Skip to Main Content

By Abraham Heifets, CEO of Atomwise

Undrugged genes represent 96 percent of the human genome and include the most challenging — and the most promising — future for pharma and human health. Of the roughly 20,000 proteins encoded in the human genome, only 754 have FDA-approved drugs. About 4,000 additional genes already have evidence linking them to human disease and the remaining human gene targets are uncharted territory for drug discovery. Many of these new targets identified by biological research do not have enough information to enable traditional drug design efforts. A lack of known drug scaffolds, imprecise structural information, or a high degree of protein complexity can interfere with a medicinal chemist’s ability to design new drugs against the target, rendering a promising disease target “undruggable.”

Artificial intelligence is the catalyst to design new drugs for these challenging targets. Using AI, medicinal chemists can run hundreds of millions of tests on proteins of interest before pursuing a potential candidate. Today, machine learning can help answer many of the challenges of drug discovery, including identifying novel hit molecules that are unencumbered by competing patents, and developing such hits into potent, selective, soluble, permeable, plasma-stable, and synthesizable molecular lead series which are safe in computational toxicity panels. A successful drug must simultaneously optimize these competing factors and, like a Venn diagram with too many circles, the intersection that satisfies all of the constraints can be very small indeed. Our AI platform, AtomNet®, can computationally screen over 16 billion molecules (which is 5,000 times larger than typical big pharma corporate collections) in less than two days, and discover novel molecules that land in that needed intersection. By effectively pairing AI and physical testing, medicinal chemists can ensure that the molecules on which they choose to devote their labor, time, and money have a maximal probability of success.

Not all AI is created equal, and to address the most challenging targets requires an AI platform that has been challenged with, and succeeded on, diverse experimental data. Atomwise created the first convolutional neural networks (CNN) for drug discovery, and we have spent years developing and improving our technology through efficient engineering, algorithmic insights, careful data cleaning, deep domain expertise, and uncompromising and rigorous performance analysis. We work with academic research teams on projects that test and improve our AI models across all major protein classes and hundreds of disease areas, further enhancing the capabilities of AtomNet®. Our Artificial Intelligence Molecular Screening Program (AIMS) provides academic researchers with a carefully-selected set of compounds to test in their laboratory. From this we have gained a valuable breadth of experimental data for the performance of AtomNet®, including the largest diversity of drug target sites, homology models, protein classes, and disease areas of any AI platform. We have a track record of discovering hits across all of these projects above 75%, a testament to the training which our commercial partners also benefit from.

Significant advancements have come out of our AIMS program across many disease areas: cancer, neurology, immunity, infectious disease, inherited disorders, and others. A research team from the University of North Carolina, Chapel Hill, was able to validate a new target for pediatric cancer, Ewing’s Sarcoma, and find a potential new drug lead even without having prior structural information for the target. Another research team from Stanford University was able to validate Miro1 as a new target for Parkinson’s disease, and AtomNet® was able to discover a novel small molecule, now called the Miro1 reductase, and work towards the treatment of a disease that has long evaded drug development efforts. Working with a lab at the University of California, Riverside, we found hits for HTRA-1 that were extremely promising as drug leads for Age-Related Macular Degeneration (AMD), and we formed a joint-venture company, Theia Biosciences, to pursue development of the compound as a novel therapy.

Sick patients shouldn’t have to be patient, but cures are lacking for hundreds of millions of disease sufferers. We must make better progress in settings from rapidly-emerging zoonotic diseases (like COVID-19, SARS, and MERS), to rapidly-evolving infectious diseases (like XDR tuberculosis), to long-standing chronic diseases (like cancer and Alzheimer’s), but the drug discovery process as-is leaves no choice. Without AI allowing us to effectively scale targets prior researchers have deemed impossible, we’re only skimming the surface of what can be effectively cured or treated.

Learn more about our technology and how we are running the largest screens in human history to drug the undruggable at