ArtiDock: AI Docking That Beats Glide And AlphaFold

Development

Development of ArtiDock: Receptor.AI developed ArtiDock, an advanced AI docking model designed to deliver rapid and accurate predictions of ligand binding poses within protein pockets.
Data augmentation: The proprietary algorithmic technique for generating artificial pockets with statistical distributions of non-bond interactions from experimental structures was employed with the ensembles of representative conformations for most known protein-ligand complexes.

Lightweight architecture: The advantage of a much larger augmented dataset allowed the creation of a very lightweight model architecture that only uses the data about the binding pockets instead of the whole protein.

Benchmarking

ArtiDock Model Superiority: A detailed evaluation places ArtiDock against top AI docking methods and traditional programs like Vina, Gold, and Schrodinger's Glide.
Outstanding Performance: ArtiDock surpasses AlphaFold-latest, the previous front-runner in predicting protein-ligand complexes, and exceeds classical docking algorithms in precision, significantly enhancing throughput.
Throughput Comparison: Notably faster, with physics-based docking algorithms and the Alpha-Fold model showing 20-600 times lower throughput compared to ArtiDock.
Comprehensive Performance: Excels across all RSMD thresholds, ensuring accurate binding poses and identification of relevant chemical structures of ligands.

Market Leadership: ArtiDock has the best precision and speed within methods providing relevant binding predictions.

Benchmark Datasets Overview

PoseBusters v1 Dataset: Specifically tailored to assess the quality of AI docking algorithms, focusing on their precision and reliability.
‍PoseBusters v3 Dataset: Aimed at further challenging AI docking models through more complex screening scenarios, testing their robustness and adaptability.

Original article

Find out more about ArtDock in the preprint.

Updates

Advancement to ArtiDock 2.5: Building upon the success of the initial version, ArtiDock 2.5 incorporates all features of its predecessor along with significant enhancements. Key upgrades include simulating binding site variability through random pocket extraction during training and accounting for ions, cofactors, and non-standard residues.

Learn more about ArtiDock 2.5 here.