genes "selected from a predefined gene set." The '271 Patent's specification defines the predefined gene set in Table 1, which comprises 46 candidate genes associated with prostate cancer biology — including genes involved in cell cycle progression, stromal response, and androgen receptor signaling. Nakamura's 18 marker genes were selected for breast and colorectal cancer prognosis and are drawn from a different biological context. A comparison of Nakamura's marker genes with the '271 Patent's Table 1 gene set reveals that only 7 of Nakamura's 18 marker genes appear in the '271 Patent's predefined gene set. This overlap is insufficient to meet the limitation that the genes be "selected from a predefined gene set" as defined in the patent specification.
(c) "Applying a trained machine-learning classifier to the measured expression levels to generate a risk score."
Helios argues that Nakamura's elastic net model qualifies as a "trained machine-learning classifier." As discussed above, this is a claim construction question. Elastic net regularization is a form of regularized linear regression that involves supervised learning — the model parameters are estimated by minimizing a penalized loss function on training data. In the broadest sense, any supervised learning method could be characterized as "machine learning." However, the '271 Patent's specification describes the classifier in terms of ensemble methods (random forests, support vector machines, gradient-boosted trees), and the prosecution history suggests that the applicant intended the term "machine-learning classifier" to denote something more than simple regression. During prosecution, GenomeDx's counsel argued that the claims "recite specific computational steps performed by a trained machine-learning classifier" to overcome the § 101 rejection, invoking the complexity and non-linearity of the claimed classifier as evidence of a specific technical implementation.
Moreover, Nakamura's elastic net model generates a "prognostic index" — a continuous score. Whether this "prognostic index" constitutes a "risk score" within the meaning of the claims is a question of claim construction, but there appears to be minimal basis for distinction on this element alone.
(d) "Classifying the patient into a risk category based on the risk score."
Nakamura describes dichotomizing patients into "high-risk" and "low-risk" groups based on a threshold applied to the prognostic index. This appears to satisfy the limitation of classifying the patient into a risk category based on the score. However, it is noteworthy that Nakamura uses a binary classification (two categories), whereas the '271 Patent's specification contemplates multi-category classification (e.g., low, intermediate, and high risk). This distinction is unlikely to be dispositive because Claim 1 broadly recites "a risk category" without specifying the number of categories.
Claim 3 Analysis. Dependent Claim 3 adds the limitations that (i) the biological sample is an FFPE tissue sample and (ii) the expression levels are normalized using a reference gene panel prior to application of the machine-learning classifier. While Nakamura mentions FFPE tissue in passing, its primary embodiments use fresh-frozen tissue, and Nakamura does not describe the specific FFPE-optimized normalization protocols required for reliable gene expression measurement from degraded FFPE-derived RNA. This is a technically significant distinction. RNA extracted from FFPE tissue is typically degraded, fragmented, and chemically modified, requiring specialized normalization approaches to produce reliable expression measurements. The '271 Patent's specification describes specific protocols for FFPE RNA extraction and normalization that address these challenges, including the use of a defined reference gene panel comprising genes with stable expression across a range of FFPE tissue conditions. Nakamura's specification does not provide this level of detail for FFPE tissue processing, and its normalization approach is designed for the higher-quality RNA typically obtained from fresh-frozen specimens.
C. Distinguishing Arguments Against Nakamura Anticipation
Based on the element-by-element analysis above, GenomeDx has the following distinguishing arguments available to defeat Helios's anticipation theory:
1. Gene Panel Size and Composition. The strongest distinguishing argument is that Nakamura's gene panel does not satisfy the "at least 22 genes selected from a predefined gene set" limitation. This argument has two components. First, Nakamura's 18 marker genes fall short of the 22-gene minimum if reference genes are excluded from the count — and there are strong reasons to exclude them, as the reference genes are used for normalization purposes and are not prognostic markers. Second, even if all 22 of Nakamura's genes are counted, only 7 overlap with the '271 Patent's predefined gene set (Table 1), meaning the genes are not "selected from" the predefined gene set as the claim requires. GenomeDx should seek a claim construction that defines "predefined gene set" by reference to Table 1 in the specification, which would effectively foreclose Nakamura as anticipatory art on this element.
