Biophysics-informed computational framework for predicting antisense oligonucleotide knockdown efficacy under small-data constraints

A central strategy in treating neurodegenerative disorders, including amyotrophic lateral sclerosis (ALS), is lowering levels of disease-driving proteins. Antisense oligonucleotides (ASOs) offer a direct RNA-targeted approach to reduce protein expression, but experimental discovery and optimization are constrained by small, heterogeneous datasets and target-specific variability. Computational models that predict ASO knockdown efficacy directly from sequence and context could improve candidate selection, yet generalization across gene families and experimental settings remains challenging. In this study, we present a biophysics-based computational framework for predicting RNase H-active ASO knockdown efficacy under small-data constraints. We curate and integrate heterogeneous literature-derived ASO efficacy measurements into a unified dataset and evaluate model transfer using independently generated experimental ASOs targeting ALS-relevant genes DAZAP1 and FAM69C. Knockdown efficacy is defined at the transcript level (%KD), providing a mechanistic measure of RNase H-mediated activity. To our knowledge, this is the first framework that combines ASO feature engineering with one-shot family calibration for prediction under realistic small-data conditions. We integrate sequence representations with physics-based features capturing RNA-DNA hybridization thermodynamics, targeting RNA accessibility. Compared to sequence-only baselines, this mechanistic feature augmentation reduces external wet-lab prediction error, improving generalization to independent experimental data. Despite these gains, uncalibrated predictions exhibit systematic target-family-specific scale bias. We address this challenge by using a minimal one-shot family calibration strategy that applies an additive offset derived from a single experimentally measured reference ASO per new target family. We further assess the model's behavior using a leave-one-family-out simulation on literature-derived ASO families, repeatedly applying one-shot calibration to quantify its expected benefit and variability under controlled small-n conditions. One-shot calibration markedly improves absolute prediction accuracy while preserving within-family ranking, without retraining or increasing model complexity. Out-of-distribution evaluation shows stable performance for DAZAP1 and increased uncertainty for FAM69C, consistent with the expected effects of very small sample sizes. Permutation-based interpretability analyses show that predictive signal is dominated by global sequence architecture, with thermodynamic and accessibility features providing correlated, secondary contributions. Together, these results establish a computational workflow in which literature-trained models can be transferred to new target genes using minimal experimental input, supporting systematic ASO prioritization in small-data models.