StructRSMA: A Proof-of-Concept Framework for PDB-Derived Contact-Supervised Transfer in RNA-Small Molecule Binding Affinity Prediction

Abstract RNA-small molecule affinity data sets usually provide only pair-level labels, making it difficult for multiview models to learn local nucleotide-atom recognition patterns. We present StructRSMA, a proof-of-concept contact-supervised transfer framework built on the DeepRSMA backbone. StructRSMA preserves the RNA sequence, RNA graph, molecule sequence, molecule graph, cross-fusion, and base affinity components, while introducing PDB-derived nucleotide-atom contact supervision and a Structural Contact Adapter (SCA). In Stage 1, RNA-ligand complexes are converted into binary contact maps using a 4 Å cutoff, and the shared backbone and contact head are trained to rank local contacts. In Stage 2, the contact-pretrained parameters are transferred to R-SIM, where only pKd labels are used; the frozen contact head supplies inferred contact priors to SCA for residual affinity calibration. Under a fixed validation-selected independent-test protocol, the reproduced DeepRSMA baseline obtained PCC 0.387 ± 0.022 and RMSE 0.970 ± 0.012. Contact500 pretraining without SCA yielded PCC 0.418 ± 0.048 and RMSE 0.951 ± 0.023, whereas full StructRSMA reached PCC 0.441 ± 0.037 and RMSE 0.918 ± 0.020. No-pretraining, shuffled-contact, and parameter-matched adapter controls did not reproduce the full improvement. Feature-removal and view-randomization analyses further indicated that contact statistics and multiview context contribute jointly to SCA. Contact prediction showed probability enrichment above contact density, but weak thresholded classification, and improvements became smaller and statistically uncertain under deoverlapped and cold evaluations. StructRSMA therefore provides evidence that PDB-derived contact supervision can support affinity calibration under covered data distributions, while explicitly identifying the limits of structural transfer.