Search

Objective/Background We performed a retrospective analysis of longitudinal real-world data (RWD) from patients with breast cancer to replicate results from clinical studies and demonstrate the feasibility of generating real-world evidence. We also assessed the value of transcriptome profiling as a complementary tool for determining molecular subtypes. Methods De-identified, longitudinal data were analyzed after abstraction from records of patients with breast cancer in the United States (US) structured and stored in the Tempus database. Demographics, clinical characteristics, molecular subtype, treatment history, and survival outcomes were assessed according to strict qualitative criteria. RNA sequencing and clinical data were used to predict molecular subtypes and signaling pathway enrichment. Results The clinical abstraction cohort (n = 4000) mirrored the demographics and clinical characteristics of patients with breast cancer in the US, indicating feasibility for RWE generation. Among patients who were human epidermal growth factor receptor 2-positive (HER2+), 74.2% received anti-HER2 therapy, with ∼70% starting within 3 months of a positive test result. Most non-treated patients were early stage. In this RWD set, 31.7% of patients with HER2+ immunohistochemistry (IHC) had discordant fluorescence in situ hybridization results recorded. Among patients with multiple HER2 IHC results at diagnosis, 18.6% exhibited intra-test discordance. Through development of a whole-transcriptome model to predict IHC receptor status in the molecular sequenced cohort (n = 400), molecular subtypes were resolved for all patients (n = 36) with equivocal HER2 statuses from abstracted test results. Receptor-related signaling pathways were differentially enriched between clinical molecular subtypes. Conclusions RWD in the Tempus database mirrors the overall population of patients with breast cancer in the US. These results suggest that real-time, RWD analyses are feasible in a large, highly heterogeneous database. Furthermore, molecular data may aid deficiencies and discrepancies observed from breast cancer RWD.

Background: Breast cancer demonstrates heterogeneity in biological features, and the therapeutic strategy depends on tumor subtype. African-Ancestry (AA) patients experience a disproportionally high rate of triple negative breast cancer (TNBC) and worse outcomes than European-Ancestry (EA) patients. However, the biological drivers causing this disparity between ancestral populations are not deeply understood. To address the issue, we performed genomic and transcriptomic sequencing of breast tumors for comparison between AA and EA patients according to breast cancer molecular subtypes. Materials and Methods: The study included 221 AA and 341 EA patients. Collected samples underwent the Tempus xT next-generation sequencing panel. Following DNA-panel and whole-transcriptome RNA-sequencing, we compared gene mutation rates, homologous recombination deficiency (HRD) scores, degree of immune infiltration and tumor mutational burden (TMB) between ethnicities and molecular subtypes. Additionally, differences between the activity of relevant signaling pathways were evaluated from RNA-sequencing data. Results: Relative to EA TNBC, AA TNBC tumors exhibited higher mutation rates in TP53 (94% vs 86%), KMT2C (17% vs 9%), APOB (19% vs 10%), BRCA2 (11% vs 5%), EP300 (8% vs 2%), NOTCH1 (12% vs 4%), and EGFR (11% vs 4%). Conversely, AA TNBC tumors had relatively lower rates of PIK3CA (10% vs 18%), RB1 (8% vs 15%), and NF1 (5% vs 11%) mutations. Among patients with HR+/HER2- breast cancer, AA tumors had higher mutation rates in CCND1 (23% vs 10%) and FGF3 (16% vs 10%) than EA tumors, but lower rates in TP53 (32% vs 39%). HRD scores were higher in TNBC and HR-/HER2+ tumors compared with the other subtypes (P Citation Format: Minoru Miyashita, Joshua SK Bell, Yonglan Zheng, Toshio Yoshimatsu, Padma Sheila Rajagopal, Anna Woodard, Jean Baptiste Reynier, Elisabeth Sveen, Galina Khramtsova, Fang Liu, Abiola Ibraheem, Gini Fleming, Nora Jaskowiak, Rita Nanda, Benjamin Leibowitz, Nike Beaubier, Kevin White, Dezheng Huo, Olufunmilayo I Olopade. Comprehensive genomic and transcriptomic profiling of molecular subtypes reveal ancestral differences in the activity of signaling pathways between patients with African and European ancestry [abstract]. In: Proceedings of the 2020 San Antonio Breast Cancer Virtual Symposium; 2020 Dec 8-11; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2021;81(4 Suppl):Abstract nr SS1-04.

