Jul 2026· Cancer Research· Vol 86, pp. PR006-PR006· 0 citations
TL;DR
It is demonstrated that machine learning applied to transcriptomic data can uncover novel cancer vulnerabilities and actionable targets in individual tumors, even in the absence of functional screening, which may represent a scalable approach to advance precision oncology in rare and/or under-characterized cancer types.
Abstract
Identifying cancer gene dependencies is essential for nominating new therapeutic targets. However, because most experimental models do not sufficiently represent the full diversity of tumors—especially for rare cancers—it remains challenging to use functional screening on experimental models to infer the dependency landscape of individual tumors. We used machine learning to infer gene dependencies from tumor transcriptional profiles, applying our model to the TCGA (>11000 tumors across 28 lineages), rare cancers (>1,000 samples, including multiple rare kidney cancer subtypes), and >500 previously unscreened cancer cell lines. In addition to validating our approach via recovery of known dependencies previously identified in functional genetic screens, we were able to directly infer drug response and synthetic essential relationships from tumor data, highlighting associations with RB1 inactivation, KRAS mutations, and microsatellite instability. Via dependency prediction, we discovered and validated a shared reliance on oxidative phosphorylation in two previously unscreened rare cancers both driven by TFE3 gene fusions: translocation renal cell carcinoma (tRCC) and alveolar soft part sarcoma (ASPS). We also nominate potentially actionable vulnerabilities across other rare cancers, most of which lack in experimental models, but for which RNA-Seq data from tumors are available. These findings demonstrate that machine learning applied to transcriptomic data can uncover novel cancer vulnerabilities and actionable targets in individual tumors, even in the absence of functional screening. This may represent a scalable approach to advance precision oncology in rare and/or under-characterized cancer types.
Ananthan Sadagopan, Bingchen Li, Jiao Li, Yantong Cui, Riva Deodhar, Di Yang, Yuqianxun Wu, Prathyusha Konda, Christy Biji, Dharma Thapa, Meha Thakur, Cary Weiss, Toni Choueiri, Jaime Cheah, John Doench, Benjamin Drapkin, Srinivas Viswanathan. Target discovery in rare cancers enabled by transcriptome-based virtual CRISPR screening [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Breaking Barriers in the Fight against Rare Cancers; 2026 Jul 18-20; Philadelphia, PA. Philadelphia (PA): AACR; Cancer Res 2026;86(14_Suppl):Abstract nr PR006.
Rare cancers collectively account for ∼25% of all cancer diagnoses yet remain underserved due to limited biological understanding, a lack of preclinical models, and challenges associated with conducting clinical trials. Overcoming these barriers requires both the development of rare cancer research resources and new approaches for identifying therapeutic vulnerabilities from limited patient samples. This presentation will highlight how rare cancer biorepositories, patient-derived tumor models, large-scale polypharmacology datasets, and machine learning approaches can be integrated to accelerate therapeutic discovery and functional precision oncology. A rare cancer biorepository containing cryopreserved patient specimens, molecular profiling data, patient-derived microtumors, and xenograft models provides a foundation for studying tumor biology across diverse rare malignancies and generating clinically relevant experimental systems. To enable systematic therapeutic discovery, we developed KIRHub, a comprehensive functional atlas of FDA-approved kinase inhibitors and oncogenic kinase variants that supports characterization of kinase inhibitor polypharmacology, drug repurposing, and mechanism discovery. This approach identified Polo-like kinase 1 (PLK1) as a therapeutic vulnerability in fibrolamellar carcinoma, an ultra-rare liver cancer driven by the DNAJB1-PRKACA fusion, illustrating how functional and computational approaches can uncover actionable dependencies in rare cancers. Patient-derived microtumor models that preserve key features of the native tumor microenvironment further enable rapid ex vivo drug testing and identification of therapeutic vulnerabilities that may not be captured in conventional cancer cell lines. Building on these models, SmartMatch integrates microtumor screening with machine learning-based prediction of therapeutic response. By leveraging responses to a focused set of compounds, SmartMatch prioritizes treatment options across thousands of drugs while delivering clinically actionable results within seven days. Together, these examples demonstrate how experimental and computational modeling approaches can be combined to accelerate therapeutic discovery, identify actionable vulnerabilities, and expand treatment opportunities for patients with rare and treatment-refractory cancers.
Taran Gujral. Functional Precision Oncology in Rare Cancers: From Biobank to Drug Discovery [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Breaking Barriers in the Fight against Rare Cancers; 2026 Jul 18-20; Philadelphia, PA. Philadelphia (PA): AACR; Cancer Res 2026;86(14_Suppl):Abstract nr IA004.
The function of many genes is still unknown, and conventional driver-discovery methods, which rely on how frequently a gene is mutated, cannot assess genes that are only rarely affected. Here we pair Evo~2-based genome analysis with routine clinical imaging to identify gene--phenotype associations at genome-wide scale. For every somatic mutation across three TCGA cohorts (cRCC=clear cell renal cell carcinoma, HCC=hepatocellular carcinoma, and BC=breast cancer; $n = 340$ total), Evo~2 predicts a severity score, with no task-specific training. Per-gene severity summaries are then correlated with radiomic features extracted from paired tumor segmentations, controlling for total mutation burden. In TCGA-cRCC ($n = 162$), this sweep recovers established renal-cancer drivers and identifies 46 additional genes reaching false discovery rate (FDR) significance absent from curated cancer-gene panels, several of which are Mendelian ciliopathy and cytoskeletal-disease genes. These results demonstrate that pairing a genomic language model with widely available clinical imaging can serve as a hypothesis-free discovery tool for gene--imaging associations invisible to conventional approaches.
