NitroXAI: An Interpretable Deep Learning Framework for Human S-Nitrosylation Site Prediction

Abstract

NitroXAI v0.1.1 is an interpretable framework for residue-level S-nitrosylation (SNO) site prediction. The release provides two frozen ensemble predictors for identifying candidate S-nitrosylated cysteineresidues from complete protein sequences: • SNO-CLIM: a lightweight CNN-BiLSTM baseline using handcrafted biochemical and positional features.• NitroXAI: a hybrid model integrating handcrafted features with full-protein contextual ESM-2 residue embeddings through attention-based fusion. Both packages support residue-level prediction, all-cysteine scanning, FASTA and optional UniProt workflows, batch prediction, and Integrated Gradients-based interpretation. The bundled models, fold-specific scalers, ensemble rules, and decision thresholds are frozen from the final training notebooks. Inference does not retrain models, refit scalers, optimize thresholds, or fine-tune ESM-2. NitroXAI is intended for computational prioritization and hypothesis generation. Predictions represent candidate S-nitrosylation sites and require experimental validation. This release contains frozen inference assets and interpretability workflows. The complete training pipeline, processed datasets, and associated materials will be released separately upon publication. The MIT License applies to software and original released assets in this record. Third-party resources, including UniProt retrieval and ESM-2 weights, remain subject to their respective terms.