Research map / Probabilistic and scientific modeling / Molecules and proteins

MAPE-PPI: Towards Effective and Efficient Protein-Protein Interaction Prediction via Microenvironment-Aware Protein Embedding

ICLR 2024 · Probabilistic and scientific modeling › Molecules and proteins · approach: Protein fitness and variant effect prediction · 10 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MAPE-PPI, which learns a large fine-grained microenvironment codebook via a VQ-VAE variant on sequence-structure protein graphs, pre-trains it with masked codebook modeling, and uses frozen codebook-based protein embeddings with a GIN for scalable PPI type prediction.

Problem

Structure-based PPI prediction is accurate but computationally expensive and does not scale to millions of PPIs, while sequence-based methods are efficient but less accurate; existing experimental microenvironment vocabularies are too small and coarse.

Approach

Learn a large discrete microenvironment codebook (ego-subgraph residue contexts) with a VQ-VAE variant, pre-train it by masking codebook entries (MCM) rather than input features or codes, then freeze the encoder and codebook as an off-the-shelf embedder so PPI prediction is decoupled from protein encoding.

Key results

• MAPE-PPI reports micro-F1 of 90.42 on SHS27k Random and 73.15 on SHS27k DFS, outperforming HIGH-PPI (86.23 and 70.24) and the best pre-trained baseline KeAP.
• Average rank 1.11 across the 10 settings among pre-trained methods, versus 2.78 for KeAP and GearNet-Edge.
• Reports roughly 50x less training time than HIGH-PPI on SHS27k and scalability to million-scale PPI datasets.

Measured effects reported in the paper

Surprising findings

• ESM-1b and ProBERT fail to match vanilla MAPE-PPI, which uses no additional unlabeled data, on several splits.
• Fine-tuning the pre-trained encoder gives only a small gain (+0.39%) over frozen embeddings while being computationally prohibitive (ESM-1b over 10 hours vs ~300s).

Limitations

• Only predicts PPI types; extension to interaction interfaces or multi-protein conformations is left open.
• Fine-tuning pre-trained encoders is too computationally expensive to evaluate fully at large scale, so the frozen setting is the main focus.

Experimental setup

Datasets: STRING SHS27k SHS148k CATH 4.2

Benchmarks: SHS27k SHS148k STRING

Metrics: micro-F1 AUPR training time

Baselines: DPPI DNN-PPI PIPR GNN-PPI SemiGNN-PPI HIGH-PPI ProBERT ESM-1b GearNet-Edge KeAP

Builds on: VQ-VAE HIGH-PPI GIN AttMask GraphMAE

Closest papers by meaning

• Learning to Predict Mutational Effects of Protein-Protein Interactions by Microenvironment-aware Hierarchical Prompt Learning (ICML 2024 · similarity 0.87)
• Learning to design protein-protein interactions with enhanced generalization (ICLR 2024 · similarity 0.85)
• PRING: Rethinking Protein-Protein Interaction Prediction from Pairs to Graphs (NeurIPS 2025 · similarity 0.85)
• PPDiff: Diffusing in Hybrid Sequence-Structure Space for Protein-Protein Complex Design (ICML 2025 · similarity 0.85)
• MMSite: A Multi-modal Framework for the Identification of Active Sites in Proteins (NeurIPS 2024 · similarity 0.84)
• PSC-CPI: Multi-Scale Protein Sequence-Structure Contrasting for Efficient and Generalizable Compound-Protein Interaction Prediction (AAAI 2024 · similarity 0.84)