2. Classifier Type. Nakamura's elastic net model is a penalized linear regression technique, not a non-linear ensemble machine-learning classifier of the type described in the '271 Patent's specification. GenomeDx should argue — and seek a claim construction — that "trained machine-learning classifier" encompasses methods such as random forests, support vector machines, and gradient-boosted decision trees (all specifically listed in dependent Claim 5), but does not encompass simple regularized linear regression. This construction is supported by the specification, which describes the machine-learning classifier in terms of these non-linear methods and does not discuss elastic net or other penalized regression techniques.
3. Post-Radical Prostatectomy Clinical Context. The '271 Patent's Claim 1 is specifically directed to "prognosing the likelihood of clinical recurrence following radical prostatectomy." Nakamura does not describe this clinical context. Nakamura's examples are drawn from breast cancer and colorectal cancer, and although the specification mentions prostate cancer in a list of potential applications, it does not describe the post-prostatectomy clinical scenario or the specific types of clinical outcomes (clinical recurrence, metastasis) associated with that context. This is not merely a difference in the patient population but a fundamental difference in the clinical question being addressed — post-surgical prognosis involves different biological processes and clinical endpoints than the initial diagnostic or staging context.
4. FFPE-Specific Methods (Claim 3). For dependent Claim 3, GenomeDx can argue that Nakamura does not disclose the FFPE-specific sample preparation and normalization methods required by the claim. Nakamura's passing reference to FFPE tissue does not constitute a disclosure of the specific normalization protocols necessary for reliable expression profiling from degraded FFPE-derived RNA.
Prosecution History Support. It is notable that the Examiner did not raise any prior art rejection during prosecution of the '271 Patent, despite having access to Nakamura and other references. While the absence of a rejection is not dispositive — the Examiner may not have identified Nakamura during the '271 prosecution — it suggests that the Examiner did not view Nakamura as anticipatory of the '271 claims. Moreover, the fact that the Examiner's § 102 rejection of the later-filed '893 Patent based on Nakamura was overcome by distinguishing the random-forest classifier training methodology and minimum cohort size reinforces the conclusion that the Examiner recognized material differences between Nakamura and the GenomeDx inventions.
D. Nakamura in the Context of '893 Patent Claims
Nakamura was the basis for a § 102 rejection during prosecution of the '893 Patent, and that rejection was the most significant prior art event in the prosecution history of the three patents at issue. The Examiner initially rejected the '893 Patent's claims as anticipated by Nakamura, but prosecution counsel overcame the rejection by amending the claims to add two specific limitations: (i) the classifier must be a "random-forest classifier" (not merely any computational method), and (ii) the classifier must be "trained on a cohort of at least 500 patient samples."
Nakamura does not disclose either of these features. Nakamura's classifier is an elastic net regularized regression model, which is fundamentally different from a random forest in terms of its algorithmic structure, the way it processes features, and its ability to capture non-linear interactions. A random forest is an ensemble method that constructs multiple decision trees trained on bootstrapped samples and produces predictions by aggregating the trees' outputs (majority vote for classification, averaging for regression). By contrast, elastic net is a linear model that assumes a linear relationship between features and the response variable, with regularization to prevent overfitting. These are categorically different approaches.
Additionally, Nakamura's training cohort of 235 patients is less than half the minimum 500-patient threshold required by the '893 Patent's claims. The 500-patient minimum is not an arbitrary number; it reflects the inventors' finding that the random-forest classifier requires a sufficiently large and diverse training set to avoid overfitting and to achieve clinically meaningful predictive accuracy. The specification of the '271 and '893 Patents describes validation studies demonstrating that classifiers trained on fewer than 500 patients exhibited significantly higher variance and lower AUC in independent validation cohorts.