Genomic analysis of paired tumor–normal samples and clinical data can be used to match patients to cancer therapies or clinical trials. We analyzed 500 patient samples across diverse tumor types using the Tempus xT platform by DNA-seq, RNA-seq and immunological biomarkers. The use of a tumor and germline dataset led to substantial improvements in mutation identification and a reduction in false-positive rates. RNA-seq enhanced gene fusion detection and cancer type classifications. With DNA-seq alone, 29.6% of patients matched to precision therapies supported by high levels of evidence or by well-powered studies. This proportion increased to 43.4% with the addition of RNA-seq and immunotherapy biomarker results. Combining these data with clinical criteria, 76.8% of patients were matched to at least one relevant clinical trial on the basis of biomarkers measured by the xT assay. These results indicate that extensive molecular profiling combined with clinical data identifies personalized therapies and clinical trials for a large proportion of patients with cancer and that paired tumor–normal plus transcriptome sequencing outperforms tumor-only DNA panel testing. Genomic profiling of cancer samples yields more therapeutic options by including germline, RNA-seq and immunotherapy biomarkers.

Summary: Patient-derived tumor organoids (TOs) are emerging as high-fidelity models to study cancer biology and develop novel precision medicine therapeutics. However, utilizing TOs for systems-biology-based approaches has been limited by a lack of scalable and reproducible methods to develop and profile these models. We describe a robust pan-cancer TO platform with chemically defined media optimized on cultures acquired from over 1,000 patients. Crucially, we demonstrate tumor genetic and transcriptomic concordance utilizing this approach and further optimize defined minimal media for organoid initiation and propagation. Additionally, we demonstrate a neural-network-based high-throughput approach for label-free, light-microscopy-based drug assays capable of predicting patient-specific heterogeneity in drug responses with applicability across solid cancers. The pan-cancer platform, molecular data, and neural-network-based drug assay serve as resources to accelerate the broad implementation of organoid models in precision medicine research and personalized therapeutic profiling programs.

INTRODUCTIONWe performed a retrospective analysis of longitudinal real-world data (RWD) from breast cancer patients to replicate results from clinical studies and demonstrate the feasibility of generating real-world evidence. We also assessed the value of transcriptome profiling as a complementary tool for determining molecular subtypes.PATIENTS AND METHODSDe-identified, longitudinal data were analyzed after abstraction from U.S. breast cancer patient records structured and stored in the Tempus database. Demographics, clinical characteristics, molecular subtype, treatment history, and survival outcomes were assessed according to strict qualitative criteria. RNA sequencing and clinical data were used to predict molecular subtypes and signaling pathway enrichment.RESULTSThe clinical abstraction cohort (n=4,000) mirrored U.S. breast cancer demographics and clinical characteristics indicating feasibility for RWE generation. Among HER2+ patients, 74.2% received anti-HER2 therapy, with ~70% starting within 3 months of a positive test result. Most non-treated patients were early stage. In this RWD set, 31.7% of patients with HER2+ IHC had discordant FISH results recorded. Among patients with multiple HER2 IHC results at diagnosis, 18.6% exhibited intra-test discordance. Through development of a whole-transcriptome model to predict IHC receptor status in the molecular sequenced cohort (n=400), molecular subtypes were resolved for all patients (n=36) with equivocal HER2 statuses from abstracted test results. Receptor-related signaling pathways were differentially enriched between clinical molecular subtypes.CONCLUSIONRWD in the Tempus database mirrors the overall U.S. breast cancer population. These results suggest real-time, RWD analyses are feasible in a large, highly heterogeneous database. Furthermore, molecular data may aid deficiencies and discrepancies observed from breast cancer RWD.

