Benchmarking Large Language Models for Predicting Therapeutic Antisense Oligonucleotide Efficacy

Antisense oligonucleotides (ASOs) are a promising class of therapeutic agents capable of selectively modulating gene expression and treating a wide range of genetic and neurological disorders. Accurate prediction of ASO efficacy is essential for accelerating drug discovery and reducing experimental costs, yet remains a challenging computational task due to complex sequence-function relationships. In this study, we benchmark large language models (LLMs) and molecular embedding-based regression models for predicting therapeutic ASO efficacy across three publicly available biological datasets: PFRED, openASO, and ASOptimizer. We evaluate multiple transformer-based molecular embedding models, including ChemBERTa and MolFormer, alongside prompt-engineered LLM configurations such as GPT-3.5-Turbo, LLaMA-2-7B, and Galactica-6.7B. Our results demonstrate that DNA sequence-based representations combined with gene context outperform SMILES-based molecular representations for efficacy prediction. Among the evaluated approaches, GPT-3.5-Turbo using few-shot prompting achieves strong predictive performance, reaching coefficient of determination (R squared) values up to 0.6381 and substantially outperforming baseline regression models. These findings highlight the potential of general-purpose large language models as effective tools for biomolecular prediction and computational drug discovery. This work provides a systematic benchmarking framework and establishes a foundation for integrating large language models into therapeutic antisense oligonucleotide design pipelines.