Conclusion: Nakamura does not anticipate the '893 Patent's claims. The Examiner's acceptance of the distinctions drawn during prosecution provides strong evidentiary support for this conclusion. Helios would need to persuade the Court that the Examiner erred — a difficult argument given the clear factual differences between Nakamura's elastic net model and the claimed random-forest classifier.
E. Nakamura in the Context of '647 Patent Claims
The '647 Patent's claims add a dimension not present in Nakamura: the combined genomic-clinical model requiring integration of a genomic risk score with at least two clinical covariates selected from Gleason score, PSA level, surgical margin status, and pathological stage to generate a composite risk metric.
Nakamura does not disclose this integration. Nakamura's specification describes a gene-expression-only prognostic model; clinical variables are not incorporated into the classifier or the output. There is no mention in Nakamura of Gleason score, PSA levels, surgical margins, pathological staging, or any other clinical covariates being combined with the genomic data to produce a composite metric. Nakamura does not generate a "composite risk metric" as required by the '647 Patent's claims, nor does it contemplate the transmission of an integrated prognostic report to a healthcare provider.
Conclusion: Nakamura does not anticipate the '647 Patent's claims. The combined genomic-clinical model is a fundamental feature of the '647 Patent that is entirely absent from Nakamura's disclosure.
F. Overall Assessment of Nakamura
Nakamura is the strongest single prior art reference in Helios's invalidity case, but it has identifiable gaps with respect to each patent's claims. The primary risk is to Claims 1 and 3 of the '271 Patent under § 102, where Nakamura presents the closest call. However, distinguishing arguments are available on multiple elements — gene panel size and composition, classifier type, and clinical context — and these arguments are supported by the prosecution history of the '893 Patent.
The risk to the '893 Patent is low because the Examiner already considered and rejected the anticipation argument based on Nakamura, accepting the distinctions drawn by GenomeDx's prosecution counsel. The risk to the '647 Patent is very low because Nakamura does not disclose the combined genomic-clinical model.
I recommend that Dr. Franklin Ogawa address Nakamura's limitations in detail in his expert report, with particular attention to the differences in gene panel composition, classifier architecture, clinical context, and training methodology. Dr. Ogawa's report is due May 15, 2026, and this analysis should be a priority.
IV. Detailed Analysis of Prior Art Reference 2: Gagnon (WO 2018/065413)
A. Overview of Gagnon
Title: "Multi-Gene Prognostic Assays for Urological Cancers"
Publication Date: April 12, 2018
Applicant: Centre de Recherche du CHU de Québec (Québec, Canada)
Named Inventor: Marc-André Gagnon
Gagnon is a PCT (Patent Cooperation Treaty) international patent application published by the World intellectual Property Organization (WIPO). It was not published as a U.S. patent or patent application, but it qualifies as prior art under 35 U.S.C. § 102(a)(1) as a printed publication available to the public before the effective filing date of the claims to which it is applied.
Prior Art Status Relative to Each Patent:
• '271 Patent (filed June 14, 2016): Gagnon was published on April 12, 2018, which is after the '271 Patent's filing date. Under pre-AIA law, Gagnon would not be prior art to the '271 Patent. Under AIA § 102(a)(1), Gagnon is prior art because it was publicly available before the '271 Patent's issue date, and the relevant question is whether it was available before the effective filing date. Because the '271 Patent was filed on June 14, 2016, and Gagnon was published on April 12, 2018 — after the filing date — Gagnon is not prior art to the '271 Patent under § 102(a)(1) unless GenomeDx cannot establish an effective filing date prior to Gagnon's publication. Since the '271 Patent's filing date of June 14, 2016 predates Gagnon, Gagnon does not qualify as prior art for the '271 Patent. This significantly limits Gagnon's utility to Helios.
