Source-linked AI summary
Pre-training of Graph Neural Network for Modeling Effects of Mutations on Protein-Protein Binding Affinity
Xianggen Liu, Yunan Luo, Sen Song, Jian Peng
TL;DR
Predicting mutation-induced changes in protein binding affinity matters for protein engineering, but experimental evaluation is labor-intensive and existing engineered features can generalize poorly. GraphPPI pre-trains a GNN to learn structural descriptors and combines them with gradient-boosting trees. It reports state-of-the-art performance across five benchmark datasets and accurate predictions for SARS-CoV-2 antibody complexes.
Problem
Experimental evaluation of mutation-induced affinity changes is labor-intensive, and manually engineered features in existing machine-learning methods may generalize poorly across protein structures.
Source: Introduction · pp. 1–2Method
GraphPPI pre-trains a GNN to reconstruct complexes from disturbed structures, then combines its graphical descriptors with gradient-boosting trees to predict mutation-induced affinity changes.
Source: Introduction · p. 2Results
GraphPPI achieved new state-of-the-art performance across five benchmark datasets for single- and multi-point mutations and accurately predicted affinity changes for SARS-CoV-2 antibody complexes.
Source: Introduction · p. 3Takeaways & Limitations
The authors present GraphPPI as a computational tool with potential applications in antibody design, identifying disease-driver mutations, and studying protein biosynthesis.
Source: Introduction · p. 3
Abstract
from arXiv · showhide
Modeling the effects of mutations on the binding affinity plays a crucial role in protein engineering and drug design. In this study, we develop a novel deep learning based framework, named GraphPPI, to predict the binding affinity changes upon mutations based on the features provided by a graph neural network (GNN). In particular, GraphPPI first employs a well-designed pre-training scheme to enforce the GNN to capture the features that are predictive of the effects of mutations on binding affinity in an unsupervised manner and then integrates these graphical features with gradient-boosting trees to perform the prediction. Experiments showed that, without any annotated signals, GraphPPI can capture meaningful patterns of the protein structures. Also, GraphPPI achieved new state-of-the-art performance in predicting the binding affinity changes upon both single- and multi-point mutations on five benchmark datasets. In-depth analyses also showed GraphPPI can accurately estimate the effects of mutations on the binding affinity between SARS-CoV-2 and its neutralizing antibodies. These results have established GraphPPI as a powerful and useful computational tool in the studies of protein design.
Introduction
Evaluating mutation-induced binding-affinity changes is difficult at scale, and existing approaches face computational, sampling, or feature-generalization limits. GraphPPI combines pre-trained GNN features with gradient-boosting trees and reports strong benchmark and antibody-focused results.
Introduction
- Mutation effects are difficult to evaluate experimentally, while the large space of possible antibody mutants makes much of the search ineffective.The paper focuses on binding affinity, ΔG, and mutation-induced changes in it, ΔΔG.Source: Introduction · pp. 1–2
- Prior approaches trade computational cost, sampling sensitivity, and feature generalization: simulations can be resource-intensive, energy methods sampling-sensitive, and engineered machine-learning features may generalize poorly.Source: Introduction · p. 2
- GraphPPI pre-trains a GNN to reconstruct a complex from a version with randomly rotated residue side chains, then uses its graphical descriptors with gradient-boosting trees to predict affinity changes.Source: Introduction · p. 2
- GraphPPI is designed to learn structural features automatically and use large-scale complex training to capture rules shared across structures, supporting generalization without manual feature engineering.Source: Introduction · p. 2
- Without annotated labels, the pre-trained GNN detected abnormal atom interactions and distinguished interface from non-interface residues.Source: Introduction · pp. 2–3
- Across five benchmark datasets (three single-point, two multi-point), GraphPPI achieved new state-of-the-art performance and accurately predicted binding-affinity differences in SARS-CoV-2 antibody–spike complexes with more mutated residues than in training.Source: Introduction · p. 3
Results
GraphPPI combines structural reconstruction pre-training with GNN descriptors and gradient-boosted prediction. Experiments report state-of-the-art benchmark performance across single- and multi-point mutations, structural diagnostics, and applications to SARS-CoV-2 antibodies.