151 Background: The use of PARP inhibitors (PARPi) may trigger synthetic lethality of tumor cells in the context of deficient homologous recombination repair (HRR). Approximately 10-20% of patients with prostate cancer harbor mutations in the HRR pathway, but HRR-associated mutations do not consistently predict the response to PARPi. Considering alternative methods to define Homologous Recombination Deficiency (HRD)—the inability to repair double strand breaks—may aid in identifying additional tumors that are sensitive to PARPi. Here, we evaluate the relationship between HRD status and HRR mutations amongst a large cohort of patients with prostate cancer. Methods: Retrospective analysis of 1,022 de-identified patients with prostate cancer that underwent next generation sequencing (NGS) with the Tempus|xT assay (DNA-seq of 648 genes at 500x coverage, whole-exome capture RNA-seq) was performed. Comparison groups were defined based on HRR alterations—either mono- or bi-allelic alterations of BRCA1 or BRCA2 ( BRCA1/2), ATM, or other HRR pathway genes. HRD status was determined via the Tempus RNA-based HRD algorithm. Results: Among this cohort, mono (-/+) or biallelic (-/-) alterations of HRR genes were found in 432 patients: BRCA1/2 -/- (n = 31), BRCA1/2 -/+ (n = 87), ATM-/- (n = 24), ATM-/+ (n = 67), other HRR-/- (n = 54), other HRR-/+ (n = 169) or no HRR alterations (n = 590). The BRCA1/2-/- group had a higher frequency of Asians (17% vs. < 7% in all other groups) and were diagnosed at younger ages (median 62 years vs. > 65 for all other groups). We identified 130/1022 (13%) patients with prostate cancer to be HRD-positive (HRD+) and observed significant differences in HRD positivity according to the type of HRR alteration observed (Table). Notably, 54% (70/130) of HRD+ patients had no mutations in any genes associated with the HRR pathway. Conversely, 89% (278/314) of patients with non- BRCA1/2 HRR mutations were HRD negative. Amongst all individuals with biallelic loss of any HRR gene, HRD positivity was most enriched for BRCA2-/- (16/29, p < 0.001) and PALB2-/- (3/4, p = 0.082). Conclusions: By using an RNA-based HRD algorithm, we found 13% of patients with prostate cancer are HRD+. This RNA-based HRD signature not only captures patients with HRR mutations but also identifies a substantial population of HRD+ patients who are currently undetectable by methods based solely on sequencing HRR genes. Further research is needed to assess the clinical response to PARPi in this HRD+ population, as well as the response to PARPi in the HRD negative population who harbor HRR gene alterations.[Table: see text]

