Using AI for Protein-Protein Interactions
AI can be more cost effective and efficient than traditional screening approaches for the assessment of difficult, intractable and atypical targets such as protein protein interactions. By partnering with Charles River, clients gain access to AtomNet™, a patented AI platform developed by Atomwise for the prediction of small molecule binding to protein targets. AtomNet™ is the world’s first deep neural network for structure-based drug discovery applications.
Artificial intelligence is applied to predict binding affinity by incorporating structural information about the target, the ligand, and their mode of interaction to make predictions. This binding affinity prediction by the platform delivers in vivo success in blocking protein-protein interactions and target.
Drug discovery success using AtomNet™ AI
AtomNet™ is a convolutional neural network (CNN) – a computational mirror of how the brain labels, analyzes, and processes information. It is a statistical approach that extracts insight from millions of experimentally-determined small molecule binding affinity measurements and thousands of protein structures to predict binders to your protein of interest.
The following video illustrates a virtual screen in which AtomNet™ evaluates, one-by-one, millions of molecules to identify those that bind with high affinity to the human ROCK2 kinase.
The convolutional architecture of AtomNet™ allows it to learn multi-dimensional, universal features that dictate the favorable and unfavorable association between small molecules and proteins. The output of a virtual screen includes a list of compounds that are predicted to bind with high affinity. Further, the highest scoring compounds are clustered at the hit identification stage to increase the probability of identifying diverse scaffolds, each a distinct starting point for hit optimization.
When partnering with Charles River and Atomwise, clients can expect shorter timelines, reduced costs and the most commercially viable drug candidate. Our experienced scientists collaborate across teams to optimize all phases including early hit identification, hit-to-lead, lead optimization, patent strategy, and preparation for IND filing.
Artificial Intelligence in Drug Discovery Frequently Asked Questions (FAQs)
How is AtomNet™ different from other artificial intelligence and machine-learning drug discovery technology platforms?
AtomNet™ is a deep convolutional neural network for structure-based drug discovery. It is a single global model that can be used for small molecule discovery on any protein structure. Unlike ligand-based machine-learning methods, AtomNet™ does not require a diverse set of known actives and/or inactives for the target of interest. In addition, a critical difference is that the Atomwise technology is integrated into the full Charles River portfolio, allowing for a seamless experience from target discovery to IND studies.
How is AtomNet™ different from other traditional computational drug discovery methods?
Traditional computational chemistry drug discovery approaches often rely on physics-based scoring functions. Methods such as Free Energy Perturbation (FEP) require long simulation times and expensive computational resources, making them unattractive for the screening of large molecular libraries. Docking and related virtual screening methods make approximations to simplify the calculations, but effectively introduce error into binding affinity predictions. They are parametrized using limited structural protein-ligand complex data that is heavily skewed towards positives (binders) and inherently discounting the negatives (non-binders).
Instead, AtomNet™ uses a statistical approach which yields more robust predictions without sacrificing the speed of the calculations. It is trained on millions of experimentally determined small molecule binding affinity measurements and thousands of protein structures, learning from both positives (binders) and negatives (non-binders). It is capable of evaluating more than 100 million compounds per day. Screens can be performed against both high resolution X-ray crystal structures and homology models, which greatly expands the diversity of targets and off-targets that can be assessed. AtomNet™ employs a new paradigm which goes beyond the traditional molecular recognition strategies, circumventing the limitations and challenges of traditional physics-based computational chemistry approaches.
How is artificial intelligence used in drug discovery?
In silico, or computational, approaches can often be more effective than traditional approaches during the hit identification, hit-to-lead and lead optimization drug discovery stages. They allow for a significantly faster evaluation of a significantly larger chemical space than HTS or cell-based methods. Artificial intelligence can be especially useful for historically difficult epigenetic and atypical targets, as well as protein-protein interactions where no known ligands have been identified.
What is required when initiating a drug discovery program with AtomNet™?
At the minimum, the initiation of a screening program requires a protein sequence of the target of interest. Additionally, AtomNet TM performance benefits from the availability of an X-ray crystal or cryo-electron microscopy structure, and a characterized binding site. Available structures of homologous proteins will suffice in cases where no structures of the target are available.
What is AtomNet™ and what does it do?
AtomNet is an artificial intelligence platform, the world’s first deep convolutional neural network for structure-based drug discovery created by Atomwise. It allows for the rapid virtual screening of large small molecule libraries for the prediction of binders to a protein target of interest. It can be used effectively for the assessment of difficult, intractable and atypical targets such as protein-protein interactions.