Frederik Hauke, Jeremias Krause, P. Wienholt et al.· arXiv.org· 0 citations
The interplay between somatic mutations and copy number alterations influences tumor evolution and prognosis. These alterations are often treated independently, overlooking gene mutant dosage (GMD)—a key property of their interaction. Here we develop a computational framework that infers mutation copy number and multiplicity from targeted sequencing panels without requiring matched normal samples. We derive GMD for over 500,000 mutations across 60,000 pan-cancer samples. By stratifying more than 20,000 patients according to GMD across multiple genes, we identify 46 tumor-type-specific biomarkers predictive of survival, 13 of which were undetectable using binary mutant/wild-type models, 26 were associated with metastatic spread and 20 predicted metastatic tropism. Our method reveals GMD patterns as independent predictors of disease prognosis, metastatic potential and site-specific dissemination across diverse tumor types. This augmented insight into genomic drivers enhances our understanding of cancer progression and metastasis and holds the potential to substantially enhance biomarker discovery. The authors present INCOMMON, an open-source Bayesian inference tool that determines the multiplicity and copy number of driver mutations from tumor sequencing datasets.
N. Calonaci, E. Krasniqi, D. Čolić et al.· Nature Genetics· 0 citations
It is demonstrated that integrated Perturb-seq experiments spanning diverse contexts enable hypotheses about gene function specific to tissue types or cancer subtypes – suggesting large-scale, genome-wide datasets would offer invaluable insight into the highly context-dependent nature of cancer biology.
Samuel Maffa, Isabella Boyle, Lie Ward et al.· bioRxiv· 0 citations
The N-glycosylation module is demonstrated as a recurrent hub of context-dependent genetic interactions and combinatorial CRISPR screening as a scalable approach for identifying therapeutic targets and resistance mechanisms in oncogene-driven cancers.
Subin Kim, Chen-Chu Lin, Sabriyeh Alibai et al.· Clinical Cancer Research· 0 citations
Targeted therapies have transformed the management of lung adenocarcinoma (LUAD), yet further clinical progress is hindered by its molecular complexity. Previous efforts to characterize this heterogeneity have been constrained by small cohorts, limiting the discovery of therapeutic targets. We leveraged a large-scale real-world dataset (RWD) to define a LUAD subpopulation with high unmet clinical need. We integrated gene network modeling and spatial transcriptomics to identify key molecular drivers and functionally validated candidate gene targets using CRISPR screening in subtype-matched patient-derived organoids (PDOs).
De-identified clinico-genomic records from LUAD patients profiled with Tempus xT (DNA) and xR (RNA)-seq assays were analyzed (n=7956). Non-negative matrix factorization applied to gene expression data was used to identify molecular subtypes. Progression free survival (PFS), defined as the time from the start of first metastatic therapy to the date of first progression or death from any cause, was assessed in a subset of patients with available data (17.6% of total samples). A classifier trained on tumor-intrinsic features was used to assign subtypes to 36 LUAD PDOs and 19 LUAD tissue samples profiled with the Visium HD spatial transcriptomics platform. A gene regulatory network was modeled using the Priori algorithm by weighting known TF-target associations from the DoRothEA database with xR data. CRISPR/Cas9 screening was performed in PDOs, followed by CellTiter-Glo viability and xR assays to evaluate the functional and transcriptomic impact of gene knockouts.
Six distinct molecular subtypes of LUAD were identified (C1-C6). The C5 subtype (23.15% of total samples) exhibited the worst prognosis (PFS 5.6 months, 95% CI 4.3-7.4) with high prevalence of STK11 (36.85%), KEAP1 (19.67%), and SMARCA4 (15.95%) mutations. Spatial transcriptomics analysis revealed that the C5 subtype was significantly enriched for three distinct niches (localized cellular micro-environments): tertiary lymphoid structure, macrophage-dominated myeloid, and mucinous malignant. Gene co-expression modules associated with the C5 subtype were enriched for NFE2L2 (NRF2) transcription factor pathways. The NFE2L2 gene regulatory network was most transcriptionally active in the C5 subtype. Spatially, NFE2L2 network activity was highest in the mucinous malignant niche and was driven primarily by club cells and malignant cells. Knockdown (KD) of key network nodes (NFE2L2, GPX2, NQO1) in C5 subtype-matched PDOs led to a reversal of the NFE2L2 signature. Additionally, GPX2 KD led to significantly higher viability loss in C5 versus C1 subtype-matched PDOs (p < 0.001).
The integrated multi-omic approach presented here identified GPX2 as a critical vulnerability in a well-defined LUAD subpopulation. These findings provide a rationale for targeting the NFE2L2/GPX2 axis, offering a strategic opportunity for therapeutic intervention.
Akul Singhania, Kayla R. Bastian, Yajas Shah, Swati Kaushik, Brandon L. Mapes, Lee F. Langer, Yaakov E. Stern, Prerna Jain, Ezgi Karaesmen Rizvi, Samantha Cowher, Chi-Sing Ho, Brian Roberts, Nick Callamaras, Richard A. Klinghoffer, Justin Guinney, Radia M. Johnson. Integrated multi-omic analysis and CRISPR screening identify GPX2 as a critical vulnerability in high-risk lung adenocarcinoma [abstract]. In: Proceedings of AACR Drug Discovery and Development (AACR D3) Conference; 2026 Jul 21-24; Boston, MA. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(14_Suppl):Abstract nr A083.
A. Singhania, K. Bastian, Y. Shah et al.· Clinical Cancer Research· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
