DiffDock-L

What is DiffDock-L?

DiffDock-L is the latest version of the state-of-the-art molecular docking tool developed at MIT that uses diffusion models to predict how small molecule ligands bind to protein targets. Building upon the original DiffDock published at ICLR 2023, DiffDock-L represents a paradigm shift in computational docking by treating the problem as a generative modeling task rather than traditional optimization.

DiffDock-L vs. DiffDock

Released in February 2024 (ICLR 2024), 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%, representing a substantial improvement in prediction accuracy. On the DockGen benchmark specifically designed to test generalization across protein domains, DiffDock-L improved success rate 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. This addresses a key limitation where the original DiffDock struggled with unseen protein targets.
• Larger model & more training data: DiffDock-L scales from 4M to 30M parameters and benefits from increased training data, larger model sizes, 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, focusing on quality over quantity and reducing computation time by approximately 2x while maintaining superior accuracy.

How does DiffDock-L work?

Unlike traditional docking methods that use physics-based scoring functions and optimization algorithms, DiffDock-L applies diffusion models—the same technology behind AI image generators like DALL-E and Stable Diffusion. The key innovation is treating molecular docking as a generative modeling task over a non-Euclidean manifold rather than an optimization problem.

Diffusion process on SE(3) manifold

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.

Graph representation

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.

Embedding message-passing layers

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.

SE(3) equivariant neural networks

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.

Confidence model

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.

Input requirements

DiffDock-L requires two inputs:

• Protein structure: A PDB file or PDB ID from RCSB PDB database. The protein should be prepared (protonated, cleaned) for best results (you can use tools like PDB Fixer to prepare it).
• Ligand: SMILES string (most common for small molecules), SDF or MOL2 file format are also supported.

Understanding the results

DiffDock-L generates multiple binding poses (default: 10) ranked by confidence.

Rank

Poses are sorted by confidence score, with Rank 1 being the most confident prediction.

Confidence score

A model-predicted score indicating confidence in the binding pose. Higher (less negative) scores indicate higher confidence. Typical ranges:

• Score > -1.5: High confidence
• Score -1.5 to -2.5: Moderate confidence
• Score < -2.5: Lower confidence

The confidence model estimates the probability that a pose has RMSD <2Å to the true binding mode.

Multiple poses

Examining multiple top-ranked poses is recommended:

• Different poses may represent alternative binding modes
• The highest-ranked pose isn't always correct
• Ensemble analysis provides more robust predictions

Best practices

Number of poses

• Start with 10-20 poses for general use
• Increase to 30-40 if exploring diverse binding modes
• More poses increase computation time proportionally

Protein preparation

• Remove water molecules unless they're critical for binding
• Add missing hydrogens and residues
• Ensure proper protonation states at physiological pH
• Use tools like PDB Fixer for automated preparation

Result validation

• Examine top 3-5 ranked poses, not just the top one
• Check for reasonable protein-ligand interactions (H-bonds, hydrophobic contacts)
• Validate critical predictions with molecular dynamics or experimental methods

Limitations

• Does not account for protein flexibility (rigid receptor assumption)
• Performance may vary for large or macrocyclic ligands
• Metalloprotein binding requires special consideration
• Trained primarily on x-ray structures; performance on computationally folded structures is reduced but still superior to traditional methods

Comparison to traditional docking

Advantages of DiffDock-L

• Superior accuracy: 43% success rate (RMSD <2Å) versus 23% for traditional methods on PDBBind
• Better handling of flexible ligands: Generative approach naturally models conformational flexibility
• No manual parameter tuning: Works out-of-the-box without exhaustive search parameters
• Diverse poses: Diffusion sampling generates multiple binding modes automatically
• Confidence estimates: Built-in confidence model for pose ranking

Traditional methods (AutoDock Vina, GOLD)

• Faster computation: Optimization-based approaches can be quicker for simple systems
• More interpretable: Physics-based scoring functions provide mechanistic insights
• Established validation: Decades of validation in drug discovery pipelines
• Specialized cases: Better support for metals, covalent binding, and specific constraints

For most drug discovery applications, DiffDock-L provides the best balance of accuracy and ease of use.

Cost

Using DiffDock-L through ProteinIQ costs 100 credits per docking job, regardless of the number of poses generated.

---

References

• Corso, G., Stärk, H., Jing, B., Barzilay, R., & Jaakkola, T. (2023). DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking. International Conference on Learning Representations (ICLR). https://openreview.net/pdf?id=kKF8_K-mBbS
• Corso, G., et al. (2024). Deep Confident Steps to New Pockets: Strategies for Docking Generalization (DiffDock-L). International Conference on Learning Representations (ICLR). https://arxiv.org/abs/2402.18396