IN SILICO GENERATIVE DESIGN OF CHEMICALLY MODIFIED RNA SEQUENCES FOR FUNCTIONAL PREDICTION

Mahule Roy, Subhas Roy

Keywords: RNA modifications, epitranscriptomics, generative modeling, conditional variational autoencoder, sequence design, motif conservation, RNA folding

TL;DR: We present a cVAE framework that generates chemically modified RNA sequences in silico, preserving motifs and thermodynamic stability while exploring underrepresented epitranscriptomic sequence space.

Abstract: RNA chemical modifications play a central role in regulating RNA stability, translation, and function. While generative machine learning has been widely applied to canonical biomolecules, generative design of chemically modified RNAs remains largely unexplored. We present a fully in silico framework for conditional generation of RNA sequences with specified epitranscriptomic modifications. Using 987,654 modification sites from RMBase and MODOMICS, we train a conditional variational autoencoder (cVAE) with 32D latent space to model RNA sequence context conditioned on modification type. The model generates diverse (94.3% unique), novel (novelty score: 0.78) sequences while preserving known motifs (similarity: 0.87) and thermodynamic plausibility (ΔMFE: 0.8 kcal/mol, p=0.12). Generated sequences exhibit modification-specific patterns with 92.5% conditional accuracy. Our results demonstrate that generative models can explore underrepresented regions of epitranscriptomic sequence space without experimental data, providing a computational foundation for future RNA modification design.

Presenter: ~Mahule_Roy1

Format: Yes, the presenting author will attend in person if this work is accepted to the workshop.

Funding: Yes, the presenting author of this submission falls under ICLR’s funding aims, and funding would significantly impact their ability to attend the workshop in person.

Submission Number: 10