← Projects Molecular Scale

MutaGraph

Graph Neural Network Protein Mutation Effect Predictor

A GAT-based platform that predicts how single-point mutations affect protein stability and antibody-antigen binding affinity, using graph attention over protein structure.

Prediction Targets
Protein Stability ΔΔG Binding Affinity Single-point Mutations
Core Architecture
Graph Attention Network PyTorch Geometric Protein Structure Graph

The Problem

Protein engineering and therapeutic antibody design require understanding how mutations affect protein behavior. Experimental measurement of mutational effects is slow and expensive — a computational model that predicts stability changes (ΔΔG) and binding affinity from structure alone can dramatically accelerate the design cycle.

The Approach

MutaGraph represents proteins as graphs — nodes are residues, edges encode spatial proximity and biochemical interactions. A Graph Attention Network learns to propagate information across the structure, capturing long-range dependencies that sequence-only models miss.

01
Structure Graph Construction

Protein structure parsed into a residue-level graph. Edges defined by spatial distance cutoff and biochemical interaction type. Node features include amino acid physicochemical properties.

02
Graph Attention Network

Multi-head GAT layers learn attention weights over neighboring residues, capturing which structural contacts matter most for stability and binding. Mutation encoded as node feature perturbation.

03
ΔΔG / Affinity Prediction

Graph-level readout predicts the thermodynamic stability change upon mutation, and optionally the change in antibody-antigen binding affinity for therapeutic design applications.

Applications

Protein Engineering

Screen thousands of candidate mutations in silico before committing to expensive wet lab synthesis and assays.

Antibody Design

Predict how mutations in the CDR loops affect antigen binding, guiding affinity maturation campaigns.

Disease Variant Analysis

Assess the functional impact of clinically observed missense variants on protein stability and interactions.

Stack

Model
PyTorch Geometric GAT PyTorch
Data
Protein Structure BioPython PDB
Deployment
FastAPI HuggingFace Spaces Docker