Beyond conventional sequence-based encoding: a generalized Machine Learning pipeline for oligonucleotide property prediction

The development of RNA-based therapeutics has significantly expanded the landscape of drug discovery by enabling precise modulation of gene expression. These approaches offer the potential to target previously "undruggable" genes, overcoming limitations inherent to traditional small molecule therapies. However, the rational design of effective oligonucleotides remains challenging partially due to the limitations of existing computational models which rely predominantly on sequence-based descriptors. These models are generally restricted to naturally occurring oligonucleotides and do not account for the chemical modifications that are essential for enhancing pharmacological and pharmacokinetic properties. To address this gap, we propose a novel chemical-based representation of oligonucleotides exploiting SMILES notation and molecular fingerprints, mainly exploited for small molecules. By applying them to both natural and chemically modified oligonucleotides, we trained machine learning (ML) models that achieved predictive performance comparable to state-of-the-art sequence-based approaches. Furthermore, we curated the first-ever collection of chemically modified oligonucleotides from patent literature, that allowed us to highlight the broad applicability and generalizability of our approach. This work lays the foundation for a unified pipeline capable of modeling both natural and modified oligonucleotides, advancing the computational design of next-generation RNA therapeutics.