DLRNA-BERTa: A transformer approach for RNA-drug binding affinity prediction

RNA-based therapies are a rapidly expanding field, offering treatments for a wide range of diseases, including many rare conditions. To date, 24 RNA therapeutics have received FDA approval, with 131 more in clinical trials, underscoring RNA growing role in modern medicine. In this context, Bidirectional Encoder Representations from Transformers (BERT) models provide a cost-effective and accurate virtual screening strategy for accelerating RNA-targeted drug discovery. These models take RNA FASTA sequences and compound SMILES strings as inputs and generate predicted binding affinities in nanomolar units. In this study, we introduce DLRNABERTa, a RoBERTa based framework combining RNA-BERTa, pretrained on 9.76 million RNA sequences, with ChemBERTa for predicting small molecule RNA interactions. The framework includes six class-specific models, aptamers, repeats, ribosomal RNAs, riboswitches, microRNAs (miRNAs), and viral RNAs, plus a general model for cases where the RNA class is unknown. Proposed DLRNABERTa consistently outperforms existing RNA drug interaction prediction methods. Pearson correlation coefficients achieved are: 0.94 (aptamers), 0.95 (repeats), 0.93 (ribosomal RNAs), 0.94 (riboswitches), 0.95 (viral RNAs), 0.98 (miRNAs), and 0.92 (general model), demonstrating robust performance across RNA classes. Benchmarking against four independent datasets from the ROBIN repository further confirms generalizability. Application of DLRNABERTa to 3,492 approved drugs from the ChEMBL database identified 2,859 compounds with predicted affinities (pKd ≥ 6) across 294 RNA targets. As proof of concept, bleomycin is highlighted, supported by literature evidence of RNA binding activity. A publicly accessible web application is available at https://huggingface.co/spaces/IlPakoZ/DLRNA-BERTa, in alignment with FAIR principles.