3D protein structure prediction with AlphaFold2, OpenFold, and ESMFold in NVIDIA's BioNeMo.

3D protein structure prediction with AlphaFold2, OpenFold, and ESMFold in NVIDIA’s BioNeMo. [NVIDIA]

Chipmaker NVIDIA and cloud behemoth AWS have been partnering for years, and now the two companies are announcing that NVIDIA’s drug discovery generative AI platform, BioNeMo, is now available on AWS. Additionally, plans are underway for BioNeMo to be offered on AWS on NVIDIA DGX Cloud.

The alliance was announced at the AWS re:Invent event. Startups including Evozyne, Etcembly and Alchemab are early AWS users using BioNeMo for generative AI-accelerated drug discovery and development.

Evozyne focuses on creating novel proteins for therapeutic development, Etcembly is building a large machine learning database for immunology and TCR immunotherapies, and Alchemab specializes in identifying protective antibodies for hard-to-treat diseases.

Last week, NVIDIA announced a collaboration with Genentech with BioNeMo also playing a key role. NVIDIA has been ramping up its promotion of BioNeMo since earlier this year after formally launching it in 2022.

Accelerating drug research with NVIDIA GPUs and AWS infrastructure

BioNeMo offers an assortment of tools and models for drug discovery tasks such as target identification, protein structure prediction, and screening of drug candidates. This integration with AWS means that researchers can now tap the scalability of NVIDIA’s GPU-accelerated cloud servers.

For drug developers, BioNeMo can shed light on complex biomolecular structures while offering predictive modeling tools enabling pharma and biotech companies to explore molecular interactions, understand disease mechanisms.

In addition, NVIDIA and AWS announced an enhanced ability to train models on NVIDIA H100 Tensor Core GPUs. These GPUs offer almost linear scaling, slashing the time required for complex model training. This is part of a broader strategic collaboration which includes the development of a new supercomputing infrastructure with the NVIDIA Grace Hopper Superchip and AWS UltraCluster scalability.

The Grace Hopper Superchip combines the NVIDIA Grace and Hopper architectures, providing a CPU+GPU coherent memory model designed for giant-scale AI and HPC applications. It features a new 900 GB/s coherent interface. AWS UltraCluster, on the other hand, is a high-performance computing environment within Amazon EC2 that scales to thousands of GPUs or machine learning accelerators, such as AWS Trainium. AWS says it democratizes access to supercomputing-class performance for ML, generative AI, and HPC developers.

BioNeMo capabilities and AWS hardware news

BioNeMo supports diverse modeling capabilities thanks to its inclusion of models such as MegaMolBART and ProtT5. Both are deep learning models for bioinformatics with the former used for small molecule drug discovery and cheminformatics. ProtT5, conversely, is a protein language model that can translate between protein sequence and structure.

AWS also announced enhanced computing with AWS ParallelCluster and Amazon SageMaker. This includes the first cloud AI supercomputer featuring NVIDIA’s Grace Hopper Superchip, which is optimized for generative artificial intelligence applications. AWS ParallelCluster simplifies the deployment and management of High Performance Computing (HPC) clusters, allowing users to focus on computational tasks such as complex computations and simulations in scientific research, including pharmaceutical research and development. SageMaker is a managed service for machine learning.

Outside of NVIDIA and AWS, several big tech companies and smaller players have unveiled AI and machine learning offers relevant to drug discovery over the past roughly decade. For instance, Atomwise, founded in 2012, applies deep learning to structure-based small molecule drug discovery. DeepMind, established two years earlier and later acquired by Alphabet, has made waves with AlphaFold, a deep learning program for predicting protein structures. Insilico Medicine, founded in 2014, focuses on end-to-end drug discovery using deep learning and generative AI, targeting various stages of the drug development pipeline. Finally, BenevolentAI, which emerged in 2013, applies machine learning and deep learning to analyze and interpret biomedical knowledge graphs, enabling the discovery of novel drug targets, the prediction of drug efficacy, and the identification of potential new uses for existing drugs.