609 Background: Homologous recombination repair (HRR) is critical for limiting DNA damage arising from double strand breaks. Disrupting HRR has emerged as a therapeutic strategy in pancreatic cancer given the development of PARP inhibitors that can specifically target tumors with homologous recombination deficiency (HRD). While HRD is strongly associated with loss-of-function mutations in BRCA1/2, a broad range of pathogenic alterations have been identified in other HRR genes. Identification of pathogenic alterations that predict response to PARP inhibition remains a critical knowledge gap in the field. Methods: We retrospectively analyzed de-identified records from 895 patients with advanced pancreatic cancer that underwent next generation sequencing (NGS) with the Tempus|xT assay (DNA-seq of 648 genes at 500x coverage, matched normal, full transcriptome RNA-seq). HRD-positivity and genome-wide loss of heterozygosity (LOH) were compared across groups, which were defined based on alterations to BRCA1/2 or other HRR pathway genes. Results: We identified 293/895 (33%) pancreatic cancer patients to be HRD-positive. Among HRD positive patients, 23/895 (2.6%) tumors harbored two alterations in BRCA1/2 (BRCA++), 179/895 (20%) harbored a single alteration in BRCA1/2 (BRCA+), 25/895 (2.8%) harbored two alterations in at least one other non-BRCA HRR gene (other HRR++), 367/895 (41%) harbored a single alteration in one other HRR gene (other HRR+), and 143/895 (16%) had pathogenic alterations in neither BRCA1/2 nor other HRR genes. In terms of HRD status, 23/23 (100%) of BRCA++ cases, 179/337 (53%) of BRCA+ cases, 4/25 (16%) of other HRR++ cases, and 87/367 (24%) of other HRR+ cases were HRD-positive. We observed no cases that were HRD-positive without an alteration in either BRCA1/2 or other HRR gene. Within the other HRR+ group, we found that single alterations in CHEK1/2, FANCA/L, MRE11, PALB2, and RAD51B were all significantly associated with HRD-positivity whereas CDK12, for instance, was not (Table). Conclusions: Comprehensive genomic profiling demonstrates that approximately 1 in 4 patients without BRCA alterations may potentially benefit from PARP inhibition due to alterations in other HRR genes. Future studies may examine the role of PARP inhibition in this population of HRR+ pancreatic cancer patients.[Table: see text]

Background: In patients with early breast cancer, neoadjuvant therapy is widely performed as standard of care. While the molecular targeting strategy makes progress in HER2 positive breast cancer, the optimal regimens for ER+/HER2- breast cancer (BC) and triple negative breast cancer (TNBC) remain undecided. Furthermore, there are few strategies for patients who do not achieve complete pathological response (pCR) who have worse prognosis. To address these unmet clinical needs, there is urgent need to examine the genomic alterations and immune microenvironment of residual tumors after neoadjuvant therapy. Here, we conducted a comprehensive analysis integrating genomic, transcriptomic, and clinical data to investigate the difference between primary breast cancer and residual disease. Materials and Methods: Using the large prospectively ascertained ethnically diverse Chicago Multi-Ethnic (ChiMEC) cohort of 562 participants with integrated genomic data, we identified 176 patients with breast cancer who underwent sequencing with Tempus xT next-generation sequencing panel including DNA- and whole-transcriptome RNA-sequencing. These included 131 primary breast tumors and 45 residual tumors after neoadjuvant therapy. We compared mutation rates between primary tumor and residual tumor in ER+/HER2- BC and TNBC. We also investigated homologous recombination deficiency (HRD) scores, tumor mutational burden (TMB), degree of immune infiltration, and microsatellite instability (MSI), and their association with survival. Results: Out of the 176 patients, there were 72 ER+/HER2- BC, 42 TNBC, and 44 HER2 positive cases. Among patients with HR+/HER2- BC, residual tumors had higher mutation rates in PIK3CA (61% vs 33%), CDH1 (33% vs 17%), CCND1 (39% vs 7%), FGF3 (28% vs 0%), FGF4 (33% vs 2%), FGF19 (33% vs 2%), and GATA3 (44% vs 7%) than primary tumors, but lower rates in TP53 (28% vs 48%), MAP3K1 (6% vs 40%), KMT2D (0% vs 33%), MCL1 (0% vs 30%), SPEN (6% vs 20%), ZFHX3 (0% vs 20%), ARID1A (11% vs 16%), KMT2C (0% vs 19%), LZTR1 (0% vs 19%), BCORL1 (6% vs 17%), NOTCH3 (6% vs 17%), FAT1 (0% vs 17%) and LPR1B (0% vs 17%). Relative to primary TNBC tumors, residual TNBC tumors exhibited higher mutation rates in PTEN (10% vs 3%), CCNE1 (10% vs 3%), CIC (10% vs 0%), and KMT2D (10% vs 3%). Conversely, residual TNBC tumors had relatively lower rates of MCL1 (0% vs 21%), RB1 (0% vs 12%), CDH1 (0% vs 12%), KMT2C (11% vs 18%), and PIK3CA (0% vs 15%), CDKN1B (0% vs 12%) and ETV6 (0% vs 12%). There was a significant trend of higher TMB (>5.0 mutations/megabase, m/MB) associated with improved disease-free survival among the 176 patients (P=0.031). TMB was not significantly different between primary tumors and residual tumors, or across subtypes. Microsatellite instability (MSI) status was high in 3 patients (1.6%) and equivocal in 5 patients (2.7%). TMB in MSI-high or MSI-equivocal tumors was significantly higher than TMB in tumors with microsatellite stability (37.7m/MB vs 5.3m/MB, P Citation Format: Minoru Miyashita, Masaya Hattori, Yonglan Zheng, Toshio Yoshimatsu, Joshua SK Bell, Padma Sheila Rajagopal, Anna Woodard, Jean Baptiste Reynier, Elisabeth Sveen, Galina Khramtsova, Fang Liu, Abiola Ibraheem, Gini Fleming, Nora Jaskowiak, Rita Nanda, Benjamin Leibowitz, Nike Beaubier, Kevin White, Dezheng Huo, Olufunmilayo I Olopade. Unstable mutational profile and heterogeneity of residual breast tumor following neoadjuvant therapy from comprehensive genomic and transcriptomic sequencing [abstract]. In: Proceedings of the 2020 San Antonio Breast Cancer Virtual Symposium; 2020 Dec 8-11; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2021;81(4 Suppl):Abstract nr PD7-06.

