IBM Research has open sourced two biomedical foundation models, offering the scientific community powerful tools for drug discovery.
The models include several different modalities (including sequences, graphs and images) and cover multiple domains (including targets, small molecules and biologics), giving them an advantage over other models in the open space.
“The main challenge and gap in existing models is addressing how to aggregate information from multiple modalities and connect them across multiple domains,” says Michal Rosen-Zvi, Director of AI for Drug Discovery at IBM Research. “We are addressing these gaps with the models we are open sourcing and the methodologies behind them.” Generative AI can streamline traditional drug development, which is often lengthy and costly. Typically, the development process involves four critical phases: identifying a target protein linked to a disease, confirming that interacting with this target can prevent symptoms or cure the condition, discovering a small molecule or biologic therapeutic that interacts with the target and optimizing the candidate therapy.
“Ultimately, we believe our models have the potential to drive scientific discovery with greater speed and at scale,” says Jianying Hu, an IBM Fellow and Director of Healthcare and Life Sciences Research.
The researchers say that the decision to open source both models also grew out of a conviction that safe and responsible AI foundation models need to be developed out in the open. To that end, IBM Research has started working in collaboration with Boston University, Red Hat and the Cleveland Clinic on a new AI Alliance working group dedicated to open-source AI applications in drug discovery. It will kick off at Boston University on 30 October 2024. “Open sourcing will allow the community to gather the best and brightest minds and work together to tackle the most challenging questions by addressing current gaps in computational methods for drug discovery,” says Rosen-Zvi.