Overall, DiffDock should be used with caution. You can't really rely on every prediction it makes. For example, if the predicted poses look like those in the paracetamol case (spread around the surface), then the model is not confident with its results and you shouldn't trust them. However, the general trend shows that when all the poses are close to one another and placed at the same region on the target molecule surface (like in the ibuprofen example), you can trust the results more. Thanks to the authors of the model, we have a rough evaluation of DiffDock confidence scores in practical terms: you can treat the scores as good (as good prediction) when they are above 0, and from moderate to low when the scores are bad (the more negative, the less confident and the worse the prediction).

Architecture insights

The main idea behind the DiffDock is “Any ligand pose consistent with a seed conformation can be reached by a combination of (1) ligand translations, (2) ligand rotations, and (3) changes to torsion angles”. Hence, the authors construct the score model and the confidence model to take as input the current ligand pose and the protein structure in 3D space. Both models use SE(3)-equivariant convolutional networks over point clouds, but the score model operates on a coarse-grained representation of protein structures (C-alpha), while the confidence model utilizes an all-atom structure. The representation involves heterogeneous geometric graphs, language model embeddings, and convolution operations for translational, rotational, and torsional scores. The multiscale setup improves performance and speeds up the process compared to atomic-scale approaches. The architectural components include initial features, distance-based connections, and specific convolutional operations tailored to each model's requirements.

This architecture achieves a higher success rate in docking into both experimental and predicted target structures. The evaluation was made in terms of the percentage of predictions with a heavy-atom RMSD <2Å and the median heavy-atom RMSD between the predicted and the experimental ligand atoms. The experimental structures were obtained from 363 PDBBind (Wang et al. 2005) complexes from 2019 that were used for testing of the model. DiffDock outperforms state-of-the-art tools such as SMINA (Koes, Baumgartner, and Camacho 2013), QuickVina-Q (Alhossary et al. 2015), GLIDE (Yang et al. 2021), GNINA (McNutt et al. 2021), Autodock Vina (Eberhardt et al. 2021), EquiBind (Stärk et al. 2022), and TANKBind (Lu et al. 2022), demonstrating remarkably faster speeds (3 to 12 times faster than the best search-based method), but with only moderate improvement in accuracy. This speed is particularly valuable for applications like virtual screening and reverse screening for drug candidates or protein targets. 

The availability of Google Collab and 310 copilot notebooks makes running DiffDock convenient. As the input you provide a ligand in SMILES (Weininger 1988) notation, and a structure of your protein target in PDB format. Another strength lies in its compatibility with predicted structures, yielding similar accuracy results. DiffDock’s “ability to generalize to imperfect structures, even without retraining, can be attributed to a combination of (1) the robustness of the diffusion model to small perturbations in the backbone atoms, and (2) the fact that DiffDock does not use the exact position of side chains in the score model and is therefore forced to implicitly model their flexibility”. It highlights the importance of dynamics in understanding and predicting protein-ligand complexes, hence methods like molecular dynamics (MD) simulations or the new AI tools (Jing, Berger, and Jaakkola 2024) that are trying to replace it or speed it up are of paramount importance for the next (after AlpfaFold2 (Jumper et al. 2021)) leap in structural biology.