Background: Tumors of unknown origin account for up to 5% of newly diagnosed cancers and the average survival time is 9 to 12 months from diagnosis. Establishing tumor type and subtype guides standard of care treatment for several NCCN targeted therapy guidelines. Methods: Targeted DNA sequencing for more than 500 cancer-associated genes and exome-capture RNA sequencing was carried out in more than 25,000 fresh frozen or paraffin embedded tumor samples, including both primary and metastatic tumors. Mutations, copy number variants, and viral sequences were detected from DNA sequencing while gene expression and fusion events were determined from RNA sequencing. We aimed to predict cancer type by utilizing multiple machine learning models trained on individual data types and harmonize predictions across multiple data types. Results: The transcriptome model predicts more than 60 unique diagnoses covering both solid and hematological cancers with >90% overall accuracy on a held-out test set. Of note, the model can accurately predict 10 subtypes of sarcoma and 6 subtypes of neuroendocrine tumors. Gene expression and splicing were the most informative data types, but a performant DNA-only model was also evaluated for application when only DNA data is available. Finally, we evaluated the model on an unlabeled cohort of poorly differentiated samples with inconclusive diagnosis. Conclusions: The incorporation of multiple modes of omics data can improve the interpretability and robustness of machine learning models to predict cancer diagnosis. Citation Format: Jackson Michuda, Benjamin Leibowitz, Shlomit Amar-Farkash, Crystal Bevis, Alessandra Breschi, Joshuah Kapilivsky, Catherine Igartua, Joshua S. Bell, Kyle A. Beauchamp, Kevin White, Martin Stumpe, Nike Beaubier, Timothy Taxter. Multimodal prediction of diagnosis for cancers of unknown primary [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 5423.

