Predicting Antisense Oligonucleotide Thermodynamics using Deep Learning

Abstract

Antisense oligonucleotides (ASOs) have emerged as a promising approach in medicine for the treatment of diseases associated with disrupted protein production. However, the lengthy drug discovery process poses a significant challenge. This project aimed to address this challenge by developing a baseline model to accurately predict the binding affinity of ASOs to mRNA, thereby expediting the drug development timeline.

Drawing upon the successful deep learning approach by Buterez in DNA hybridization [1], the baseline model was tailored to handle short ASO sequences. Various model architectures and features were explored, and recurrent neural networks (RNNs) based on Buterez’s approach were employed as a benchmark for performance evaluation. The accuracy of the models was assessed based on their ability to predict the ΔΔG°, representing the difference in Gibbs free energy between the ASO sequence and a perfect match.

To optimize model performance, different input embeddings were tested, and architecture modifications were implemented. As a result, the final model achieved a high accuracy of approximately 96% with an error margin of±1.0.

By enabling accurate predictions of ASO binding affinity, this research contributes to streamlining the drug development process and holds promise for the advancement of precision medicine.