UniDL4BioPep: A universal deep learning architecture for binary classification in peptide bioactivity

Abstract: Developing peptides with bioactive properties or nutraceutical/pharmaceutical benefits from food proteins is gaining increasing interests. Identification of novel peptides through model prediction can reduce benchwork in wet experiments. However, conventional model building process is complicated and time-consuming regarding peptide representation, feature dimensionality, model selection, hyperparameter tuning, etc. Recently, some cutting-edge pre-trained language models (LM) have been released for protein sequence embedding and further applied to structure/function prediction. Based on that, we proposed and developed UniDL4BioPep, a universal deep-learning model architecture that can be self-adaptive to predict any bioactivity of peptides with any length and achieve cutting-edge performance with simplified effort in model development.

All the benchmark datasets were retrieved from published state-of-the-art (SOTA) models in order to conduct a unbiased performance evaluation and comparison. Ten binary (i.e., active and in-active) datasets for different (bio)activities were collected, including ACE inhibition, DPP IV inhibition, antimicrobial, antimalarial, anticancer, quorum-sensing, tumor T cell antigens, anti-methicillin-resistant S. aureus strains, bitterness, and umami activities. The LM evolutionary scale modeling (ESM-2) was employed to generate numerical vectors to represent peptide sequences, which were then used for a convolutional neural network (CNN) model development.

For nine of the ten different bioactivity prediction tasks, the UniDL4BioPep exhibited greater performances than the respective SOTA models, with accuracy, Mathews correlation coefficient, and area under the curve of 1.6-7%, 3.6-26.4%, and 1.8-12.3% higher, respectively. For ACE inhibition prediction, the performances of UniDL4BioPep and SOTA were comparable. Furthermore, t-distributed stochastic neighbor embedding (t-SNE) distribution showed that active and inactive samples were distributed in two clusters for most datasets after being embedded by the pretrained LM. The UniDL4BioPep is the first successful attempt for large-scale evaluation of the biological LM in peptide representation for bioactivity prediction and proving its superiority.