ASO-RASAR: A Read-Across Framework for Predicting Antisense Oligonucleotide Gapmer Activity Across Target Genes

The development of predictive sequence-activity models capable of reliably identifying optimal antisense oligonucleotide (ASO) candidates across diverse target genes remains a central challenge in ASO drug discovery. Here, we curated a patent-derived data set of 59,273 gapmer ASOs (20-mer 5-10-5 MOE and 16-mer 3-10-3 cET) across 30 human genes and benchmarked gene-specific and cross-target prediction models using comprehensive sequence and target-context descriptors. Gene-specific models achieved strong predictive performance, with genomic context and sequence motifs as the most informative features. However, cross-target models evaluated by leave-one-gene-out cross-validation failed to generalize to new target genes, revealing that ASO activity is governed primarily by target-specific determinants. To address this limitation in data-scarce settings, we developed ASO-RASAR, a read-across sequence-activity relationship model that transfers predictive information from data-rich to data-poor target genes. In simulated low-data scenarios, the best ASO-RASAR strategy improved median AUPRC over standard QSAR baselines by up to 22.5%. Experimental validation using WFDC1-targeting ASOs in human bone marrow-derived mesenchymal stem cells supported that ASO-RASAR prediction scores correlated strongly with knockdown efficiency (Pearson's r = 0.83). These findings establish ASO-RASAR as a practical approach for prioritizing ASO candidates against novel targets with limited experimental data.