Abstract

Predicting gene expression changes in cancer cells following drug treatment remains a critical challenge in oncology. Traditional methods are often blind, time-consuming, and resource-intensive. With the rise of large language models (LLMs) and multi-modal deep learning, there is growing potential to model complex biological responses more effectively. This project aims to discuss existing research progress and proposes a multi-modal framework that fine-tunes large language models to predict transcriptomic outcomes in cancer cell lines after drug exposure. The model integrates drug features—derived from SMILES-based molecular fingerprints and graph structures—with baseline cell line data, including gene expression, protein abundance, and mutation profiles. These are combined with known gene expression perturbations from public databases such as LINCS. A pre-trained LLM, such as DeepSeek, will be adapted using LoRA (Low-Rank Adaptation) to learn the mapping from this rich input space to post-treatment gene expression profiles. Model performance will be evaluated using metrics like Pearson correlation, mean squared error, and biological pathway coherence. Experimental validation will be conducted to assess predictive accuracy. This integrative approach offers a scalable solution to forecast cellular transcriptional responses, potentially advancing drug development strategies.

Bio

Jinming Bai is a first-year PhD student in Cell Biology at the University of Cape Town. He completed his MSc at UCT, focusing on cancer drug responses and transcriptional regulation. He earned his Bachelor's degree in Biotechnology from Shandong Agricultural University in China. His current research integrates multi-modal deep learning to predict gene expression changes in cancer cells following drug treatment.