CORAL: Learning Amyloid Fibril Ligand Docking with Cooperative Binding Rewards

Yasheng Sun, Bohan Li, Youqi Tao, Jürgen SchmidhuberarXiv 2026 · arXiv

TL;DR

CORAL (COopeRative Amyloid Ligand docking) is the first deep learning framework designed specifically for docking small molecules to amyloid fibrils, the protein aggregates central to Alzheimer's and Parkinson's diseases. By incorporating structural periodicity priors and an online reinforcement learning loop with a cooperative binding reward, CORAL successfully models how multiple ligands stack together inside fibril grooves, drastically outperforming general-purpose protein docking models.

What Problem Does This Paper Address?

Predicting how small molecules bind to proteins (molecular docking) is a cornerstone of computational drug discovery. While deep learning models have gotten incredibly good at docking molecules into the well-defined pockets of globular proteins, they fail completely when applied to amyloid fibrils.

Amyloid fibrils present two unique challenges:

  1. Data Scarcity: Fibrils cannot be crystallized. While cryo-EM can determine fibril structures, it rarely provides enough resolution to unambiguously identify exactly how a small molecule ligand is bound. Consequently, real-world co-crystal structures are exceptionally rare, making standard supervised learning impractical.
  2. Cooperative Stacking Geometry: Unlike globular proteins where a single ligand fits into a pocket, ligands bind to fibrils collectively. They intercalate into long, longitudinal cross-$\beta$ grooves and stack on top of each other along the fibril axis. In this environment, inter-ligand interactions ($\pi-\pi$ stacking) often dominate over direct protein-ligand contacts.

Existing docking models treat ligands independently and lack any awareness of this periodic, cooperative stacking behavior.

Methods

CORAL tackles these challenges using an AlphaFold 3-style Pairformer–Diffusion architecture augmented with structural priors and a unique two-stage training recipe.

Figure 1: Overview of CORAL. A two-stage training recipe progressively shapes the model’s ligand- pose distribution on amyloid fibrils. The geometric learning stage (left to middle) supervises the model on curated stacking poses, teaching it to align ligands along the cross-β groove into periodic stacking arrangements. The energetic optimisation stage (middle to right) further refines the pose distribution via online reinforcement learning under a composite reward that combines protein–ligand binding affinity with cooperative inter-ligand stacking energy, leading to tighter contacts such as the H-bond and π–π stacking interactions highlighted in the right inset.

1. Structural Priors

Fibrils are ribbon-like assemblies with a characteristic inter-chain spacing of $\sim4.8$ Å. CORAL introduces a ConstraintEmbedder that injects this exact translational periodicity and pocket location data directly into the model's initial pair representation, biasing its attention toward the periodic ladder of the fibril architecture.

2. Two-Stage Training Pipeline

  • Stage 1: Geometric Learning (Warm-up): To bypass the lack of real-world data, the authors generated a synthetic training set of physically plausible ligand-stacking configurations. CORAL is first supervised on this dataset to learn basic groove-intercalation geometry.
  • Stage 2: Energetic Optimization (Reinforcement Learning): The model is then fine-tuned via online reinforcement learning (using the DiffusionNFT framework) to optimize the poses based on realistic binding energetics.

3. The Cooperative Binding Reward

The reinforcement learning phase is guided by a composite reward function: $$r(\mathbf{x}) = R_{\text{bind}}(\mathbf{x}) + \lambda R_{\text{sym}}(\mathbf{x})$$

  • $R_{\text{bind}}$ (Binding Energy): Combines standard protein-ligand affinity (via AutoDock Vina) with inter-ligand stacking energy (via the MMFF94 force field). Crucially, the authors apply a thermodynamic attenuation factor ($f = 0.005$) to scale the vacuum ligand-ligand interactions so they match solution-phase free energies.
  • $R_{\text{sym}}$ (Symmetry Consistency): Penalizes ligand poses that deviate from the rigid-body mathematical transformation defining the fibril's repeating units.
Figure 2: Overview of CORAL. A Pairformer–Diffusion structure model with fibril-periodicity priors generates 3D fibril–ligand complexes from protein sequences and ligand graphs. Generated poses are scored by a reward that combines protein–ligand binding affinity with cooperative inter-ligand stacking energy, and the model is refined via online reinforcement learning.

Datasets

The authors evaluated CORAL on three distinct datasets:

  • AmyDock-CryoEM: A benchmark of 18 experimentally resolved fibril-ligand complexes extracted from the Protein Data Bank (PDB).
  • AmyDock-Synthetic: A broader evaluation set of 57 fibril-ligand pairs covering diverse chemotypes, curated via an agent-based pipeline and validated by human experts.
  • $\alpha$Syn-Bind: A set of 47 small molecules with experimentally measured dissociation constants ($K_d$) against $\alpha$-synuclein fibrils, used exclusively to test binding-affinity correlation (it contains no ground-truth structural poses).

Key Insights and Ideas

  • Cooperative Stacking is the Secret Sauce: General dockers fail because they try to optimize one ligand at a time. By explicitly rewarding the model when adjacent ligands stack neatly together, CORAL forces the generative diffusion process to discover the true collective binding mode of fibrils.
  • Bridging Gas and Solution Thermodynamics: Directly combining vacuum force fields (MMFF94) with solution-phase affinities (Vina) usually fails because vacuum interactions are orders of magnitude stronger. The inclusion of the physics-derived scaling factor ($f=0.005$) successfully balances the two forces by accounting for the entropic and desolvation costs of stacking in water.
  • RL as a Solution to Data Scarcity: When experimental data doesn't exist, physics-based reward functions can act as an effective surrogate supervisor, guiding a deep learning model to physically realistic outputs.

Key Results

CORAL significantly outpaced seven state-of-the-art docking baselines (including DiffDock, Chai-1, and Boltz-2):

  • Accuracy on Real Structures: On AmyDock-CryoEM, CORAL achieved a 50.0% hit rate (fraction of predictions with a heavy-atom RMSD $\le 5$ Å). In comparison, almost all baselines scored a 0% hit rate because they lacked fibril-periodicity awareness and tended to dock ligands vertically or completely miss the grooves.
  • Accuracy on Synthetic Benchmarks: On AmyDock-Synthetic, CORAL achieved a 68.4% hit rate with a median RMSD of 3.84 Å.
  • Affinity Correlation: On the $\alpha$Syn-Bind dataset, CORAL was the only method whose predicted binding scores significantly correlated with experimental $K_d$ values (Spearman $\rho = 0.419, p = 0.004$). This proves that CORAL captures underlying binding energetics rather than just matching shapes.

Limitations and Future Work

  • Fibril Scope: The evaluation was restricted to fibrils associated with tau, $\alpha$-synuclein, and amyloid-$\beta$. Future work needs to validate the model on other pathogenic aggregates like transthyretin or light-chain amyloids.
  • Reward Exploitation (Cheating): In a few failure cases, the model "cheated" the reward function by folding a ligand back on itself or rotating it sideways to maximize both protein contact and stacking energy simultaneously, resulting in distorted geometries.
  • Closed-Loop Discovery: The authors intend to integrate CORAL into a fully automated, AI-orchestrated closed-loop pipeline that combines de novo ligand generation, docking, and filtering to discover new drug candidates with minimal human intervention.

Summary updated

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.
BibTeX
@misc{sun2026coral,
  author = {Yasheng Sun and Bohan Li and Youqi Tao and Jürgen Schmidhuber},
  title = {{CORAL: Learning Amyloid Fibril Ligand Docking with Cooperative Binding Rewards}},
  year = {2026},
  eprint = {2607.17412},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2607.17412},
}
Figure