DiffDock-L is the latest version of the molecular docking tool developed at MIT that uses diffusion models to predict how small molecule ligands bind to protein targets. It treats molecular docking as a generative modeling task rather than traditional optimization, using the same technology that powers image generators like DALL-E and Stable Diffusion.
Building upon the original DiffDock published at ICLR 2023, DiffDock-L (released at ICLR 2024) achieves significantly better accuracy and generalization to novel protein targets.
DiffDock-L introduced several enhancements over the original DiffDock:
Enhanced accuracy: DiffDock-L achieves 43% success rate for ligand RMSD <2Å compared to the original DiffDock's 38%. On the DockGen benchmark designed to test generalization across protein domains, success rate improved from 10% to 24%.
Better generalization: The model uses ECOD (Evolutionary Classification of protein Domains) structure-based cluster sampling during training to ensure better performance on proteins with novel binding pocket architectures.
Larger model & more training data: DiffDock-L scales from 4M to 30M parameters and benefits from increased training data and novel synthetic data generation (extracting side chains from protein structures as ligands).
Confidence bootstrapping: A novel self-training scheme where the diffusion model and confidence model work together to fine-tune on unseen protein domains without structural data.
Optimized output: By default, DiffDock-L outputs 10 high-quality predictions compared to the previous 40, reducing computation time by approximately 2x while maintaining superior accuracy.
Unlike traditional docking methods that use physics-based scoring functions and optimization algorithms, DiffDock-L applies diffusion models to molecular docking. The key innovation is treating docking as a generative modeling task over a non-Euclidean manifold rather than an optimization problem.
DiffDock-L defines a diffusion process over the degrees of freedom involved in docking: the ligand's position relative to the protein (translation), its orientation in the binding pocket (rotation), and the torsion angles describing its conformation. This maps to the SE(3) manifold—the Special Euclidean group in 3D space that represents all possible rigid body transformations.
The model starts with a random ligand pose and iteratively denoises it through learned reverse diffusion steps to converge on plausible binding poses. Unlike traditional methods that optimize a single pose, this generative approach naturally produces diverse binding modes.
Protein and ligand structures are represented as heterogeneous geometric graphs. Ligand atoms and protein residues serve as nodes, with residues receiving initial features from language model embeddings trained on protein sequences.
Nodes are sparsely connected based on distance cutoffs that depend on node types and diffusion timestep, allowing efficient message passing while capturing relevant interactions.
DiffDock-L adds embedding message-passing layers that independently process protein and ligand structures before cross-attention. Unlike DiffDock's purely cross-attentional layers, this allows increased architectural depth with minimal runtime overhead.
Under the rigid protein assumption, protein embeddings are computed once and cached across all diffusion steps and samples, providing efficiency gains.
The model uses geometric deep learning with SE(3) equivariance—rotations and translations of the input don't change the predicted binding physics, only the coordinate frame. This physical symmetry is encoded through specialized graph neural networks that process geometric information while respecting 3D symmetries, improving generalization and data efficiency.
For each generated pose, DiffDock-L predicts a confidence score estimating the likelihood that the pose has RMSD <2Å to the true binding mode. This confidence model is trained jointly with the diffusion model and enables automatic ranking of generated poses without requiring external scoring functions.
DiffDock-L requires two inputs:
Number of poses: How many binding poses to generate. We recommend starting with 10-20 poses for general use. Increase to 30-40 if exploring diverse binding modes—more poses increase computation time proportionally.
Inference steps: Number of diffusion denoising steps. The model was trained with 20 steps. Research suggests 15 is optimal for speed/accuracy balance.
Actual steps: Number of steps actually executed (usually inference_steps - 2). This is an advanced parameter for optimization.
No final step noise: Removes noise from the final denoising step for more stable and deterministic predictions. We recommend keeping this enabled.
Save reverse-diffusion visualization: Writes an additional PDB trajectory for each ranked pose showing the reverse diffusion process. This is useful for debugging and qualitative inspection, but it increases output size.
DiffDock-L generates multiple binding poses ranked by confidence.
Poses are sorted by confidence score, with Rank 1 being the most confident prediction.
A model-predicted score indicating confidence in the binding pose. Higher (less negative) scores indicate higher confidence:
The confidence model estimates the probability that a pose has RMSD <2Å to the true binding mode.
We recommend examining multiple top-ranked poses:
DiffDock-L returns each ranked ligand pose as an SDF file, which preserves small-molecule bond orders and atom typing more faithfully than ligand-only PDB exports.
If reverse-diffusion visualization is enabled, DiffDock-L also returns a PDB trajectory for each ranked pose showing the denoising path used to reach the final docking configuration.
Clean structures yield better results:
DiffDock-L has known limitations that may affect certain use cases:
For most drug discovery applications, DiffDock-L provides the best balance of accuracy and ease of use.
ProteinIQ offers several molecular docking tools, each with different strengths:
For analyzing your ligand properties before docking, see ADMET-AI for pharmacokinetic predictions or Lipinski's Rule of Five for drug-likeness assessment.