Transcriptome profiling of tumors by standard RNA-seq measures the average effect of the tumor microenvironment. This expression profile is largely shaped by the tumor architecture where tumor purity can directly influence the genomic interpretation and consequent associations with clinical outcomes. This challenge is most pronounced in metastatic cancers, where the primary tumor and non-cancerous background tissue have distinct expression profiles. We developed a novel machine learning method, Correlated Composition AdMixture model (CoCoAMix), that learns to distinguish tumor and background transcriptomes and deconvolve metastatic mixtures. Our method first performs differential expression analysis to automatically select the genes that best discriminate between cancer and adjacent-normal transcriptomes. We then apply an admixture model that uses the selected genes to estimate tumor purities and builds prototypical cancer and adjacent-normal transcriptome-wide profiles. Our model utilizes pure primary cancer samples and pure background tissue samples, but does not require further information about the metastatic cancer samples. Finally, CoCoAMix produces sample-specific tumor and adjacent-normal deconvoluted transcriptomes which approximately reconstitute the observed bulk RNA. We optimized and validated this approach on 1,166 liver metastases for six different cancer types. The most frequent liver metastases were from colorectal (28.6%), pancreatic (26.6%), and breast (23.4%) cancers. Our RNA-based tumor purity estimates were correlated with pathology-based and DNA-based estimates in all six cancer types. Pure primary cancer samples inputted into CoCoAMix as negative controls were correctly identified as high-purity, with very small errors introduced by deconvolution. We also validated our approach by deconvoluting synthetic metastatic samples produced by mixing read counts. We found that deconvolution of metastatic cancers resulted in substantial changes to their expression profiles. Low-dimensional embeddings of metastases revealed that low-purity metastases cluster near normal liver transcriptomes, but deconvolution shifted these samples to cluster among primary samples of their same type while maintaining metastatic signal. Differential expression analyses between pre- and post-deconvolution transcriptomes also uncovered a large number of differentially-deconvoluted genes. Furthermore, we discovered that many gene-gene correlations arise from tumor purity, and that this confounding is reduced in deconvoluted transcriptomes. Integration of deconvolution into the Tempus Origin (tm) tumor of unknown primary prediction pipeline resulted in classification accuracy of 87% in liver metastases. These findings demonstrate a robust and versatile method that controls for tumor-adjacent tissue contributions, an absolute necessity when using RNA-seq data to inform clinical decisions. Citation Format: Calvin McCarter, Benjamin Leibowitz, Jackson Michuda, Joshua S. Bell, Catherine Igartua, Kevin P. White. Transcriptome background tissue correction in metastatic cancers using a correlated composition admixture model [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 5481.

Laboratories conducting high volumes of RNA sequencing must be extremely wary of technical batch effects if samples are to be compared across extended time periods, which is imperative for the most well-powered analyses of cancer transcriptomes. Changes in reagents, protocols, or technologies used in nucleic acid extraction, library preparation, and sequencing can alter transcriptomes in ways that invalidate or complicate comparisons of samples from different batches, necessitating continuous monitoring. This monitoring can be particularly difficult when analyzing samples from distinct tissue sites as tumor type is the major biological determinant of transcriptome variance in cancer. Brain and liver cancer transcriptomes, for example, are expected to differ so drastically that their comparison is not informative for batch effect detection. Detection methods must also be robust to disparate batch effects that can manifest as minor changes in expression among many genes or major changes in a subset of genes making ad hoc detection unfeasible. To overcome these challenges, we developed MaCoBED (matched cohort batch effect detection), a novel method that evaluates technical batch effects in a set of transcriptome samples (e.g., a flow cell) by pooling them with a set of validated reference samples matched by cancer type and tissue site. This pooled set of transcriptomes is then subjected to low-dimensional embedding using Uniform Manifold Approximation and Projection (UMAP), and each component is tested for deviation from the reference set using a Wilcox test. Matching new and legacy samples by cancer type and tissue site ensures that any differences in UMAP clustering are not driven by known biological contributions. We found that UMAP was preferable to Principal Components Analysis (PCA). UMAP can capture variability in just two dimensions, accentuating modest but consistent transcriptome differences among batches that would otherwise be manifested among multiple minor principal components, making batch effects more obvious and readily detectable. This approach was able to detect a number of simulated batch effects with high specificity and sensitivity relative to randomly sampled validated legacy samples. Thus, we propose MaCoBED as a simple and rapid approach for batch effect monitoring of high-throughput RNA sequencing datasets that is versatile in detecting distinct kinds of batch effects, easily automatable, readily interpretable upon visualization, and extensible to small or large batch sizes. Citation Format: Joshua Drews, Joshua Bell, Wesley Munson, Saksham Saini, Benjamin Leibowitz, Jackson Michuda, Calvin McCarter, Lee Langer, Catherine Igartua, Kevin White. Robust detection of sequencing batch effects in RNA through low dimensional embedding with subtype-matched reference samples [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr 5466.

