DiffDock Protein-Ligand Docking: When to Use It?

In the long chain of drug discovery, molecular docking has always been one of the most central and time-consuming steps. Given a protein target and a small-molecule ligand, how can we predict their three-dimensional binding conformation? Traditional methods rely on search algorithms and scoring functions, repeatedly sampling, scoring, and optimizing within a vast conformational space in an attempt to find the "optimal solution." This process typically takes hours to days and is heavily dependent on predefined binding pocket locations.

At the end of 2022, an international research team offered a fundamentally different answer—treating docking as a generative problem rather than a search problem. That answer was named DiffDock.

I. A Paradigm Shift: From "Search" to "Generation"

Conventional molecular docking methods essentially perform a search task: within a predefined binding pocket, they continuously adjust the ligand's position, orientation, and conformation using Monte Carlo, genetic algorithms, and other techniques, evaluating each candidate pose with a scoring function and ultimately selecting the highestscoring conformation. This is like looking for a needle in a vast ocean—the target direction is clear, but the conformational search space is enormous, the scoringfunction energy landscape is rugged, and the algorithm easily gets trapped in local optima.

DiffDock adopts a completely new approach: no longer searching, but generating. It frames molecular docking as a generative task, learning a mapping from random noise to valid binding conformations. Given a protein and a ligand structure, DiffDock does not "seek" the correct binding pose; instead, it directly generates a set of plausible binding conformations and then selects the most reliable ones.

Behind this conceptual shift lies a key insight: traditional regression methods attempt to predict a single "correct answer," but molecular docking essentially requires learning a probability distribution—given the protein and ligand, all feasible binding conformations collectively form a distribution, and generative models are naturally suited for such modeling.

II. How Does DiffDock Work?

How Diffusion Works

At its core, DiffDock is a diffusion generative model operating on the non-Euclidean manifold of ligand poses.

Ligand–protein binding involves three types of degrees of freedom: translational position relative to the protein, rotational orientation within the pocket, and torsional conformation of the ligand's rotatable bonds. DiffDock maps these three degrees of freedom into a unified mathematical space and builds a diffusion process on that space.

Forward diffusion process: Starting from a true binding conformation, noise is gradually added, progressively scrambling the ligand's position, orientation, and torsion angles until the conformation degenerates into a completely random noise distribution. Reverse diffusion process: The model learns to denoise through training, starting from a random initial pose and progressively removing noise to reconstruct a physically plausible binding conformation. During inference, the model repeatedly executes the reverse diffusion process to output multiple candidate binding poses.

Critically, DiffDock is a blind docking model—it does not require predefined binding pocket locations. It can sample ligand poses across the entire protein surface, autonomously identifying potential binding sites. This feature is highly valuable for studying novel targets and discovering allosteric sites.

Denoising to Binding Poses

III. Performance: 38% Top1 Success Rate

On the standard PDBBind test set, DiffDock achieves a 38% top-1 success rate (predicted conformation with RMSD < 2 Å from the experimental structure), significantly outperforming traditional docking methods (23%) and early deep-learning approaches (20%).

In more challenging scenarios that better reflect real-world drug discovery—docking against computationally predicted protein structures (such as ESMFold outputs) rather than experimentally determined ones—DiffDock's advantage becomes even more pronounced. Previous traditional docking and deep-learning methods achieved at most 10.4% accuracy (from TANKBind) in this setting, while DiffDock reaches 21.7%. This blind-docking setup with apo-like protein structures more closely mimics the actual conditions of early-stage drug discovery.

DiffDock's inference pipeline consists of two steps: first, generating a set of candidate binding poses via the diffusion model; second, ranking and filtering these poses with a built-in confidence module to select the most reliable binding mode. The model is relatively fast, and its confidence estimates show good discriminative power.

IV. Evolution and Applications of DiffDock

Since its release in 2022, DiffDock has spawned several variants tailored to diverse research scenarios:

DiffDock-L: An enlarged version that leverages synthetic data strategies to expand the dataset and model scale, supporting large-scale virtual screening.

DiffDock-Pocket: A pocketrestricted version suitable for docking when the binding pocket is already known.

DiffDock-Site: Incorporates a bindingsite recognition module to further improve pose prediction accuracy.

DiffDock-L-Allo: Fine-tuned for allosteric kinase inhibitors, showing strong performance on Type III/IV allosteric inhibitor pose prediction.