• '893 Patent (filed March 22, 2019; claims priority to '271 Patent family, effective date June 14, 2016): Gagnon was published before the '893 Patent's actual filing date of March 22, 2019, but after its claimed priority date of June 14, 2016. The question is whether the '893 Patent's claims are entitled to the benefit of the June 14, 2016 priority date for purposes of prior art analysis. Because the '893 Patent is a continuation of the '271 Patent and shares a common specification, the '893 Patent's claims should be entitled to the benefit of the June 14, 2016 priority date provided that the claims are supported by the '271 Patent's specification under 35 U.S.C. § 112. If so, Gagnon is not prior art to the '893 Patent. We should confirm that each asserted claim of the '893 Patent has adequate written description support in the '271 Patent's original specification.
• '647 Patent (filed September 9, 2021): Gagnon was published on April 12, 2018, well before the '647 Patent's filing date. The '647 Patent does not claim priority to the '271 Patent family (it has a new priority chain). Gagnon therefore qualifies as prior art to the '647 Patent under § 102(a)(1).
General Disclosure. Gagnon describes multi-gene prognostic assays for urological cancers, including bladder cancer, kidney cancer, and prostate cancer. The specification discloses methods for measuring the expression levels of selected gene panels in tissue samples from patients diagnosed with urological cancers and using the expression data to generate prognostic information. Gagnon describes several gene panels of varying sizes, including panels of 15, 24, and 31 genes, for different urological cancer types. For prostate cancer specifically, Gagnon describes a 24-gene panel, though the genes in this panel differ substantially from those in GenomeDx's patented gene set — a comparison reveals only 11 genes in common between Gagnon's 24-gene prostate panel and the '271 Patent's Table 1 gene set.
Gagnon describes the use of statistical methods for prognostic scoring, including Cox proportional hazards regression and Kaplan-Meier survival analysis, but does not describe machine-learning classifiers such as random forests, support vector machines, or gradient-boosted trees. Gagnon does not describe a combined genomic-clinical variable model; the prognostic methods disclosed in Gagnon are based exclusively on gene expression data, without integration of clinical covariates such as Gleason score, PSA level, surgical margin status, or pathological stage.
Gagnon addresses "urological cancers" broadly and does not specifically focus on the post-radical prostatectomy clinical context. The prostate cancer examples in Gagnon appear to address pre-treatment prognosis in the context of biopsy specimens, rather than post-surgical recurrence prediction.
B. Helios's Obviousness Theory — Gagnon in Combination with Nakamura
Helios asserts that the asserted claims of all three patents are obvious under 35 U.S.C. § 103 in view of Nakamura in combination with Gagnon. Helios's theory, as articulated in its invalidity contentions, is that a POSITA would have been motivated to combine Nakamura's gene expression profiling methodology with Gagnon's multi-gene prognostic assay for prostate cancer to arrive at the claimed inventions.
The key question is what Gagnon adds to Nakamura. Nakamura describes gene expression profiling with a 22-gene panel (18 marker + 4 reference) and an elastic net classifier, but is primarily directed to breast and colorectal cancer. Gagnon describes multi-gene prognostic assays specifically for urological cancers, including prostate cancer, with a 24-gene panel. In combination, Helios may argue that Gagnon supplies the prostate-cancer-specific application and the larger gene panel that Nakamura lacks.
However, this combination has several weaknesses:
First, Gagnon does not supply the machine-learning classifier limitation. Gagnon's prognostic methods are based on traditional statistical approaches (Cox regression, Kaplan-Meier analysis), not machine-learning classifiers. Thus, even in combination, Nakamura and Gagnon do not disclose a "trained machine-learning classifier" (for the '271 Patent) or a "random-forest classifier" (for the '893 Patent) applied to gene expression data in the prostate cancer context.
Second, Gagnon does not supply the combined genomic-clinical variable model required by the '647 Patent. Gagnon's methods are gene-expression-only, without integration of clinical covariates. Therefore, the Nakamura + Gagnon combination does not satisfy the '647 Patent's requirement for combining a genomic risk score with at least two clinical covariates.
