Venom is a mixture of substances produced by a venomous organism aiming at preying, defending, or intraspecific competing resulting in certain unwanted conditions for the target organism. Venom sequences are a highly divergent class of proteins making their machine learning-based and homology-based identification challenging. Prominent applications in drug discovery and healthcare, while having scarcity of annotations in the protein databases, made automatic identification of venom an important protein informatics task. Most of the existing machine learning approaches rely on engineered features, where the predictive model is trained on top of those manually designed features. Recently, transfer learning and representation learning resulted in significant advancements in many machine learning problem settings by automatically learning the essential features. This paper proposes an approach, called ToxVec, for automatic representation learning of protein sequences for the task of venom identification. We show that pre-trained language model-based representation outperforms the existing approaches in terms of the F1 score of both positive and negative classes achieving a macro-F1 of 0.89. We also show that an ensemble classifier trained over multiple training sets constructed from multiple down-samplings of the negative class instances can substantially improve a macro-F1 score to 0.93, which is 7 percent higher than the state-of-the-art performance.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.