In order to speed up research in the areas of genomics, chemistry, biology, and molecular dynamics, NVIDIA today unveiled an enhanced set of generative AI cloud services for designing AI foundation models.
The newest BioNeMoTM Cloud service offering, a component of NVIDIA AI Foundations, speeds up the most time-consuming and expensive phases of drug development for both AI model training and inference. The ability to perform AI model inference directly in a web browser or using new cloud application programming interfaces (APIs) that are simple to incorporate into existing applications allows researchers to fine-tune generative AI systems on their own proprietary data.
According to Kimberly Powell, vice president of healthcare at NVIDIA, “the transformative power of generative AI holds enormous promise for the life science and pharmaceutical industries.” “BioNeMo Cloud Service, which is already functioning as an AI drug discovery laboratory, was developed as a result of NVIDIA’s extensive collaboration with industry pioneers. It offers pre-trained models and permits model customization using proprietary data to support all phases of the drug discovery pipeline, assisting researchers in choosing the appropriate target, designing compounds and proteins, and predicting how they will interact with the body.
Amgen Early Adopters
One of the top biotechnology businesses in the world, Amgen, is already utilizing the service to boost its R&D initiatives.
According to Peter Grandsard, executive director of Amgen’s Center for Research Acceleration through Digital Innovation’s Biologics Therapeutic Development, “BioNeMo is dramatically accelerating our approach to biologics discovery.” With it, we are able to research and create therapeutic proteins for the upcoming generation of patient-helping medicines by pretraining huge language models for molecular biology using Amgen’s confidential data.
Generated AI Accelerates the Drug Discovery Process
Pretrained AI models are part of the BioNeMo Cloud service, which assists researchers in creating AI drug development pipelines. It has been implemented to promote data-driven drug design for new therapeutic prospects by drug discovery businesses like Evozyne and Insilico Medical.
In some circumstances, generative AI models can create new substances or protein-based medicines from scratch to quickly find promising medication molecules. These models can forecast the 3D structure of a protein and how successfully a molecule would dock with a target protein after being trained on vast datasets of small molecules, proteins, DNA, and RNA sequences.
Early Access to the BioNeMo Service Unveils Novel Generative AI Models
In addition to its previously disclosed MegaMolBART generative chemistry model, ESM1nv protein language model, and OpenFold protein structure prediction model, BioNeMo now has six new optimized, open-source models. They consist of:
AlphaFold2: The DeepMind-developed AlphaFold2 deep learning algorithm, which is now being used by over a million academics, cuts the time it takes to ascertain a protein’s structure from years to minutes or even seconds.
DiffDock: This model predicts the 3D orientation and docking interaction of tiny molecules with great accuracy and computational efficiency, assisting researchers in understanding how a therapeutic molecule will bond with a target protein.
ESMFold: With the help of Meta AI’s ESM2 protein language model, ESMFold is a protein structure prediction model that can infer the three-dimensional (3D) structure of a protein from a single amino acid sequence without the need for examples of multiple similar sequences.
ESM2: This protein language model is used to infer computer representations of proteins that are helpful for later tasks including molecular docking, protein structure prediction, and property prediction.
MoFlow: This generative chemistry model generates molecules from scratch, producing a variety of chemical structures for possible treatments. It is used for molecular optimization and small molecule production.
ProtGPT-2: To assist researchers in creating proteins with distinctive architectures, characteristics, and functions, this language model develops innovative protein sequences.
The BioNeMo Service provides a browser-based interface for interactive inference and protein structure visualization, making these generative AI models simple to access. Additionally, researchers can modify their models on a fully managed software service using NVIDIA Base CommandTM Platform and the NVIDIA AI Enterprise software suite by combining BioNeMo with the supercomputing resources in NVIDIA DGXTM Cloud.
Pharmaceutical firms, startups Utilize BioNeMo to enhance AI processes.
Drug discovery startups and pharmaceutical corporations are employing BioNeMo today and, in many cases, are achieving notable results.
Amgen used its own confidential data on antibodies to pre-train and improve the ESM model architecture of BioNeMo. The time needed to train five unique models for molecule screening and optimizations was reduced from three months to a few weeks using the DGX Cloud.
Researchers from Chicago-based biotechnology business Evozyne, a participant in the NVIDIA Inception program for cutting-edge entrepreneurs, worked with NVIDIA to create the Protein Transformer Variational AutoEncoder, a BioNeMo-based deep learning model. The generative AI model may create synthetic variations with much better performance compared to naturally occurring enzymes according to Evozyne’s proprietary protein data.
The early drug development process typically takes more than four years and costs over $500 million. Insilico Medical, a leading member of NVIDIA Inception, is employing BioNeMo to speed up this process. Insilico was able to find a preclinical candidate drug in one-third the time and at one-tenth the cost by utilizing generative AI from beginning to end. The medication is anticipated to start patient phase 2 clinical trials soon.
