Montai Therapeutics Utilises NVIDIA's Multimodal AI for Advanced Drug Discovery
Montai Therapeutics, a company under the wing of Flagship Pioneering, has partnered with tech giant NVIDIA to make significant advancements in the field of drug discovery. Their collaboration centres around a cutting-edge multimodal AI platform designed to streamline the complex process of identifying new therapeutic agents.
Complexities in Drug Discovery
Developing new drugs involves finding therapeutic agents that can precisely target diseases without causing adverse side effects. Leveraging multimodal data—encompassing molecular structures, cellular images, genetic sequences, and unstructured data—offers valuable insights into potential drug candidates. However, integrating these diverse data types into a coherent AI model poses considerable challenges, such as aligning varied data and managing computational demands.
Montai's Strategy
Montai Therapeutics addresses these difficulties through the utilisation of NVIDIA’s BioNeMo platform. Central to Montai’s methodology is the aggregation and curation of the world’s most extensive annotated library of Anthromolecule chemistry. Anthromolecules include bioactive compounds that have been historically consumed by humans through foods, supplements, and herbal medicines. This collection presents a vast chemical diversity, often greater than that found in traditional synthetic chemistry libraries.
These anthromolecules and their derivatives have already provided several FDA-approved drugs for various ailments, but their potential remains largely underexplored in systematic drug development. The unique structural features of these compounds offer a wider array of possibilities for engaging complex biological targets with precision, potentially leading to more effective small molecule drugs.
Developing the AI Platform
Montai Therapeutics and the NVIDIA BioNeMo team have created a multimodal AI model designed to identify promising small molecule drugs from anthromolecule sources. Hosted on AWS EC2, the AI model is trained using several large-scale biological datasets and features NVIDIA's BioNeMo DiffDock NIM, a state-of-the-art generative model for molecular docking pose estimation. DiffDock NIM is part of NVIDIA's suite of microservices aimed at accelerating generative AI deployment across various platforms.
The collaborative efforts have resulted in significant optimisations in model architecture, leveraging a foundation built on contrastive learning. Early results indicate that the model significantly outperforms traditional machine learning methods for predicting molecular functions.
Integrating Multimodal Data
The breakthrough model integrates information across four key modalities: chemical structures, phenotypic cellular data, gene expression data, and biological pathway information. By unifying these modalities, the model demonstrates advantages over single-modality models, highlighting the efficacy of contrastive learning and foundation models in AI-driven drug discovery.
Through their proprietary CONECTA platform, Montai Therapeutics aims to leverage this model to identify and develop promising drug compounds from vast and underutilised human chemistry sources.
Future Prospects
The partnership is now set to integrate a fifth modality known as the "docking fingerprint," derived from DiffDock predictions. Utilising NVIDIA's BioNeMo platform has been crucial for scaling up the inference processes, with significant improvements in processing times. For instance, DiffDock on the DUD-E dataset, utilising eight NVIDIA A100 Tensor Core GPUs, achieves a processing speed of 0.76 seconds per ligand.
These strides in efficient GPU utilisation underscore the collaboration's role in enhancing drug discovery processes. Montai and NVIDIA's joint efforts mark a pivotal step towards creating more effective and efficient methodologies for developing new therapeutic agents.
Source: Noah Wire Services