Application of Protein Structure Encodings and Sequence Embeddings for Transporter Substrate Prediction

TL;DR

This paper explores using deep learning and protein structure data to better predict what molecules membrane transporters move across cell membranes.

Contribution

The study introduces new deep learning features combining sequence and structure data for improved transporter substrate prediction.

Findings

Deep learning features and FNN models outperformed previous methods in transporter substrate classification.

Structure encodings from FoldSeek and ProstT5 matched the performance of top sequence embeddings like ProtT5-XL.

The approach was tested on sugar and amino acid carriers in A. thaliana and human ion channels with consistent results.

Abstract

Membrane transporters play a crucial role in any cell. Identifying the substrates they translocate across membranes is important for many fields of research, such as metabolomics, pharmacology, and biotechnology. In this study, we leverage recent advances in deep learning, such as amino acid sequence embeddings with protein language models (pLMs), highly accurate 3D structure predictions with AlphaFold 2, and structure-encoding 3Di sequences from FoldSeek, for predicting substrates of membrane transporters. We test new deep learning features derived from both sequence and structure, and compare them to the previously best-performing protein encodings, which were made up of amino acid k-mer frequencies and evolutionary information from PSSMs. Furthermore, we compare the performance of these features either using a previously developed SVM model, or with a regularized feedforward neural network (FNN). When evaluating these models on sugar and amino acid carriers in A. thaliana, as well as on three types of ion channels in human, we found that both the DL-based features and the FNN model led to a better and more consistent classification performance compared to previous methods. Direct encodings of 3D structures with Foldseek, as well as structural embeddings with ProstT5, matched the performance of state-of-the-art amino acid sequence embeddings calculated with the ProtT5-XL model when used as input for the FNN classifier.

Figures