Merck and Nvidia team up on new drug discovery model

There’s no shortage of AI models aimed at discovering new drugs these days. Researchers have introduced more than 200 of these foundation models in the last three years, with 40% growth per quarter, according to a new paper published this month in the journal Drug Discovery Today.

Pharmaceutical giant Merck and Nvidia recently rolled out a new entry in this growing body with a small-molecule drug model called KERMT. The model is pretrained on more than 11 million molecules, then fine-tuned for various tasks specific to industrial drug discovery workflows, according to Merck.

Alan Cheng, Merck’s senior director of data science, told us the model could help scientists better predict how a given molecule will behave in the body, potentially catching problems before researchers invest in months of testing.

“Traditionally, scientists spend months running physical and biological tests to understand ADMET properties: absorption, distribution, metabolism, excretion, and toxicity,” Cheng said in an email interview. “These steps are essential because a promising compound can fail late in development if it is toxic or has poor exposure at the therapeutic target.”

The KERMT name is a Muppet-themed nod to the underlying GROVER model, a neural network pretrained on millions of small molecules. Merck’s scientists found that multitask-based fine-tuning of the enhanced GROVER-based KERMT improved its predictive performance across both public material and Merck’s internal datasets.

Pills + parameters. Merck isn’t the only drug company getting cozier with tech companies as AI-powered drug discovery heats up. Eli Lilly and Nvidia recently teamed up on a supercomputer for drug discovery, and Novo Nordisk has partnered with Anthropic and AWS. Johnson & Johnson is working with Nvidia on medical robotics, and Anthropic recently launched Claude for Life Sciences, aimed at expediting biotech R&D.

For Merck, working directly with Nvidia allows it to more efficiently scale models across multiple processors. With Nvidia’s infrastructure help, Merck was able to speed up KERMT’s calculation time and reduce the amount of memory it used, compared to GROVER. “These improvements are eco-friendly—KERMT is green,” Cheng wrote.

He said AI is already “speeding up the early stages of drug development dramatically.”

“AI is…sometimes cutting timelines by 30% or more, improving drug candidate quality and reducing costs,” Cheng said. “This is a very meaningful acceleration. While clinical trials remain lengthy, our models…are enabling faster identification of disease targets and optimized compounds, which should increase probability of success and shorten preclinical phases.”

Drugs pending. Yet after three years of drug discovery foundation models, the tech has yet to yield fully approved drugs. The pipeline process for new drugs can take years, however, and nine in 10 clinical drug candidates fail.

Cheng said areas like “oncology, cardiovascular, and immunology, where we have a promising pipeline, are likely to see early impact.” KERMT and TEDDY, Merck’s genomics-focused foundation model, both already “have versions deployed directly to scientists’ desktops and are embedded in real-time design cycles,” according to Cheng.

Merck is currently focused on scaling up these models further and expanding multimodality, drawing on other fields like physics. The company also wants to make them more widely accessible to the scientific community.

“In the spirit of open science, we have published our investigations with KERMT as a scientific manuscript,” Cheng said. “Since both KERMT and TEDDY are open-source, they have the potential to accelerate success across the industry and community.”