Third, and critically, the Examiner during prosecution of the '647 Patent considered the combination of Nakamura and Gagnon and raised a § 103 rejection based on this combination. That rejection was overcome through Dr. Mehta's Rule 1.132 declaration establishing unexpected results. The Examiner's acceptance of the unexpected-results evidence is part of the intrinsic record and constitutes strong evidence that the Nakamura + Gagnon combination does not render the '647 Patent's claims obvious.
C. Distinguishing Arguments Against Gagnon
GenomeDx has the following arguments available to distinguish Gagnon:
1. Field-of-Use Differences. Gagnon covers "urological cancers" broadly — bladder, kidney, and prostate — without specifically teaching the post-radical prostatectomy prognosis context that is central to the GenomeDx patents. The prostate cancer examples in Gagnon appear to involve biopsy-based analysis rather than post-surgical specimens, which implicates different biological processes and clinical endpoints. A POSITA would recognize that prognosis based on biopsy tissue (pre-treatment) and prognosis based on radical prostatectomy tissue (post-treatment) involve different tumor biology, different gene expression patterns, and different clinical outcomes.
2. Absence of Machine-Learning Classifier. Gagnon does not disclose the use of machine-learning classifiers. Its prognostic methods are based on Cox regression and Kaplan-Meier analysis, which are established survival analysis tools but are not "machine-learning classifiers" as that term is used in the GenomeDx patents and understood in the field. The absence of machine-learning methodology in Gagnon means it cannot supply this critical limitation, either individually or in combination with Nakamura.
3. Absence of Combined Genomic-Clinical Model. Gagnon does not teach the integration of genomic risk scores with clinical covariates. This is the defining feature of the '647 Patent, and Gagnon's failure to teach it means the Nakamura + Gagnon combination is insufficient to render the '647 Patent's claims obvious — a conclusion the Examiner reached during prosecution.
4. Gene Panel Differences. Gagnon's 24-gene prostate cancer panel shares only 11 genes with the '271 Patent's predefined gene set. This limited overlap undermines Helios's argument that a POSITA would have been motivated to select genes from Gagnon's panel for use in the claimed methods, because the specific gene selection is an inventive contribution.
5. Prosecution History of the '647 Patent. The Examiner's § 103 rejection of the '647 Patent based on Nakamura + Gagnon was overcome by Dr. Mehta's Rule 1.132 declaration, which established that the combined genomic-clinical model achieved unexpected improvements in predictive accuracy. This is powerful evidence against Helios's obviousness theory. While prosecution history does not bind a federal court, it is entitled to deference and shifts the burden to Helios to explain why the Examiner's determination should be disregarded.
D. Overall Assessment of Gagnon
Gagnon is primarily relevant as a secondary reference in obviousness combinations — it does not independently anticipate any asserted claim. Its principal contribution to Helios's invalidity theories is to supply a prostate-cancer-specific gene expression panel, but the panel shares limited overlap with the '271 Patent's predefined gene set, and Gagnon does not disclose machine-learning classifiers or combined genomic-clinical models.
Gagnon's prior art status is limited with respect to the '271 and '893 Patents. For the '271 Patent, Gagnon was published after the filing date and therefore does not qualify as prior art. For the '893 Patent, Gagnon is prior art only to the actual filing date, not the claimed priority date; GenomeDx should confirm that the '893 Patent's asserted claims are entitled to the June 14, 2016 priority date.
For the '647 Patent, Gagnon is prior art, and the Nakamura + Gagnon combination was considered during prosecution. The Examiner's acceptance of the unexpected-results declaration creates a strong evidentiary record against the obviousness theory.
The key vulnerability in the Gagnon analysis is that Helios may attempt to challenge the sufficiency of Dr. Mehta's Rule 1.132 declaration — for example, by arguing that the data presented were insufficient, that the improvement was not truly "unexpected," or that the comparison was flawed. GenomeDx should ensure that Dr. Mehta's testimony and the underlying data are consistent between the prosecution record and this litigation, and that Dr. Ogawa's expert report addresses and supports the unexpected-results finding.
