Genome-wide Prediction of Small Molecule Binding to Remote Orphan Proteins Using Distilled Sequence Alignment Embedding

Abstract

Endogenous or surrogate ligands of a vast number of proteins remain unknown. Identification of small molecules that bind to these orphan proteins will not only shed new light into their biological functions but also provide new opportunities for drug discovery. Deep learning plays an increasing role in the prediction of chemical-protein interactions, but it faces several challenges in protein deorphanization. Bioassay data are highly biased to certain proteins, making it difficult to train a generalizable machine learning model for the proteins that are dissimilar from the ones in the training data set. Pre-training offers a general solution to improving the model generalization, but needs incorporation of domain knowledge and customization of task-specific supervised learning. To address these challenges, we develop a novel protein pre-training method, DIstilled Sequence Alignment Embedding (DISAE), and a module-based fine-tuning strategy for the protein deorphanization. In the benchmark studies, DISAE significantly improves the generalizability and outperforms the state-of-the-art methods with a large margin. The interpretability analysis of pre-trained model suggests that it learns biologically meaningful information. We further use DISAE to assign ligands to 649 human orphan G-Protein Coupled Receptors (GPCRs) and to cluster the human GPCRome by integrating their phylogenetic and ligand relationships. The promising results of DISAE open an avenue for exploring the chemical landscape of entire sequenced genomes.

DeepREAL: A Deep Learning Powered Multi-scale Modeling Framework Towards Predicting Out-of-distribution Receptor Activity of Ligand Binding

“…To answer Q1, the same model architecture is trained on the same IID and OOD settings using four procedures: 1) from total scratch without any pretraining, i.e., Stage 3 only, 2) going through Stage 1 whole Pfam pretraining but not the Stage 2 binary DTI classification pretraining, which is equivalent to the DISAE model[2], 3) going through only Stage 2 but not Stage 1, and 4) complete three stage pretraining/fine-tuning as DeepREAL. As shown in Figure 2 on the evaluation cross three classes (no-binding, agonist, antagonist), the model without any pretraining (i.e., only Stage 3) has the worst performance.…”

“…There are four data sets involved: Pfam used to pretrain protein descriptor [17]; GLASS data [3] with a large GPCR protein-ligand binding binary data set; agonist/antagonist data are downloaded from the International Union of Basic and Clinical Pharmacology/British Pharmacological Society (IUPHAR/BPS) Guide to Pharmacology; Opioid receptor activity data from [19]. Protein descriptor pretraining exactly follows DISAE[2] hence is not explained in details here. In brief, DISAE builds up a distilled triplet sequence dictionary for the whole Pfam proteins based on multiple sequence alignments (MSA).…”

“…Intuitively, DeepREAL leverages large data sets to hierarchically inform predictions on the receptor activity whose labeled data are scarce along a three-stage end-to-end pipeline as illustrated in Figure 1B. By pretraining protein sequences on whole Pfam in Stage 1, DeepREAL equips itself with genome-scale protein representations that capture novel relationships between proteins beyond sequence homology [2]. By pretraining on a large scale of binary DTI data in Stage 2, DeepREAL builds knowledge of chemical-protein interactions which is the initial step in the ligand binding event.…”

“…Finally, we integrate chemical fingerprint model, sequence embedding model, and DTI latent representation model to train an end-to-end deep learning model for predicting the receptor activity of ligand binding using limited data. In the rigorous benchmark studies that simulate real-world applications, DeepREAL significantly improves the generalization ability in the OOD setting compared with the state-of-the-art methods[19][2]. Data and code is available on github https://github.com/XieResearchGroup/DeepREAL.…”

“…The information on the receptor activity following the ligand binding will fill in a critical knowledge gap in correlating DTIs to clinical outcomes. Although a great deal of efforts have been devoted to predict genome-wide DTIs using deep learning [2], few large-scale experimental and computational studies have been able to specify the receptor activity of ligand binding, i.e., the functional selectivity of the ligand as an antagonist or an agonist.…”