UGC/FDS16/E16/23
← Research from the project

Binding-affinity prediction

Which atomic contacts should a model learn from to predict binding strength?

AGIMA-Score investigates this question using graphs whose edges connect protein atoms to ligand atoms.

AGIMA-Score: method and findings ↓Affinity scoring ↓

The interface between protein and ligand

Binding affinity describes the strength of association between a ligand and its target protein. A binding structure records where the ligand sits within the protein. Our AGIMA-Score study asks which connections in that structure a model should learn from: every nearby atomic contact, or contacts across the binding interface?

Atomic connections in AGIMA-Score. What counts as an edge changes what a graph can learn. Panel D retains contacts between the binding partners; panels A–C show alternative adjacency definitions. Purple identifies protein atoms and ochre identifies ligand atoms. Figure 2, Wang and Huang (2025) · CC BY 4.0.

Atomic connections in AGIMA-Score

Four adjacency diagrams: A covalent bonds; B distance contacts; C a mixture; D only intermolecular distance contacts between ligand and protein atoms.

What counts as an edge changes what a graph can learn. Panel D retains contacts between the binding partners; panels A–C show alternative adjacency definitions. Purple identifies protein atoms and ochre identifies ligand atoms. Wang and Huang (2025).

Making the graph reflect the question

In panel D, edges connect protein atoms to ligand atoms. Distance ranges distinguish different neighbourhoods. The graph therefore emphasises how the two binding partners meet.

AGIMA-Score: learning intermolecular adjacency

Scoring protein-ligand binding structures through learning atomic graphs with inter-molecular adjacency ↗
Debby D. Wang and Yuting Huang
PLOS Computational Biology 21(5), e1013074 (2025)

AGIMA-Score architecture. Atomic features F and the intermolecular adjacency tensor A enter two graph-learning branches. The aggregated features are joined to estimate binding strength. Figure 3, Wang and Huang (2025) · CC BY 4.0.

AGIMA-Score architecture

Published AGIMA-Score architecture: feature input F and adjacency input A, graph convolution and aggregation for two adjacency matrices, dense layers, concatenation, and a scalar output.

Atomic features F and the intermolecular adjacency tensor A enter two graph-learning branches. The aggregated features are joined to estimate binding strength. Wang and Huang (2025).

From atomic contacts to a score

AGIMA-Score combines atom features with distance-dependent, intermolecular adjacency matrices. Graph convolutions and aggregation produce features used to estimate binding strength.

The study examines alternative atom-feature sets and adjacency definitions. The default model uses two intermolecular contact ranges: 2–3 Å and 3–4 Å.

The research code and trained model and data accompany the paper.

Testing on independent complex sets

Models were trained on PDBbind v2020 and tested on two CSAR sets. AGIMA-Score18 achieved the lowest test RMSEs in the comparison: 1.5806 and 1.6279. The eight-feature variant achieved the highest Test1 correlation, 0.7414.

AGIMA-Score scoring comparison. Models use PDBbind v2020 for training and validation and two CSAR sets for testing. PC is Pearson correlation; lower RMSE means smaller prediction error. Table 2, Wang and Huang (2025) · CC BY 4.0.

AGIMA-Score scoring comparison

Table 2, scoring results. AGIMA-Score18 has Test1 PC 0.7339 and RMSE 1.5806, Test2 PC 0.6698 and RMSE 1.6279. AGIMA-Score8 has the highest Test1 PC, 0.7414. Full training, validation and comparison-model results are shown.

Models use PDBbind v2020 for training and validation and two CSAR sets for testing. PC is Pearson correlation; lower RMSE means smaller prediction error. Wang and Huang (2025).

Investigating what the model uses

Feature masking measures how validation performance changes when one feature type is removed. In the eight-feature variant, excluded volume has the largest effect.

Feature masking in AGIMA-Score. Masking a feature type across the validation set reveals its effect on correlation and prediction error. For the eight-feature model, masking excluded volume has the largest effect. Figure 9, Wang and Huang (2025) · CC BY 4.0.

Feature masking in AGIMA-Score

AGIMA-Score eight-feature importance plots. Excluded volume gives the largest correlation drop and RMSE increase; hydrophobic feature masking gives a small effect.

Masking a feature type across the validation set reveals its effect on correlation and prediction error. For the eight-feature model, masking excluded volume has the largest effect. Wang and Huang (2025).