V. Detailed Analysis of Prior Art Reference 3: Thornberry et al.
A. Overview of Thornberry
Citation: Thornberry, R.M., Chen, L., and Patel, S.K., "Machine Learning Approaches to Genomic Risk Stratification in Prostate Cancer," Journal of Computational Oncology, Vol. 14, No. 3, pp. 287–301 (March 2015).
Publication Date: March 2015
Prior Art Status: Thornberry was published in March 2015, which predates the filing dates of all three patents-in-suit ('271 Patent filed June 14, 2016; '893 Patent filed March 22, 2019; '647 Patent filed September 9, 2021). Thornberry therefore qualifies as prior art under 35 U.S.C. § 102(a)(1) for all three patents.
General Disclosure. Thornberry is a review and research article published in the Journal of Computational Oncology. The article surveys machine learning approaches applied to genomic data for the purpose of risk stratification in prostate cancer. The authors describe the general framework of applying supervised machine learning algorithms to gene expression datasets to predict clinical outcomes, including biochemical recurrence, metastasis, and cancer-specific mortality.
Thornberry discusses several machine learning methodologies, including random forests, support vector machines (SVMs), neural networks, and gradient-boosted decision trees. For each methodology, the article describes the general principles of the algorithm, its strengths and weaknesses, and its application to genomic data. The article includes a comparative analysis of these methods using a publicly available gene expression dataset (GEO Series GSE21032), reporting performance metrics including AUC, sensitivity, and specificity.
Critically, Thornberry is a methodological review article — it does not disclose a specific, implementable multi-gene classifier for prostate cancer prognosis. The article does not identify a specific gene panel, does not define a predefined gene set, does not describe specific training protocols or cohort requirements, and does not describe sample preparation methods (such as RNA extraction from FFPE tissue or reference-gene normalization). Thornberry operates at a level of generality that is one step removed from the specific, practical methods claimed in the GenomeDx patents.
Level of Specificity. On the critical question of whether Thornberry discloses a specific multi-gene panel of at least 22 genes: it does not. Thornberry describes using publicly available gene expression datasets containing thousands of probe sets, but it does not identify a specific panel of 22 or more genes selected for prognostic purposes. The article's analysis uses the full gene expression datasets rather than a curated, predefined subset.
On the question of whether Thornberry describes a complete method including sample preparation, gene expression measurement, classifier training, and risk classification: it does not. Thornberry describes computational experiments performed on pre-existing gene expression data. The article does not address the laboratory steps of obtaining biological samples, extracting RNA, measuring expression levels, or any of the sample preparation and normalization steps recited in the GenomeDx patents' claims.
On the question of classifier training methodology: Thornberry describes the general principles of random forests and other machine learning algorithms but does not specify training parameters such as the number of trees, the minimum cohort size, or the cross-validation protocols. The article does not describe training a random-forest classifier on a cohort of at least 500 patient samples, as required by the '893 Patent.
B. Helios's Obviousness Theory — Thornberry as Teaching Machine Learning Classifiers
Helios is likely to rely on Thornberry primarily as a supporting reference to supply the "trained machine-learning classifier" limitation of the '271 Patent and, more specifically, the "random-forest classifier" limitation of the '893 Patent. Helios's theory, as best understood from its invalidity contentions, is that Thornberry demonstrates that random forests and other machine learning methods were known in the art and had been applied to genomic data for prostate cancer risk stratification prior to GenomeDx's inventions. In combination with Nakamura (which supplies gene expression profiling methodology) and/or Gagnon (which supplies a prostate-cancer-specific gene panel), Helios argues that it would have been obvious to a POSITA to apply Thornberry's machine learning methodology to Nakamura's or Gagnon's gene expression data to arrive at the claimed inventions.
