Scouter: Predicting Transcriptional Responses to Genetic Perturbations with LLM embeddings

Abstract

This paper addresses the challenging problem of predicting transcriptional outcomes-- the expression levels of all genes--in gene perturbation experiments and introduces a novel method called Scouter. By leveraging the capabilities of large language models and employing a neural network that facilitates easy training, Scouter overcomes key limitations of current approaches and accurately predicts the outcomes of single-gene and two-gene perturbations, reducing the error of state-of-the-art methods by half or more.