CORAL: Learning Amyloid Fibril Ligand Docking with Cooperative Binding Rewards

Paper walkthrough

5:27 · English · Captioned

A short animated introduction to CORAL. Download video · Read transcript

CORAL's two-stage training: geometric learning organizes ligands into periodic stacks along an amyloid groove; energetic optimization refines the interactions between the fibril and neighboring ligand copies. — **Learning geometry, then cooperative interactions.** Geometric warm-up teaches the model to place ligand copies along the cross-β groove. Online reinforcement learning then refines the pose distribution using both protein–ligand and inter-ligand interaction scores.

Abstract

A hallmark of neurodegenerative diseases such as Alzheimer's and Parkinson's is the aberrant aggregation of proteins into amyloid fibrils, and small molecules that selectively bind to these fibrils hold promise as diagnostics, imaging probes, and therapeutics. Predicting how such ligands bind to fibril targets, however, presents two fundamental challenges. First, resolved co-crystal structures of amyloid–ligand complexes are exceptionally scarce; even with recent advances in cryo-EM only a handful have been structurally characterized, making supervised training of docking models impractical for this target class. Second, amyloid fibrils present a binding mode fundamentally different from globular proteins: ligands intercalate into longitudinal cross-β grooves and stack cooperatively along the fibril axis, a geometry that existing docking models are not designed to capture.

To address these challenges, we present CORAL (COopeRative Amyloid Ligand docking), a reinforcement learning framework that trains a generative docking model to produce ligand pose distributions tailored to the cross-β groove geometry. Our reward explicitly incorporates cooperative ligand–ligand stacking energy alongside protein–ligand docking affinity, directly capturing the distinctive binding geometry of amyloid fibrils. We further introduce a curated evaluation set of amyloid–ligand complexes constructed from model-generated poses validated by domain experts. Experiments on both experimentally resolved structures and this evaluation set demonstrate improved pose quality and binding affinity correlation over existing docking baselines.

Method

Key observation. Amyloid grooves repeat along the fibril axis. Modeling the interactions between neighboring ligand copies gives the model a signal for this cooperative binding geometry.

CORAL architecture and training pipeline: a Pairformer–Diffusion model uses fibril-periodicity priors and is trained with geometric supervision followed by cooperative-reward reinforcement learning. — **Overview of CORAL.** A Pairformer–Diffusion structure model jointly predicts coordinates for peptide chains and ligand copies. Fibril-periodicity priors encode the repeating architecture, while two training stages shape the generated pose distribution.

1. Geometric warm-up

Approximately 2,000 curated stacking poses provide supervision for the model's diffusion, distogram, and confidence objectives, establishing a prior for groove-aligned ligand geometry.

2. Cooperative-reward refinement

Online DiffusionNFT training uses computed protein–ligand and inter-ligand energy terms. A frozen-reference velocity regularizer helps retain the warmed-up geometry during refinement.

Results

We evaluate docking on AmyDock-CryoEM (18 experimentally resolved complexes) and AmyDock-Synthetic (57 evaluation pairs with computational reference poses curated by experts). CORAL reaches a Hit@5 Å of 50.0% and 68.4%, respectively, with median RMSDs of 4.89 Å and 3.84 Å.

Docking accuracy and affinity rank correlation. Mean and median report best-pose heavy-atom RMSD in Å; Hit@5 Å is the fraction of pairs with RMSD ≤ 5 Å. αSyn-Bind contains 47 compounds with measured affinities; ρ is Spearman correlation with experimental log(K_d/M). Values are reported in the paper.

Main results for eight docking methods on AmyDock-CryoEM, AmyDock-Synthetic, and αSyn-Bind.
	AmyDock-CryoEM			AmyDock-Synthetic			αSyn-Bind
Method	Mean ↓	Median ↓	Hit@5 Å ↑	Mean ↓	Median ↓	Hit@5 Å ↑	ρ ↑	p
DiffDock	17.67	13.68	0.0%	19.19	19.68	3.5%	0.191	0.197
DynamicBind	26.63	29.02	0.0%	18.90	17.93	1.8%	0.102	0.544
FABind	26.50	24.90	0.0%	30.34	26.87	0.0%	0.157	0.292
NeuralPLexer	17.43	19.00	5.6%	19.78	16.56	0.0%	−0.291	0.047
FlowDock	19.95	21.52	0.0%	20.69	19.88	0.0%	0.148	0.322
Boltz-2	28.27	28.67	0.0%	37.13	33.87	0.0%	0.223	0.132
Chai-1	19.85	19.92	0.0%	19.40	16.52	0.0%	−0.079	0.597
CORAL	8.70	4.89	50.0%	7.35	3.84	68.4%	0.419	0.004

Qualitative comparison on two AmyDock-CryoEM targets, showing ground-truth structures and predictions from seven baselines and CORAL in axial and oblique views. — **Predicted fibril–ligand poses.** Each target is shown in axial and oblique views, with close-ups of the ligand-binding region. CORAL models periodic ligand stacks along the groove. View the full-resolution figure.

Analysis

Both training stages contribute. Geometric warm-up supplies a useful stacking prior; cooperative-reward refinement improves the reported pose and energy metrics. Removing the stacking reward mainly degrades the computed interaction score, while removing the stability regularizer also worsens pose accuracy.

Ablation study on AmyDock-Synthetic. Each variant changes one component of the full model. RMSD is in Å. ΔG_eff is the paper's computed effective binding score (kcal/mol); lower values are better.

Ablation study of the reward components and two-stage training recipe.
	RMSD ↓		Computed ΔG_eff ↓
Variant	Mean	Median	Mean	Median
CORAL (full)	7.35	3.84	−2.24	−2.44
No symmetry term	7.92	4.31	−2.07	−2.21
No cooperative term	8.21	4.46	−0.78	−0.95
No stability regularizer	12.07	8.13	−1.32	−1.45
Warm-up only (no RL)	8.67	5.18	−1.41	−1.58
RL only (no warm-up)	13.45	9.27	−0.72	−0.88

Molecular Dynamics Visualizations

Simulation trajectories of candidate ligand stacks along amyloid fibrils. The gallery retains the original structure and system identifiers for each clip.

8A9L · rJCU

7WMM · System 11

More simulations (6)

7WMM · System 29

7WMM · System 10

7WMM · System 17

9UFF · c29

9UFF · c13

7UPG · System 10

BibTeX

Download .bib

@inproceedings{sun2026coral,
  title     = {{CORAL}: Learning Amyloid Fibril Ligand Docking with Cooperative Binding Rewards},
  author    = {Sun, Yasheng and Li, Bohan and Tao, Youqi and Schmidhuber, J{\"u}rgen},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026},
  url       = {https://neurips.cc/virtual/2026/loc/sydney/poster/154426}
}