Conventional high throughput screening for clinical candidates cannot evaluate over 99% of commercially available molecules, thereby restricting the accessible chemical search space for therapeutic applications. With the rise of artificial intelligence (AI) and machine learning (ML) tools to provide improved speed and failure rate reductions for drug discovery, a new study from the Atomwise AIMS (Artificial Intelligence Molecular Screen) initiative presents computational screening as a viable alternative to physical high-throughput screening (HTS) for the first step of small molecule drug discovery.
Atomwise is a technology-enabled pharmaceutical company leveraging deep learning for structure-based drug design. Their proprietary AI/ML drug discovery platform, AtomNet, applies a virtual HTS approach that searches a chemical library of more than 15 quadrillion synthesizable compounds to find hits in new chemical space.
In the paper published in Scientific Reports titled, “AI is a viable alternative to high throughput screening: a 318-target study,” AtomNet was applied to 318 targets identified from collaborations from over 250 academic labs across 30 countries. The platform successfully identified structurally novel hits for 235 of 318 targets evaluated, and provided a success rate of 74%, an improvement upon literature estimated HTS success rates of 50%.
“It’s important to be able to go into novel parts of chemical space because if you want to help patients, you have to be meaningfully differentiated in the clinic,” said Abraham Heifets, CEO of Atomwise, highlighting that such differentiation comes from providing first-in-class strategies when no treatment exists or improving upon existing strategies.
Heifets emphasized that AtomNet’s strength comes from the platform’s versatility across “hundreds of targets and in different people’s hands.” Hits were robust across a wide breadth of protein classes and major therapeutic areas including, oncology, infectious disease, neurology, immunology, cardiovascular disease, and more. Enzymes represented 59% of the target protein classes. Additional protein classes included GPCRs, transporters, ion channels, and DNA/RNA-binding proteins.
Example therapeutic directions from the AIMS initiative included the first reducer for Miro1, a new target in Parkinson’s disease, the first inhibitors for OTUD7A and OTUD7B, challenging deubiquitinase targets for solid and hematological tumors, and small molecule inhibitors for CTLA-4, a well-established oncology target.
“Generally the predictive power of virtual screening platforms has been extremely limited,” said Gregory Bowman, professor at the University of Pennsylvania, who used AtomNet to uncover a cryptic site for phosphatase PPM1D, a therapeutic target in oncology. “Results of the AIMS study show that AtomNet has a high success rate in finding hits for traditionally challenging biology, like allosteric or protein-protein interactions.”
Heifets attributes AtomNet’s success to a paradigm shift from a per target model to a global model.
“In a lot of AI/ML approaches, if you’re working on protein A, you build a ML model for protein A. If you move on to protein B, you build a new ML model for protein B. In that world, it matters whether you have training data for A versus B versus C. If you have no training data, it’s not clear that you can build a model,” said Heifets.
Alternatively, AtomNet is pre-trained on a range of molecular data across the proteome, allowing the platform to be more generalizable across targets.
Looking ahead, Atomwise has been using AtomNet to enter the inflammatory disease market. The company plans to file an IND application this year for its lead candidate, a novel allosteric TYK2 inhibitor discovered using AtomNet.
