Overview
Artificial intelligence (AI) and machine learning (ML) are revolutionizing drug discovery, substituting traditional high throughput screening (HTS) with computational methods. Conventional HTS fails to assess over 99% of commercially available molecules, significantly limiting the chemical search space. Atomwise, through its AIMS (Artificial Intelligence Molecular Screen) initiative, provides a breakthrough with its deep learning platform, AtomNet. This platform uses virtual HTS to navigate a synthetic chemical library of over 15 quadrillion compounds, identifying promising hits for various therapeutic targets.
Achievements in Computational Screening
According to a recent study published in Scientific Reports, AtomNet achieved a remarkable success rate of 74% in identifying novel hits for 235 out of 318 targets. This rate surpasses the conventional HTS success rate of about 50%. Here are some highlights from the study:
- Number of targets assessed: 318
- Number of successful hits: 235
- Success rate: 74%
This success was achieved in collaboration with over 250 academic labs across 30 countries, emphasizing the versatility and robustness of AtomNet.
Therapeutic Areas and Protein Classes
Atomwise’s technology shows promise across numerous therapeutic areas and protein classes. The platform’s hits spanned widely differing areas like oncology, infectious disease, neurology, immunology, and cardiovascular disease. The majority of target protein types included enzymes, GPCRs, transporters, ion channels, and DNA/RNA-binding proteins. Enzymes alone represented a significant 59% of the targets.
Major Therapeutic Areas and Protein Classes
| Therapeutic Areas | Protein Classes |
|---|---|
| Oncology | Enzymes |
| Infectious Disease | GPCRs |
| Neurology | Transporters |
| Immunology | Ion Channels |
| Cardiovascular Disease | DNA/RNA-binding Proteins |
Specific Case Studies
Examples of significant discoveries via AtomNet include inhibitors for challenging targets such as OTUD7A, OTUD7B, and CTLA-4. In Parkinson’s disease, AtomNet identified the first reducer for Miro1. Notable highlights also encompass the discovery of allosteric sites for protein-protein interactions, which are traditionally difficult to target.
The Unique Approach of AtomNet
The secret to AtomNet’s high success rate lies in its generalized model, pre-trained on extensive molecular data across the proteome. Unlike traditional AI/ML approaches that construct a new model for each protein, AtomNet’s broad training data allows it to handle various targets seamlessly.
-
Traditional AI/ML Approach:
- Create new ML models per protein
- Depend heavily on specific training data
-
AtomNet Approach:
- Pre-trained on diverse molecular data
- Generalizable across numerous targets
Future Prospects
Atomwise is set to break new ground in the inflammatory disease market. The company plans to file an IND application this year for its leading candidate, a novel allosteric TYK2 inhibitor, discovered using AtomNet. This move promises to bring innovative solutions to patients battling inflammatory diseases.
In conclusion, computational methods like those employed by Atomwise’s AtomNet are positioning themselves as vital alternatives to traditional HTS. By navigating a vastly expanded chemical space and providing robust, reproducible results across various therapeutic areas and protein classes, Atomwise is poised to significantly influence the future of drug discovery.
