Drug Discovery Aims to Develop New Therapeutic Agents
Drug discovery aims to develop new therapeutic agents that effectively target diseases while minimizing side effects for patients. Using multimodal data—such as molecular structures, cellular images, sequences, and unstructured data—can be highly valuable in identifying novel and safe drug candidates.
Creating a Multimodal AI Drug Discovery Platform
Creating multimodal AI models for computer-aided drug discovery is challenging. These models must align diverse types of data and handle significant computational complexity. Multimodal AI models often require advanced deep learning techniques, which are computationally demanding. Ensuring that these models use information from all data types effectively without introducing bias is a major difficulty.
Montai Therapeutics’ Approach
Montai Therapeutics, a Flagship Pioneering company, is tackling these challenges using the NVIDIA BioNeMo platform. At the core of Montai’s innovation is the aggregation and curation of the world’s largest, fully annotated library of Anthromolecule chemistry.
The Role of Anthromolecules
Anthromolecules and their derivatives, with scaffolds designed to modulate biology, offer far greater chemical structural diversity than traditional synthetic combinatorial chemistry libraries. This diversity is critical for two reasons: Anthromolecule chemistry has already been proven to be a source of FDA-approved drugs for various diseases, yet it remains largely untapped for systematic drug development. The rich topological structures across this diverse chemistry offer a far wider range of vectors to engage complex biology with precision and selectivity and can unlock small molecule pill-based solutions for targets that have historically eluded drug developers to transform therapeutic offerings across chronic diseases.
Creating a Multimodal AI Model
In a recent collaboration, Montai and the NVIDIA BioNeMo solution team have made significant strides in developing a multimodal model aimed at virtually identifying potential small molecule drugs from Anthromolecule sources. The model, built on AWS EC2, is trained on multiple large-scale biological datasets. It incorporates NVIDIA BioNeMo DiffDock NIM, a state-of-the-art generative model for blind molecular docking pose estimation. BioNeMo DiffDock NIM is part of NVIDIA NIM, a set of easy-to-use microservices designed to accelerate deployment of generative AI across cloud, data center, and workstations.
Model Architecture Optimization
The collaboration between Montai and NVIDIA has produced notable model architecture optimization on the backbone of a contrastive learning foundation model. Initial results are promising, with the model demonstrating superior performance to traditional machine learning methods for molecular function prediction.
Multimodal Model
The multimodal model unifies information across four modalities:
- Chemical structure
- Phenotypic cell data
- Gene expression data
- Information about biological pathways
The combined use of these four modalities has resulted in a model that outperforms single-modality models, demonstrating the benefits of contrastive learning and foundation model paradigms in the AI for drug discovery space.
Conclusion
Currently, the collaborative efforts are focused on incorporating a fifth modality, the “docking fingerprint,” derived from DiffDock predictions. The role of NVIDIA BioNeMo has been instrumental in scaling up the inference process, enabling more efficient computation. For example, DiffDock on the DUD-E dataset, with 40 poses per ligand on eight NVIDIA A100 Tensor Core GPUs, achieves a processing speed of 0.76 seconds per ligand.
FAQs
Q: What is the goal of drug discovery?
A: The goal of drug discovery is to develop new therapeutic agents that effectively target diseases while minimizing side effects for patients.
Q: What is the challenge in creating a multimodal AI model for drug discovery?
A: The challenge lies in aligning diverse types of data and handling significant computational complexity.
Q: What is Anthromolecule chemistry, and how does it differ from traditional synthetic combinatorial chemistry libraries?
A: Anthromolecule chemistry refers to the rigorously curated collection of bioactive molecules humans have consumed in foods, supplements, and herbal medicines. It offers far greater chemical structural diversity than traditional synthetic combinatorial chemistry libraries.
Q: What is the role of NVIDIA BioNeMo in the development of the multimodal AI model?
A: NVIDIA BioNeMo played a crucial role in scaling up the inference process, enabling more efficient computation, and providing a state-of-the-art generative model for blind molecular docking pose estimation, BioNeMo DiffDock NIM.
