ToxiFusion: Integrating Contextual and Compositional Sequence Representations for Toxic Peptide Prediction

Abstract Motivation Accurate prediction of toxic peptides from sequence information is important for peptide safety assessment and therapeutic development. However, reported performance of computational predictors can be affected by representation limitations and evaluation bias. Results We present ToxiFusion, a leakage-controlled framework integrating protein language model-derived contextual representations with interpretable sequence descriptors for toxic peptide prediction. ToxiFusion combines ProtBERT embeddings with amino acid composition and dipeptide composition features using an XGBoost classifier. Preprocessing, model selection, and threshold optimization used training-derived information only. On the locked ToxTeller independent test set, ToxiFusion achieved a Matthews correlation coefficient of 0.718. External evaluation on an overlap-audited ToxIBTL cohort achieved a Matthews correlation coefficient of 0.835 after removing exact training-sequence overlaps. Ablation analyses supported incremental contributions from the representation blocks, while feature-attribution analyses showed broadly consistent patterns for contextual and compositional features across evaluation cohorts. ToxiFusion provides a reproducible framework for toxic peptide prediction and highlights the importance of leakage-controlled validation. Availability and implementation Source code, frozen model artifacts, processed data, and analysis outputs are available at https://github.com/ameneayyoubi-hub/ToxFusion.