e15664 Background: Homologous recombination deficiency (HRD) is the primary biomarker for sensitivity to PARP inhibitors, but identifying the genetic and transcriptomic characteristics that fully capture all HRD patients has remained difficult. For example, DNA-based approaches are limited to patients with pathogenic mutations and loss of heterozygosity (LOH) events, and can fail to properly classify patients with variants of unknown significance. To capture more dynamic cellular processes that arise immediately upon induction of HRD through silencing or loss of BRCA 1/2, a more integrated approach that includes both RNA and DNA based models is necessary. Methods: Using DNA sequencing we developed a genome-wide LOH score that combines pathogenic mutation status and LOH at the BRCA1/2 loci, and the proportion of bases sequenced in the Tempus xT panel that undergo LOH. We also developed three independent RNA-based models to predict BRCA deficiency: 1) An elastic net transcriptome model to predict DNA-based HRD scores derived from exome and SNP array data for each tumor type represented in TCGA; 2) A logistic model to detect BRCA1 promoter hypermethylation from the transcriptome in TCGA data; 3) A model that leveraged the mSigDB annotated gene sets to conduct single sample gene set enrichment analysis (ssGSEA) on Tempus-sequenced patients, selecting over a hundred gene sets that were predictive of BRCA-deficiency. These 4 features were combined to develop a stacked, linear-regression model to distinguish BRCA-intact from BRCA-deficient patients. Results: We found that the genome-wide LOH score alone is predictive of BRCA deficiency. However, our integrated model was highly accurate at distinguishing between BRCA-intact and BRCA-deficient patients and outperformed any single RNA- or DNA-based model. Using this model, we identified many patients that are likely to respond to PARP inhibitors that would have been overlooked using RNA or DNA-based inferences alone. Conclusions: Our approach highlights the strength of integrating diverse molecular features to refine diagnosis and enable oncologists to deliver the most effective therapies to patients.

e16097 Background: Using gene-expression, consensus molecular subtypes (CMS) divide colorectal cancers (CRC) into four categories with prognostic and therapy-predictive clinical utilities. These subtypes also manifest as different morphological phenotypes in whole-slide images (WSIs). Here, we implemented and trained a novel deep multiple instance learning (MIL) framework that requires only a single label per WSI to identify morphological biomarkers and accelerate CMS classification. Methods: Deep learning models can be trained by MIL frameworks to classify tissue in localized tiles from large ( > 1 Gb) WSIs using only weakly supervised, slide-level classification labels. Here we demonstrate a novel framework that advances on instance-based MIL by using a multi-phase approach to training deep learning models. The framework allows us to train on WSIs that contain multiple CMS classes while further identifying previously undiscovered tissue features that have low or no correlation with any subtype. Identification of these uncorrelated features results in improved insights into the specific tissue features that are most associated with the four CMS classes and a more accurate classification of CMS status. Results: We trained and validated (n = 735 WSIs and 184 withheld WSIs, respectively) a ResNet34 convolutional neural network to classify 224x224 pixel tiles distributed across tumor, lymphocyte, and stroma tissue regions. The slide-level CMS classification probability was calculated by an aggregation of the tiles correlated with each one of the four subtypes. The receiver operating characteristic curves had the following one-vs-all AUCs: CMS1 = 0.854, CMS2 = 0.921, CMS3 = 0.850, and CMS4 = 0.866, resulting in an average AUC of 0.873. Initial tests to generalize to other data sets, such as TCGA, are promising and constitute one of the future directions of this work. Conclusions: The MIL framework robustly identified tissue features correlated with CMS groups, allowing for a more efficient classification of CRC samples. We also demonstrated that the morphological features indicative of different molecular subtypes can be identified from the deep neural network.