Deep Origin Unveils Breakthrough in Virtual Screening with DODock

Deep Origin has recently announced a groundbreaking advance in virtual screening, as detailed in a preprint published on bioRxiv. The company has developed a novel framework called DODock, which, when combined with DOScore, maintains predictive accuracy on new biological targets, outperforming conventional machine-learning models. "Virtual screening has been widely used for more than 40 years, yet it has rarely delivered an actual drug,” remarked Garegin Papoian, Ph.D., co-founder and chief scientific officer of Deep Origin.

The challenge with existing models is their accuracy, which significantly drops when faced with new biology and chemistry. Leading co-folding models often see accuracy plummet from mid-80s to below 25%. This inadequacy is evident in the independent Runs N’ Poses benchmark, where AI co-folding models achieve 75% to 88% pose accuracy on familiar targets but drop below 25% for unfamiliar ones. In contrast, DODock achieves 89% accuracy for familiar complexes and maintains over 50% pose accuracy on novel complexes.

To combat data leakage that inflates model performance, Deep Origin employs a strict double-similarity filter, eliminating proteins with over 30% sequence identity and ligands exceeding 0.4 Tanimoto similarity to test sets. Under these conditions, DODock reaches an impressive 80% pose accuracy on OpenBind, while co-folding models only achieve between 4% and 28%.

Deep Origin’s methodology combines AI with physics, addressing the limitations of traditional models. DODock uses a diffusion model to generate 3D pose proposals, which are then refined by a physical energy engine (DOFast) that calculates fundamental molecular forces. Finally, an AI model ranks these poses based on atomic contact points. Michael Antonov, co-founder and CEO of Deep Origin, stated, "What matters is how a screen behaves on a target no one has solved yet, so that is where we tested our model with strict splits and cases built to make it fail."

The research also highlighted the significant improvements in hit rates from prospective virtual screenings across four therapeutic targets. For instance, the screening for CD73 yielded a hit rate of approximately 30.6%, a remarkable improvement compared to a previous AI screen that reported only 0.3%.

Deep Origin has made its methodology publicly available, encouraging the scientific community to replicate and test their results. "We published the whole method - every algorithm, the data strategies, the full supplement - and left it open for others to test it on molecules we have not seen,” said Papoian. This transparency is seen as a crucial step in advancing the field of drug discovery.

The research was co-led by Head of AI Garik Petrosyan, Ph.D., and Dr. Papoian, with experimental validation conducted in Deep Origin's laboratory and partner organizations. The company aims to bridge the gap between preclinical predictions and clinical outcomes by leveraging advanced computational discovery systems.

About DODock and DOScore: DODock predicts molecular binding to proteins by integrating machine learning with physics, while DOScore assesses binding affinity across extensive chemical spaces. For more information on DODock, visit https://deeporigin.com/docking/.