Abstract 730: Can we overcome the scarcity of data for building predictive models in cancer?

Abstract Background: A major aim in precision oncology is establishing predictive biomarkers to enable improved treatment matching for patients. mRNA expression is one of the most available and explored modalities to enable generation of such models. Several works have demonstrated that supervised learning using transcriptomic data can generate predictive biomarkers. However, the scarcity of datasets that couple mRNA with drug response severely limit this approach. In contrast, unsupervised models offer a potential for a broadly applicable approach to develop predictive markers. We have previously described ENLIGHT, an unsupervised approach that generates predictive biomarkers to virtually any targeted cancer drug without any drug-specific data requirements. Here, we perform a large-scale analysis, comparing the supervised and unsupervised approaches. We show that ENLIGHT is able to generate predictive markers on a far larger number of drugs than the supervised approach, and that for those drugs where a supervised approach is feasible, ENLIGHT’s performance is comparable. Methods: We established a list of the 114 approved targeted oncology drugs, and systematically searched for available training and validation data. For all 114 drugs, we generated an unsupervised biomarker using ENLIGHT. For drugs where more than one dataset was available, we trained and validated a logistic regression model with LASSO in a leave-one-dataset-out fashion, and compared its performance with ENLIGHT, which was not trained on any of the datasets. For drugs where only one dataset was available, we only tested ENLIGHT. We evaluated both methods by the odds ratio (OR) to respond at a score threshold that yields 50% sensitivity in each dataset. We then averaged these values across datasets to a single value for comparison. Results: Of the 114 drugs tested, only 7 had more than one dataset with mRNA and objective responses available, to allow appropriate supervised training and validation. Thus, while we could generate ENLIGHT biomarkers for all 114 drugs, supervised methods allowed generation of a biomarker for only 7 drugs. We could test the ENLIGHT biomarkers for 17 drugs for which at least one dataset was available for validation, and found it was predictive (OR>1) in 16 of those. When focusing on the 7 drugs for which we could also generate supervised biomarkers, we found that ENLIGHT was roughly on par with the supervised models (aggregated OR of 2, p=0.003 on 6 datasets vs. aggregated OR of 2.52, p=0.0001 on all 7 datasets, respectively). Conclusion: We demonstrate that ENLIGHT can overcome the inherent lack of data for generating predictive models for cancer drugs using a powerful unsupervised approach. While less than 10% of approved targeted cancer drugs have minimal data enabling generation of supervised models, ENLIGHT can generate predictive models for all the targeted drugs, and shows predictive value in almost all cases where validation data exists. Citation Format: Gal Dinstag, Yaron Kinar, Eden Maimon, Tuvik Beker, Ranit Aharonov. Can we overcome the scarcity of data for building predictive models in cancer [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 730.