AI-Driven RNA Drug Discovery — Helmholtz Institute of Computational Biology

Resources:

• 📑 Report

Summary. In collaboration with the TUM Data Innovation Lab (MDSI) and Helmholtz Munich (Computational RNA Biology Lab), we built and evaluated deep learning models that predict RNA–small-molecule binding affinity directly from RNA sequence and SMILES—bypassing the need for solved RNA 3D structures. We curated and cleaned R-SIM data, designed interpolation and extrapolation splits, and compared sequence-only deep models against the RSAPred baseline.

Highlights

• Encoders: RNA-FM (frozen / LoRA fine-tuned) and a 1D-CNN for RNA; GIN / Graph Diffusion / MolCLR for molecules.
• Combination layers: Concatenation vs cross-attention; pocket-aware pretraining from PDB-derived RNA–ligand interactions.
• Data work: Deduplication, family assignment fixes, and robust split design to test generalization (no RNA overlap in extrapolation).

Key results

• Interpolation (easier): Best MAE ≈ 0.75 with RNA-FM (frozen) + GIN; deep models outperform RSAPred.
• Extrapolation (hard): Best MAE ≈ 1.34 with RNA-FM (LoRA) + Graph Diffusion; still challenging but better than RSAPred and mean-predictor baselines.
• Classification: Fixed-threshold AUROC ≈ 0.5; ranking accuracy ~60% with RNA-FM+MolCLR—better than chance but leaves headroom.
• Pocket pretraining: Helpful for sparsity intuition; no clear downstream gain at current scale.

Method at a glance

1. Encode RNA (RNA-FM / 1D-CNN) and molecules (GIN / Graph Diffusion / MolCLR).
2. Fuse via concatenation or cross-attention; optional pocket-aware pretraining for attention.
3. Head: MLP regressor on pKd; LoRA for parameter-efficient RNA-FM adaptation.