A Multimodal Human Protein Embeddings Database: DeepDrug Protein Embeddings Bank (DPEB)

TL;DR

DPEB tackles the challenge of representing human proteins by assembling a multimodal embedding bank that combines AlphaFold2 structural features, BioEmbeddings, ESM-2, and ProtVec for 22,043 proteins, integrated into a unified PPI graph with over 11 million interactions. The study benchmarks multiple graph neural networks across three tasks—PPI link prediction, semi-supervised enzyme classification, and protein-family clustering—highlighting that BioEmbeddings consistently provide strong predictive signals, with GraphSAGE delivering top PPI performance (AUROC around 87%), while AlphaFold2 embeddings offer structural insight that benefits other tasks when refined. Key findings show concatenated embeddings improve enzyme classification, and Transformer-based refinement of AlphaFold2 embeddings yields substantially better protein-family clustering and classification, underscoring the value of multimodal and task-specific embedding refinement. DPEB thus offers a scalable, multimodal resource for systems biology, drug discovery, and pathway analysis, enabling cross-embedding comparisons and flexible graph-based analyses on the human interactome.

Abstract

Computationally predicting protein-protein interactions (PPIs) is challenging due to the lack of integrated, multimodal protein representations. DPEB is a curated collection of 22,043 human proteins that integrates four embedding types: structural (AlphaFold2), transformer-based sequence (BioEmbeddings), contextual amino acid patterns (ESM-2: Evolutionary Scale Modeling), and sequence-based n-gram statistics (ProtVec]). AlphaFold2 protein structures are available through public databases (e.g., AlphaFold2 Protein Structure Database), but the internal neural network embeddings are not. DPEB addresses this gap by providing AlphaFold2-derived embeddings for computational modeling. Our benchmark evaluations show GraphSAGE with BioEmbedding achieved the highest PPI prediction performance (87.37% AUROC, 79.16% accuracy). The framework also achieved 77.42% accuracy for enzyme classification and 86.04% accuracy for protein family classification. DPEB supports multiple graph neural network methods for PPI prediction, enabling applications in systems biology, drug target identification, pathway analysis, and disease mechanism studies.

Figures (4)

• Figure 1: Overview of the DPEB pipeline for multimodal protein embedding analysis. Protein IDs and FASTA sequences are extracted from major human protein databases (UniProt, STRING, HuMap, IntAct, HPRD). Embeddings from four diverse sources---ProtVec, ESM-2, AlphaFold2, and BioEmbeddings---are computed and aggregated to build the DPEB (Deep Protein Embedding Benchmark) multimodal dataset. These embeddings can then be integrated into nodes of a protein-protein interaction (PPI) graph, where the edges denote interaction strengths represented by weighted connections. Graph neural networks (GNNs), including GraphSAGE, GCN, GTN, GAT, and GIN with edge weight support, are applied to perform downstream tasks such as supervised link prediction (PPI), semi-supervised node classification (e.g., enzyme vs. non-enzyme), and protein family clustering and classification.
• Figure 2: Distribution of human protein sequence lengths in the DPEB . Most sequences range from 113 to 1228 residues (5th–95th percentile), with the highest frequency observed between 315 and 354 residues.
• Figure 3: t-SNE comparison of protein embeddings with protein families distinguished by color. Each figure shows $2,399$ human proteins from the 17 families listed in Table \ref{['table:protein-family']}. Additional t-SNE visualizations, including comparisons of raw versus refined embeddings across other modalities are provided in the Supplementary Information.
• Figure 4: Snapshot of AWS Open Data page for the DeepDrug Protein Embeddings Bank (DPEB) dataset. The resource provides multimodal human protein embeddings (AlphaFold2, BioEmbeddings, ESM-2, ProtVec) with associated metadata, hosted under the AWS Open Data Sponsorship Program. Available at: https://registry.opendata.aws/deepdrug-dpeb/