Predicting function from sequence in venom peptide families

Toxins from animal venoms are small peptides that recognize specific molecular targets in the brains of prey or predators. Next generation sequencing has uncovered thousands of diverse toxin sequences, but the functions of these peptides are poorly understood. Here we demonstrate that the use of machine learning techniques on sequence-derived features enables high accuracy in the task of predicting a toxin’s functionality using only its amino acid sequence. Comparison of the performance of several learning algorithms in this prediction task demonstrates that both physiochemical properties of the amino acid residues in a sequence as well as noncontiguous sequence motifs can be used independently to model the sequence dependence of venom function. We rationalize the observed model performance using unsupervised learning and make broad predictions about the distribution of toxin functions in the venome. Keywords—Bioinformatics, machine learning, protein function prediction, venomics.