Zero-shot biological reasoning with open-weights large language models reproduces CRISPR screen based prediction of synthetic lethal interactions.

Abstract

Identifying clinically relevant synthetic lethal interactions has great potential for uncovering novel therapeutic vulnerabilities in cancer. Current approaches rely on machine learning models that estimate probabilities of synthetic lethal interactions, without supplying explicit knowledge of the underlying biology and lack the human-readable interpretation leading to the prediction. Large Language Models (LLMs) represent a new class of tools capable of reasoning and leveraging extensive biological knowledge acquired from relevant literature during their pretraining. Here, we tested multiple open-weight LLMs for their ability to predict known and novel synthetic lethal interactions. We found that most of the tested models were better at reconstructing the results of three known genome-wide CRISPR knockout screens than random chance and non-LLM-based methods, and observed that their performance was related to the parameter-size of the model, and on average benefited little from additional pathway and genetic information apart from what they already possess when estimating the likelihood of a synthetic lethal relationship. After selecting the model that offered the best balance of predictive performance and computational efficiency for our use case (Qwen2.5-32B-Instruct, 0.715 AUROC), we performed an in silico screen of 398,277 gene pairs from 893 clinically relevant genes. Our goal was to highlight the potential of open-weights LLMs as scalable, context-aware prioritization tools for synthetic lethal interactions, and to lay the groundwork for predicting higher-order genetic interactions.