Language Model–Based Representation Learning for Venom Protein Identification and Therapeutic Target Discovery in Cancer

Introduction:Venom is a complex mixture of bioactive molecules produced by venomous organisms for predation, defense, or intraspecific competition, often leading to specific physiological responses in target organisms.Venom-derived peptides and proteins have recently garnered attention in biomedical research for their potential therapeutic applications, including the discovery of anticancer drugs.However, venom sequences constitute a highly divergent class of proteins, making their identification using machine learning and homology-based methods particularly challenging.To address this, we propose ToxVec, a transfer learning-based framework for the automatic representation learning of protein sequences, designed to improve venom identification.Materials and Methods: Our approach leverages pre-trained protein language models to capture sequence-level information without manual feature engineering.Results: ToxVec outperforms existing feature-based models, achieving a macro-F1 score of 0.89.Furthermore, an ensemble model trained on multiple balanced subsets achieves a macro-F1 score of 0.93, representing a 7% improvement over the state of the art.Beyond benchmark performance, screening of experimentally validated anticancer peptides from the CancerPPD2 dataset revealed that many exhibit high venom-like signatures according to ToxVec, supporting the notion that toxin-inspired molecular architectures may underlie anticancer bioactivity.Conclusions: We further discuss how language model-based representation learning embodies a Cognitive Mind-Body-Inspired interpretation, linking abstract sequence semantics (the "mind") to biological function (the "body").By enabling more accurate large-scale identification of venom proteins, ToxVec provides a foundation for systematically exploring venom-derived bioactive peptides as potential therapeutic candidates, including those targeting pathways implicated in breast cancer progression and metastasis.This automated approach thus bridges computational protein informatics with translational oncology, supporting future efforts in bioactive peptidebased anticancer research.