UniMolRep addresses the complementary question of how representation combinations affect affinity prediction. See its PDBbind comparison for results using fingerprints, graphs and geometric features together.

Exploring grid-based scoring with GNINA

Our interactive exercise uses GNINA, a molecular docking package developed by McNutt and colleagues. It lets you examine how a three-dimensional grid of a protein–ligand complex becomes an affinity estimate.

Protein and ligand atoms occupy separate spatial channels. The crossdock_default2018 network produces an affinity estimate and a pose score. The two prepared examples are HIV-1 protease with MK1 (1HSG) and XK2 (1HVR).

a Prepared 1HSG complex

Prepared HIV-1 protease with the bound MK1 ligand, rendered from the supplied coordinates.

Protein and ligand retain their prepared coordinates.

b The MK1 binding pocket

Enlargement of MK1 and protein atoms within five angstroms, with carbon grey, oxygen red and nitrogen blue.

The enlarged view shows protein atoms within 5 Å of MK1.

c Protein and ligand grids

Protein
Protein aliphatic carbon hydrophobe grid channel at slice 23.
Ligand
Ligand aliphatic carbon hydrophobe channel at the same coordinates and scale.

Two of 28 channels, on the same plane in a 24 Å field.

28 × 48 × 48 × 48 input 3D convolutions and pooling learned readouts

Affinity estimate
pK units; larger values indicate stronger predicted binding.
Pose score
CNNscore, from 0 to 1, assesses the supplied ligand arrangement.
A binding pocket becomes a model input. Protein and ligand densities occupy separate grid channels. The network uses these channels to estimate affinity and assess the supplied pose. Complex: RCSB PDB 1HSG.

The exercise scores the ligand in its supplied pose. Docking searches for candidate poses; rescoring applies an additional scoring model to those candidates.

What changes when a region is masked?

Spatial occlusion examines how a prediction responds when part of the input grid is hidden. All channels within one region are set to zero, while the molecular coordinates stay fixed.

Evaluate the masked grid with the same model, then subtract its score from the original score. Positive differences mean that masking lowered the estimate; negative differences mean that it raised the estimate. The comparison measures the model’s sensitivity to each region. [6]

a Keep the complex fixed

The fixed MK1 ligand and grey protein pocket atoms within the grid bounds. The ochre twelve-angstrom cube marks region A in the lower x, y and z halves.

The ochre cube marks region A: the lower half of each grid axis. Ligand atoms retain their element colours; nearby protein atoms are shown as grey sticks.

b Original grid

Sum of all 28 original 1HSG channels at z slice 11, with unchanged density scale.

Evaluate the intact grid with GNINA.

c Masked grid

The same density cross-section with the lower x and y quadrant zeroed and hatched. All other values and the colour scale are unchanged.

Zero all 28 channels in region A. Evaluate again with the same GNINA model.

Original8.625 pK
Masked8.026 pK
Original − masked+0.599 pK
Masking region A lowers the affinity estimate for 1HSG. Both evaluations use the same complex and scoring model; only the grid values in the selected region change. Spatial occlusion: Zeiler and Fergus (2014).

Affinity scoring

Select a prepared complex and calculate its affinity and pose scores. Then choose Evaluate spatial occlusion to explore how masking different regions changes the affinity estimate.

Preparing analysis…

References

  1. Debby D. Wang and Yuting Huang. Scoring protein-ligand binding structures through learning atomic graphs with inter-molecular adjacency. PLOS Computational Biology 21(5), e1013074 (2025).

  2. Weiqing Guo, Jiawei Chen and Debby D. Wang. UniMolRep: A Python Package for AI-Oriented Molecular Representation Modeling. Journal of Chemical Information and Modeling (2026).

  3. Andrew T. McNutt, Paul Francoeur, Rishal Aggarwal, Tomohide Masuda, Rocco Meli, Matthew Ragoza, Jocelyn Sunseri and David Ryan Koes. GNINA 1.0: molecular docking with deep learning. Journal of Cheminformatics 13, 43 (2021).

  4. RCSB Protein Data Bank. RCSB Protein Data Bank: 1HSG. Protein structure record (1996).

  5. RCSB Protein Data Bank. RCSB Protein Data Bank: 1HVR. Protein structure record (1995).

  6. Matthew D. Zeiler and Rob Fergus. Visualizing and Understanding Convolutional Networks. Computer Vision – ECCV 2014, Lecture Notes in Computer Science 8689, pp. 818–833 (2014).

← Research from the project