Deep Docking

Deep Docking employs an iterative active learning paradigm to progressively refine its predictive model. In each iteration, a small, random subset of the chemical library is docked using a traditional physics-based program. These docking scores are used to train a deep neural network (DNN) , which then predicts the scores for the entire remaining library. The top-scoring virtual hits are fed back into the next iteration for experimental docking, creating a feedback loop that focuses computational resources on the most promising candidates while continuously improving model accuracy.

At its core, Deep Docking functions as a sophisticated QSAR model that learns the complex, non-linear relationship between a molecule's structure and its predicted docking score. The DNN ingests molecular fingerprints—such as extended connectivity fingerprints (ECFPs) or topological torsion fingerprints—as input features. The model is trained to regress a continuous docking score, effectively learning to approximate the behavior of the physics-based scoring function without performing the computationally expensive pose search and energy calculation for every molecule.

The primary utility of Deep Docking is the ability to triage ultra-large virtual libraries containing billions of compounds, such as the Enamine REAL Space or ZINC20. A brute-force docking campaign on a billion compounds is computationally prohibitive. Deep Docking reduces this to docking only a small fraction (typically 0.1-1%) of the library. The trained DNN then virtually screens the entire remaining space in hours, identifying a highly enriched subset of top-ranked molecules for final, confirmatory docking, making the exploration of vast chemical space feasible.

A critical performance metric for Deep Docking is its ability to recall the true top-scoring molecules that would have been identified by a full brute-force docking campaign. The methodology is designed to minimize false negatives. By retraining on progressively higher-scoring molecules in each iteration, the model becomes highly proficient at identifying the chemical features associated with favorable binding. Studies have demonstrated that Deep Docking can successfully recall over 90% of the top 50,000 virtual hits while docking less than 1% of the library.

Deep Docking is an agnostic wrapper around the docking engine and is not tied to any specific software. It can be used with popular docking programs such as AutoDock Vina, Glide (Schrödinger) , or FRED (OpenEye) . The DNN learns to predict the output of whatever scoring function is used in the backend. This flexibility allows research teams to integrate Deep Docking into their existing validated virtual screening workflows without changing the underlying physics-based methodology they trust for final pose evaluation.

Beyond simple scoring, the trained neural network's latent space can be analyzed to understand the chemical features driving predicted activity. By examining which molecular fingerprints most strongly activate the model's neurons for high-scoring predictions, medicinal chemists can gain insights into the structure-activity relationship (SAR) . This can guide the selection of compounds from different chemical series and facilitate scaffold hopping, where the model identifies structurally novel chemotypes that are predicted to have high affinity but would be missed by simple similarity searches.

The foundational methodology that Deep Docking accelerates. SBVS uses the three-dimensional structure of a biological target to computationally dock and score candidate ligands, predicting their binding mode and affinity.

• Requires a high-resolution protein crystal structure or cryo-EM map
• Evaluates ligand pose complementarity within the binding pocket
• Deep Docking replaces exhaustive SBVS with a surrogate neural network for scoring

The application of virtual screening to chemical libraries containing billions of compounds, made feasible by AI-accelerated docking and cloud computing. Deep Docking was specifically designed to tackle this scale.

• Libraries like Enamine REAL contain over 20 billion synthesizable molecules
• Exhaustive docking of billion-scale libraries would require centuries of CPU time
• Deep Docking reduces the compute burden by 100-fold or more while retaining top-scoring candidates

A computational method that predicts the preferred orientation of a small molecule when bound to a protein target to form a stable complex. Docking generates the training data that Deep Docking models learn from.

• Produces a binding pose and a numerical docking score
• Popular docking programs include AutoDock Vina, Glide, and FRED
• Deep Docking treats docking scores as ground truth labels for training the surrogate model

An iterative machine learning paradigm where a model strategically selects the most informative unlabeled compounds for evaluation. Deep Docking employs active learning to choose which molecules to dock for training.

• Balances exploration (sampling uncertain regions) and exploitation (sampling high-scoring regions)
• Each iteration refines the model on newly docked molecules
• Dramatically reduces the number of required docking calculations

A technique for encoding the structural features of a molecule into a fixed-length binary vector or real-valued descriptor. These fingerprints serve as the input features for the Deep Docking neural network.

• Extended Connectivity Fingerprints (ECFP) capture circular atom neighborhoods
• MACCS keys encode predefined structural fragments
• The quality of fingerprint representation directly impacts the model's ability to generalize across chemical space

A mathematical function that estimates the binding affinity between a protein and ligand by approximating the free energy of the complex. Deep Docking learns to predict these scores without performing explicit physics-based calculations.

• Force-field-based functions calculate van der Waals and electrostatic interactions
• Empirical functions fit weights to known binding data
• Knowledge-based functions derive potentials from statistical analysis of protein-ligand complexes
• Deep Docking's neural network acts as a learned scoring function surrogate