Abstract 3642: TLSL: Transfer learning for synthetic lethality (SL) experiment design and SL prediction

Abstract Introduction: In a synthetic lethality (SL) gene pair, the cell dies when the function of both genes is disrupted but remains viable when the function of only one of the genes is compromised. While carrying strong therapeutic potential in the treatment of cancer, SL interactions are context specific and sparse. CRISPR-Cas9 based gene combination double knockout (CDKO) experiments are purposely designed to discover novel SL gene pairs. However, none of the published CDKO studies have considered how to use existing SL data to choose gene pair targets and cell lines for new experiments. In this project, for the first time, we present a transfer learning framework for SL prediction, namely TLSL. We anticipate TLSL can illustrate whether and how existing SL data can assist in gene and cell line selection for new CDKO experiments. Methods: Signature-iSL is the foundation model for TLSL. It is designed to be interpretable, perform cell line specific SL prediction, and incorporate pathway-based co-membership of gene pairs and multi-omics signatures. In transfer learning, Signature-iSL shall transfer information from source data (i.e. existing CDKO experiment data) to target data (i.e. data from new CDKO experiment). Our transfer learning covers four different CDKO experiment design scenarios: (1) experiment keeps the same genes, but on a new cell line; (2) experiment keeps the same cell line, but selects different genes; (3) experiment targets a similar mechanism, but selects different cells and genes; and (4) big data (i.e. all other existing SL CDKO experiments in Synthetic Lethality Knowledge Base (SLKB)) to small data (i.e. targeted genes in a cell). TLSL was applied to all datasets under four scenarios in SLKB. It covers 11 studies, 22 cell lines, 6,127 genes, 16,059 SL, and 264,424 non-SL gene pairs. Each computational experiment was repeated five times and evaluated in out-of-sample gene pair prediction; 20% of genes and their gene partners were held out from training. Results: As a standalone model, Signature-iSL outperformed two previous state-of-the-art methods, MVGCN-iSL and MLEC-iSL. While MVGCN-iSL achieved 0.62 and MLEC-iSL achieved 0.65 AUC-ROC, Signature-iSL reached 0.75. On average, across all four scenarios, TLSL improved accuracy, AUC-ROC, and F-1 metrics by 15%, 9%, and 10% respectively. The most successful scenario was big data to small data transfer learning, where accuracy, AUC-ROC, and F-1 metrics improved by 23%, 18%, and 15% respectively. Conclusion: The proposed transfer learning approach, TLSL, is shown to be highly informative and beneficial for selecting genes and cells in new CDKO experiments by borrowing SL data from existing CDKO experiments. Under all four transfer learning scenarios, TLSL has significantly improved SL prediction model performance. Citation Format: Birkan Gökbağ, Lijun Cheng, Lang Li. TLSL: Transfer learning for synthetic lethality (SL) experiment design and SL prediction [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 3642.