Abstract 1047: Addressing data disparity in pre-clinical research using a generative approach

Abstract Anti-cancer drug response prediction is of paramount importance to drug development and patient treatment design. However, publicly available drug screening datasets on cell lines and patient-derived organoids or xenografts contain highly unbalanced information with most experiments corresponding to non-responsive drug and tumor combinations. While previous research explored augmenting tumor gene expression for cancer type classification, we target a more challenging problem - a guided generation of the responder genetic makeup for specified drugs and sample populations. To tackle this task, we employ a combination of cutting-edge artificial intelligence techniques - consistency models (CM), biology-informed dynamic neural networks (NN), contrastive learning (CL), and transformers. Recent work from OpenAI has addressed the convergence and speed limitation of generative adversarial networks (GAN) by introducing CMs, a novel class of generative models. This development provides a framework for generating synthetic cell line data under complex, biologically relevant conditions for subsequent analysis and optimization. We have modified the CM architecture for the efficient, one-shot generation of RNA-Seq samples. However, the raw outputs of the model mimic the original data distribution and produce unbalanced responder and non-responder data. We employ CL between RNA-seq samples and compounds to guide the model to a desirable outcome. This component integrates hierarchical biological knowledge from the KEGG database with convolutional filters to achieve parameter efficiency and transformer-based encoding of drugs to match responders’ genetic makeup with the corresponding drug to which it is susceptible. Unlike traditional fixed architectures, RNA-Seq encoder adaptively modifies its topology based on biological relationships in the data and leverages BRITE hierarchical relationships to guide the formation of neural connections in a biologically-informed manner. The proposed method allows for the generation of responder data that can be conditioned on both selected drugs and RNA-Seq samples to produce data with specific traits of demographics of interest. We apply popular clustering algorithms on a combination of real-world data and synthetic samples generated by our model. Multiple metrics, including adjusted random score, adjusted mutual information, and homogeneity score, are used for examining the mixing of real-world and synthetic data. The results demonstrate that ground truth and augmented data are nearly indistinguishable. Citation Format: Oleksandr Narykov, Alexander Partin, Yitan Zhu, Priyanka Vasanthakumari, Thomas Brettin, Varuni Sastry, Rick L. Stevens. Addressing data disparity in pre-clinical research using a generative approach [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 1047.