Cerebras Systems is working with Mayo Clinic to develop a genomic model that predicts arthritis treatment

Cerebras Systems Have cooperated with Mayo Clinic To create an AI-based genomic model that predicts the best medical treatments for people with rheumatoid arthritis.

This could also be useful in predicting the best treatment for people with cancer and cardiovascular disease, said Andrew Feldman, the company’s CEO Cerebras SystemsIn an interview with GamesBeat.

Mayo Clinic, in collaboration with Cerebras Systems, announced major progress in developing artificial intelligence tools to enhance patient care, today at the JP Morgan Healthcare Conference in San Francisco.

As part of Mayo Clinic’s commitment to transforming health care, the organization led the development of a world-class genomic foundation model, designed to support clinicians and patients.

Like Nvidia and other semiconductor companies, Cerebras has focused on supercomputing for artificial intelligence. But its approach is very different from that of Nvidia, which relies on individual AI processors. Cerebras Systems is designing an all-chip — containing many chips on a single silicon wafer — that collectively solves big AI problems and other computing tasks while consuming much less power. It took dozens of these systems to compute the genomic basis model over months of time, Feldman said. However, he said, this requires much less time, effort, power and cost than traditional computing solutions. PitchBook recently predicted that Cerebras will have an IPO in 2025.

Building on Mayo Clinic’s leadership in precision medicine, the model is designed to improve diagnosis and personalize treatment selection, with an initial focus on rheumatoid arthritis (RA). Treatment of rheumatoid arthritis represents a major clinical challenge, often requiring multiple attempts to find effective drugs for individual patients.

Traditional approaches examining individual genetic markers have shown limited success in predicting response to treatment.

The joint team’s genomic model was trained by blending publicly available human reference genome data with Mayo’s comprehensive patient exome data. The human reference genome is a digital DNA sequence that represents a “perfect” composite copy of the human genome. It serves as a standard framework against which individual human genomes can be compared, enabling researchers to identify genetic differences.

In contrast to models trained exclusively on the human reference genome, Mayo’s genomic foundation model shows much better results in classifying genomic variants because it was trained on data sourced from 500 patients at Mayo Clinic. As more patient data is incorporated into training, the team expects the quality of the model to continue to improve.

The team designed new criteria to evaluate clinically relevant model capabilities, such as detecting specific medical conditions from DNA data, addressing a gap in publicly available criteria, which primarily focus on identifying structural elements such as regulatory or functional regions.

The Mayo Clinic Genomic Model demonstrates cutting-edge accuracy in several key areas: 68-100% accuracy in criteria for rheumatoid arthritis, 96% accuracy in predicting cancer, and 83% accuracy in predicting cardiovascular phenotype. These capabilities align with Mayo Clinic’s vision to deliver world-leading health care through AI technology. Feldman said more tests will need to be done to verify the results.

“Mayo Clinic is committed to using the most advanced AI technology to train models that will transform health care,” Matthew Kallstrom, M.D., medical director of strategy and chief of radiology at Mayo Clinic, said in a statement. “Our collaboration with Cerebras has enabled us to create a state-of-the-art AI model for genomics. In less than a year, we have developed promising AI tools that will help our doctors make more informed decisions based on genomic data.”

“Mayo’s genomic foundation model sets a new standard for genomic modeling, excelling not only at standard tasks such as predicting the functional and regulatory properties of DNA but also enabling the discovery of complex associations between genetic variants and medical conditions,” said Natalia Vasilieva, field technology manager at Cerebras. Systems, in a statement. “In contrast to existing approaches that focus on single-variable associations, this model allows for the discovery of associations where combinations of variables contribute to a given condition.”

The rapid development of these models — an endeavor that typically takes several years — was accelerated by training custom Mayo Clinic models on the Cerebras AI platform. The Mayo Genomic Model represents important steps toward enhancing clinical decision support and advancing precision medicine.

Cerebras’ flagship product is the CS-3, a system powered by the Wafer-Scale Engine-3.

Development of artificial intelligence for chest X-rays

Separately, Mayo Clinic today unveiled separate groundbreaking collaborations with Microsoft Research and with Cerebras Systems in obstetric artificial intelligence (AI), designed to personalize patient care, dramatically speed time to diagnosis and improve accuracy.

The projects, announced at the JP Morgan Healthcare Conference, focus on developing and testing custom core models for various applications, leveraging the power of multimodal radiology images and data (including CT and MRI) with Microsoft Research and genomic sequencing data with Cerebras.

Innovations have the potential to change the way doctors approach diagnosis and treatment, ultimately leading to better outcomes for patients.

Foundational AI models are large, pre-trained models that are able to adapt to and perform many tasks with minimal additional training. They learn from large data sets, gaining general knowledge that can be used across diverse applications. This adaptability makes them efficient and versatile building blocks for many artificial intelligence systems.

Mayo Clinic and Microsoft Research are working collaboratively to develop core models that combine text and images. For this use case, Mayo and Microsoft Research are working together to explore the use of generative AI in radiology using Microsoft Research’s AI technology and Mayo Clinic’s x-ray data.

Enabling doctors to have immediate access to the information they need is the core of this research project. Mayo Clinic aims to develop a model that can automatically generate reports, evaluate tube and line placement on chest X-rays, and detect changes from previous images. This proof-of-concept model seeks to improve physician workflow and patient care by providing more efficient and comprehensive analysis of radiographs.

Mayo Clinic employs 76,000 people and sees a large number of patients annually.

“We started a partnership to bring AI technology to healthcare. This allowed us to kind of combine their domain expertise and great data with our AI expertise and computation,” Feldman said.

He said that large language models predict words, but genomic models predict nucleotides. When a nucleotide is flipped in a mutation or transcription error, it may be the cause of a disease or can predict the onset of the disease.

Current models can only ask whether a single nucleotide flip predicts disease. But Cerebras looks at the permutation of more than one nucleotide and comes up with a more accurate model.

“What we’re using it for, in collaboration with Mayo Clinic, is to predict which drug will work for a particular patient,” Feldman said.

It’s a billion-parameter base model, or 10 times larger than AlphaFold, and was trained on a trillion codes. This makes it more accurate, Feldman said.

Often, patients have to go through a process of trial and error to figure out which medication will work. But with this model, Feldman believes he can predict which drug will work for a particular person. The first target is rheumatoid arthritis, which affects 1.3 million Americans.

“Although it is still early days, what we have been able to show is that we have been able to predict with amazing accuracy which drug will work for a given patient,” he said.

For arthritis, prediction accuracy was 87%. The data has yet to be published and peer-reviewed.