More knowledge means more informed predictions. That’s the principle San Diego-based startup Entos is applying to revolutionize drug design with an AI-powered approach that enables a thousandfold acceleration in molecular properties prediction.
Drug discovery is a notoriously time-consuming and data-intensive process, but Entos’ OrbNet architecture changes that. It requires 30x less data to train a model for molecular drug discovery with quantum accuracy and 100x fewer experiments to find promising drug compounds — cutting through the waiting time and complexities associated with traditional therapeutic drug discovery methods.
The company, a member of the NVIDIA Inception program for startups revolutionizing industries with advancements in AI, data science and high performance computing, is advancing its work with NVIDIA Clara Discovery — a collection of state-of-the-art frameworks, applications and pretrained models built to unlock insights about how billions of potential drug molecules interact inside our bodies.
“Our physics-based approach means we include more qualities about the underlying quantum mechanics into the machine learning model,” said Tom Miller, CEO of Entos. “This enables us to make better predictions, even while using less data.”
Entos is focusing on identifying drug molecules that could deactivate proteins linked to certain forms of cancer. By including quantum mechanics calculations into its machine learning workflow, the startup can more quickly narrow the pool of potential compounds that bind with these target proteins.
Bringing AI Into the Drug Discovery Process
Machine learning is transforming the way scientists approach everything from climate science to drug discovery. Now the combination of AI and machine learning is leading to a new way of doing science, creating a hybrid of deep learning and physics-based simulation to transform the way drugs are discovered.
Drug discovery is a data-intensive process where researchers perform computationally dense calculations to simulate how molecules and proteins interact to identify the right therapeutics. Traditionally, these forms of quantum calculations are extremely expensive to perform and take weeks or months to complete.
These computational experiments benefit from the incorporation of AI and accelerated computing by allowing researchers to simulate the interaction of a drug with a protein at quantum accuracy. The simulations are far too computationally expensive to perform using traditional quantum mechanical calculations.
Entos optimizes its OrbNet drug discovery software on NVIDIA DGX A100 Tensor Core GPUs. The OrbNet AI model — developed jointly at Caltech by Entos CEO Tom Miller and Anima Anandkumar, director of machine learning research at NVIDIA — enables robotic synthesis and high-throughput experimentation to speed up therapeutic design.
“OrbNet uses a graph neural network built on domain-specific features that account for interactions between atoms. Further, we account for symmetries such as three-dimensional rotations,” said Anandkumar. “These design considerations make it possible to train OrbNet only on small molecules — those with less than 40 atoms — and directly apply the model on large protein molecules with a high degree of accuracy.”
By collaborating with NVIDIA experts, Miller’s team of scientists is able to perform high-throughput experimentation and “open new doors to what we can go after,” he said.
Unlocking the Potential of Covalent Bonds
Recent developments in machine learning are transforming the size and scale of chemical databases that researchers can trawl for promising drug compounds. AI models also allow scientists to study the innumerable chemical reactions of enzymes within the body in a new way.
Together, these advancements are enabling the study of entirely new classes of drugs that researchers were unable to investigate with previous methods.
One promising technique involves the creation of drug molecules that form covalent bonds with the target protein. If therapeutics can form these bonds with only the target protein, patients could be prescribed smaller doses and experience fewer side effects. Entos researchers plan to apply this method to disease areas including cancer, diabetes and cystic fibrosis.
Entos has forged partnerships with leaders in the pharmaceuticals, materials and chemical industries and raised $53 million in July to support its efforts to create meaningful therapies with high accuracy. The company credits active collaborations with the NVIDIA healthcare team as an asset in getting connected to technical resources and assistance in optimizing its applications on NVIDIA hardware architecture.