DiffDock-NMDN: An end-to-end blind docking and virtual screening protocol that combines DiffDock's sampling capability with affinity prediction.

In real-world drug discovery, DiffDock has been applied to natural-product virtual screening, allosteric-site discovery, and other scenarios.

V. Challenges and Controversies

Any emerging technology faces scrutiny before adoption, and DiffDock is no exception.

One point of contention is the dataset split. The original training set included approximately 98% of known protein–small-molecule complex structures, leaving only 2% for testing. Critics argue that this split may lead to test samples being highly similar to training data, potentially overestimating the model's generalization ability.

A preprint from late 2024 systematically compared DiffDock against traditional docking tools (such as Glide and AutoDock Vina). The results showed that in conventional scenarios where the binding pocket is known in advance, traditional tools outperformed DiffDock overall; even in blind-docking tasks without pocket information, DiffDock lagged behind in the number of accurate predictions.

Another major challenge is generalization beyond the training set. When removing training samples similar to test cases, DiffDock's performance on difficult tasks dropped markedly. On benchmarks like DockGen, which feature entirely new binding pockets, the original DiffDock's top-1 success rate plummeted to 7.1%; even the upgraded DiffDock-L recovered to only 22.6%, still far below its performance on the standard PDBBind set.

These findings offer important lessons: generative molecular docking holds considerable promise, but it is not a universal panacea at this stage. For routine projects with well-defined binding pockets and experimentally determined protein structures, traditional docking software remains stable and efficient. DiffDock's true value lies in expanding the boundaries of molecular docking—excelling in blind docking, novel targets, and allosteric-site prediction, where conventional methods struggle.

VI. From "Prediction" to "R&D": An AI-Driven New Paradigm in Protein Research

From Prediction to R&D Platform

DiffDock's inspiration extends beyond molecular docking. It represents a role transition for AI in protein research—from outputting a single prediction to generating a diverse set of plausible conformations, from depending on human-defined constraints to autonomously exploring uncharted protein surfaces.

This "generative exploration" mindset is continuously permeating the entire protein R&D workflow. In April 2026, the conversational protein R&D agent MatwingsVenus™ (Xiaowu™) was officially launched, extending AI capabilities from structure prediction to sequence design, functional validation, and experimental optimization in a complete closed loop. In July 2026, the platform was selected for the "Treasure of the Hall" award at the World Artificial Intelligence Conference (WAIC), the only AI for Science product to receive this honor.

On the MatwingsVenus™ (Xiaowu™) platform, researchers describe their R&D needs in natural language, and the system automatically orchestrates various specialized tools to handle complex tasks, covering sequence analysis, protein design, structure prediction, and functional optimization in an end-to-end pipeline. The platform integrates over 200 protein design tools, supports billion-scale real-labeled protein data retrieval, and is backed by more than 50 certified experts and over 30 domain-specific fine-tuned skill sets.

The platform's "Protein Generation" module has incorporated core models such as BoltzGen, LigandMPNN, and Protenix, linking structure prediction, ligand design, and sequence generation into a unified intelligent R&D pipeline. After completing molecular docking or structural analysis, users can directly initiate protein engineering and functional optimization on the same platform, without needing to migrate data across platforms or rebuild workflows.

This full-chain integration is reshaping protein research: moving from isolated predictions to a complete R&D closed loop, and from scattered independent software to a unified intelligent platform. Industry consensus indicates that the biggest bottleneck in AI-driven protein R&D is no longer the accuracy of individual models, but whether the entire pipeline from molecular design to wet-lab validation can be seamlessly connected.

VII. Summary

DiffDock is a landmark exploration in molecular docking. By redefining the classical docking problem with a generative diffusion model, it has demonstrated differentiated advantages in challenging scenarios such as blind docking, novel targets, and allosteric sites, providing a new paradigm for AI applications in structural biology. At the same time, it reminds practitioners that the ceiling of AI model performance is often constrained by the coverage of training data and the degree to which test scenarios match real-world conditions.

For those working in drug discovery and protein research, DiffDock is not meant to replace traditional computational tools but to complement them. Understanding the appropriate applicability boundaries of different methods—knowing when to use DiffDock and when to rely on conventional docking software—is more important than simply mastering how to operate the tools.