This theory is essentially a "known method, known problem" argument: the prior art shows that gene expression profiling in cancer prognosis was known (Nakamura, Gagnon) and that machine learning classifiers, including random forests, were known tools for analyzing genomic data (Thornberry); therefore, combining the two to create a machine-learning-based gene expression classifier for prostate cancer prognosis was obvious. This is a facially plausible argument under the KSR framework, which holds that combining known elements to achieve predictable results may support a finding of obviousness.
C. Distinguishing Arguments Against Thornberry
GenomeDx has several strong arguments against Helios's reliance on Thornberry:
1. General Methodological Review vs. Specific Implementation. Thornberry is a review article that discusses the theoretical applicability of machine learning methods to genomic data. It does not disclose a specific, implementable classifier for prostate cancer prognosis. The distinction between knowing that a class of algorithms could theoretically be applied to a type of data and actually developing a specific, validated classifier with defined gene features, training parameters, and clinical endpoints is precisely the kind of inventive contribution that the patent system is designed to protect. General teachings in a review article about the suitability of machine learning methods do not, without more, render obvious a specific implemented classifier with defined gene features, training methodology, and clinical application.
2. No Specific Gene Panel. Thornberry does not identify a specific panel of 22 or more genes selected from a predefined gene set. The article uses publicly available full-transcriptome datasets without curating a specific gene panel. Gene selection — the process of identifying, from thousands of candidate genes, a specific panel of genes that collectively provide clinically meaningful prognostic information — is a critical inventive step that required substantial experimentation and technical judgment. Thornberry does not perform or disclose this step.
3. No Sample Preparation or Normalization Methods. Thornberry's computational experiments are performed on pre-existing gene expression datasets. The article does not address the laboratory steps of obtaining biological samples, extracting RNA from FFPE tissue, or normalizing expression data using a reference gene panel. These are specific technical steps recited in the claims that are absent from Thornberry's disclosure.
4. No Combined Genomic-Clinical Model. Thornberry does not describe integrating genomic risk scores with clinical covariates to generate a composite risk metric, as required by the '647 Patent. Thornberry's analysis is purely genomic.
5. Motivation to Combine. To combine Thornberry with Nakamura and/or Gagnon, Helios must show why a POSITA would have been motivated to select specific machine learning methods from Thornberry's survey, apply them to specific gene expression data from Nakamura's or Gagnon's disclosure, and arrive at the specific claimed methods — with a reasonable expectation of success. The mere fact that machine learning methods existed and gene expression profiling existed does not establish that combining them in the specific manner claimed was obvious. The claim limitations define a specific implementation, not a general concept.
6. Hindsight Reconstruction. The combination of Thornberry with Nakamura and/or Gagnon requires selecting specific teachings from each reference and combining them in a specific way to arrive at the claimed inventions. This kind of multi-reference combination, guided by the claims of the patent rather than by the teachings of the prior art itself, is precisely the type of hindsight reasoning that Graham v. John Deere and KSR warn against. The Supreme Court cautioned in KSR that "[a] factfinder should be aware, of course, of the distortion caused by hindsight bias and must be cautious of arguments reliant upon ex post reasoning." 550 U.S. at 421.
D. Overall Assessment of Thornberry
Thornberry is a supporting reference that provides general background context for the use of machine learning in genomic oncology. It does not independently disclose any complete claimed method. Its primary utility to Helios is to supply the machine-learning classifier limitation — particularly the random-forest classifier for the '893 Patent — as a known technique in the art.
However, Thornberry's value to Helios is limited by its generality. A review article describing the theoretical applicability of a class of algorithms to a type of data is, at best, evidence that the general concept was known. It does not make obvious a specific implementation with defined gene features, training parameters, sample preparation methods, and clinical applications. Reliance on Thornberry for the specifics of GenomeDx's patented methods may be seen by the Court as classic hindsight reasoning — using the patentee's own invention as a template and then searching the prior art for individual elements that can be assembled to reconstruct the invention.