The GraphPPI framework
- GraphPPI uses GNN descriptors from wild-type and mutant complexes, with gradient-boosting trees predicting affinity changes; pre-training reconstructs structures disturbed by random side-chain rotations.The reconstruction task is designed to make the GNN capture patterns in atom interactions and produce predictive graphical features.Source: The GraphPPI framework · pp. 3–5, Figure 1 · p. 4
GraphPPI captures meaningful patterns of the complex structure
- Pre-training separated interface from non-interface residue descriptors and ordered disturbed-atom descriptors by disturbance level, unlike the untrained GNN.The descriptor distributions are visualized with t-SNE; the tested disturbance range was 0Å to 4Å.Source: Figure 2 · p. 6, GraphPPI captures meaningful patterns… · p. 7, Figure 2A
GraphPPI advances the state of the art in estimating effects of mutations on binding affinity
- GraphPPI achieved new state-of-the-art performance across five mutation benchmarks, including 73% improvement over Discovery Studio in Rp on S1101 and Rp 0.89 with RMSE 1.51 kcal mol−1 on M1707.Source: GraphPPI advances the state of the art… · pp. 7, 9, Figure 3 · p. 8
GraphPPI shows better prediction generalizability across proteins and structures
- Leave-one-complex and leave-one-structure tests showed better generalizability than baselines; removing pre-training reduced performance, especially across proteins.In leave-one-complex-out tests, GraphPPI improved Rp by up to 13% on S645 and M1707.Source: GraphPPI shows better prediction… · pp. 10, 12, Figure 4 · p. 11
GraphPPI accurately predicts effects of mutations on binding affinities between SARS-CoV-2 and its antibodies
- For approximated complexes of five neutralizing mAbs, GraphPPI's predicted affinity changes correlated 0.71 with measured values and had RMSE 0.67 kcal mol−1.The models were trained on the M1707 dataset before this evaluation.Source: GraphPPI accurately predicts effects… · p. 12, Figure 5 · p. 13
GraphPPI provides biological insights in designing antibody against SARS-CoV-2
- A BD-508 scan identified H47W and H59Y among three residues whose mutations could significantly improve predicted binding affinity.The scan considered 646 single-point mutations at the antibody interface.Source: GraphPPI provides biological insights… · p. 14, Figure 6 · p. 15
Methods
GraphPPI estimates mutation-induced binding free-energy changes from protein-complex graph descriptors, combining a coordinate-aware GNN with a gradient-boosting tree. Its pre-training scheme perturbs side chains and trains the GNN to reconstruct the original coordinates.
Definition of the task of predicting protein-protein binding affinity changes upon mutation
- Given a complex structure, mutated residue(s), and new residue type(s), the task is to estimate the binding free-energy change between the original and mutant complexes.The target is ΔΔG.Source: Definition of the task of predicting… · p. 16
Encoding the graph representations of a given complex
- The graph represents atoms of types C, O, N, and S as nodes, connects pairs within 3 Å, and encodes each node with 30-dimensional features.Features include element and residue type, chain, interface status, and 3D coordinates.Source: Encoding the graph representations of… · p. 16
Generation of deep graphical features by graph neural network (GNN)
- CGAT updates each atom from its neighbors using attention that incorporates coordinates, including the difference between atom coordinates in its first transformation layer.It uses multi-head attention, concatenating the outputs of K independent mechanisms.Source: Generation of deep graphical features… · pp. 16–17
- Stacked multi-attention layers and neighboring-node max-pooling produce atom descriptors with an enlarged receptive field.The GNN outputs a graphical descriptor for each atom.Source: Generation of deep graphical features… · p. 17
Prediction of binding affinity changes upon mutations by gradient-boosting tree (GBT)
- GBT predicts ΔΔG from pooled descriptors for mutated-residue and interface atoms in both the original and mutant complexes.Mean- and max-pooling summarize each selected atom set before prediction.Source: Prediction of binding affinity changes… · p. 18
Detail implementation of the pre-training scheme
- Pre-training rotates randomly selected residue side chains and teaches CGAT to reconstruct the original coordinates, using coordinate differences and mean squared error.Residues are sampled uniformly from interface residues; non-interface residues have zero selection probability.Source: Detail implementation of the… · pp. 18–20