Photo credit: venturebeat.com
Cerebras Systems has collaborated with the Mayo Clinic to develop a groundbreaking AI genomic foundation model aimed at predicting optimal medical treatments for individuals diagnosed with rheumatoid arthritis. This innovative model may also extend its predictive capabilities to patient treatment for cancer and cardiovascular diseases, according to Andrew Feldman, CEO of Cerebras Systems, who shared insights during an interview with GamesBeat.
At the JP Morgan Healthcare Conference in San Francisco, representatives from Mayo Clinic and Cerebras Systems showcased advancements in artificial intelligence tools designed to enhance patient care. These efforts reflect Mayo Clinic’s commitment to transforming healthcare through innovative approaches and technology.
While companies like Nvidia and various semiconductor firms are heavily invested in AI supercomputing, Cerebras adopts a distinctive strategy. Unlike Nvidia’s reliance on individual processors, Cerebras develops an entire wafer that houses multiple chips, enabling efficient solutions to complex AI challenges while consuming significantly less power. Feldman emphasized that deploying numerous Cerebras systems over several months to compute the genomic model was still more efficient in terms of time, effort, power, and cost compared to traditional computing methodologies. Looking ahead, PitchBook has suggested that Cerebras might pursue an IPO within the next two years.
By leveraging its AI capabilities, Cerebras Systems can now identify which treatments are likely to be effective for patients suffering from rheumatoid arthritis.
Building on Mayo Clinic’s strength in precision medicine, this genomic model is engineered to enhance diagnostics and tailor treatment options for patients, with an initial emphasis on rheumatoid arthritis (RA). Treating RA presents substantial clinical hurdles, often necessitating several attempts to identify effective medications for individual patients.
Previous methods that concentrated solely on individual genetic markers have had restricted success in anticipating treatment responses.
The collaborative team’s genomic model integrates publicly available human reference genome data along with extensive patient exome data from Mayo. The human reference genome serves as a digital representation of an “ideal” genome, allowing researchers to discover genetic variations among individuals.
Distinctively, Mayo’s genomic foundation model has outperformed models trained exclusively on reference genome data, achieving better results in genomic variant classification due to its comprehensive training data from 500 Mayo Clinic patients. The team anticipates ongoing enhancements in model efficacy as it incorporates more patient data.
To evaluate the model’s practical applications, the researchers established new benchmarks aimed at identifying specific medical conditions from DNA data. This focus addresses a gap in publicly available benchmarks that mainly target structural elements of DNA.
Cerebras Systems asserts that its predictive model for treatments exhibits notable accuracy.
The Mayo Clinic Genomic Foundation Model illustrates considerable accuracy across several critical metrics: achieving between 68% and 100% accuracy for RA benchmarks, 96% accuracy in predicting cancer predisposition, and 83% accuracy in forecasting cardiovascular phenotypes. These advancements align with Mayo Clinic’s goal of delivering premier healthcare via artificial intelligence. Feldman noted the necessity of additional tests to validate these findings.
“Mayo Clinic is dedicated to utilizing advanced AI technologies to develop models that will fundamentally change the landscape of healthcare,” stated Matthew Callstrom, Mayo Clinic’s medical director for strategy and chair of radiology. “Our partnership with Cerebras has allowed us to create a sophisticated AI model for genomics, yielding promising tools in under a year to assist our physicians with informed decision-making grounded in genomic data.”
“The Mayo genomic foundation model establishes a new benchmark in genomic models, excelling in not only standard tasks, like predicting the functional and regulatory aspects of DNA, but also in identifying intricate correlations between genetic variations and health conditions,” added Natalia Vassilieva, field CTO at Cerebras Systems. “This model empowers us to explore connections among multiple variants that may influence disease outcomes, as opposed to the current methodology that typically focuses on single variants.”
The accelerated creation of these models—often a lengthy, multi-year task—was made possible through training tailored models at Mayo on the Cerebras AI platform. The Mayo Genomic Foundation Model signifies a substantial advancement in clinical decision support and precision medicine.
Cerebras’ flagship product is the CS-3 system, which operates using the Wafer-Scale Engine-3 technology.
Advancing AI for Chest X-Rays
Additionally, during the same event, Mayo Clinic disclosed initiatives in partnership with Microsoft Research and Cerebras aimed at utilizing generative artificial intelligence (AI) to enhance patient care, expedite diagnostic processes, and boost accuracy.
These projects, announced at the J.P. Morgan Healthcare Conference, focus on creating customized foundation models that leverage multimodal data derived from various radiology images, such as CT scans and MRIs, as well as genomic sequencing information in collaboration with Cerebras.
The potential impact of these innovations could significantly alter the diagnostic and treatment landscape for clinicians, ultimately improving patient outcomes.
Foundation AI models are expansive, pre-trained structures that can adapt to various tasks with minimal additional training by learning from extensive datasets. This versatility renders them efficient components applicable across a multitude of AI applications.
Mayo Clinic and Microsoft Research are working synergistically to develop foundation models that amalgamate both text and images. For this project, teams from both institutions are examining the integration of generative AI in radiology, employing Microsoft’s advanced AI tools alongside Mayo Clinic’s extensive X-ray data.
Central to this research initiative is the goal of equipping clinicians with immediate access to crucial information. Mayo Clinic envisions a model capable of automating report generation, assessing the placement of tubes and lines in chest X-rays, and detecting changes in subsequent images. This proof-of-concept model aims to streamline clinician workflows and enhance patient care through comprehensive radiographic analysis.
With a patient base of approximately 76,000 individuals, Mayo Clinic handles a substantial number of cases annually.
“Our collaboration aims to leverage AI technology within healthcare, merging their expertise and extensive data with our AI proficiency and computing resources,” Feldman noted.
He elaborated that while large language models forecast word sequences, genomic models similarly predict nucleotide sequences. Changes in a nucleotide due to mutations or transcription errors can be linked to disease causation or its onset.
Currently available models generally inquire whether a change in a singular nucleotide can indicate a disease risk. In contrast, Cerebras’ approach evaluates multiple nucleotide variations, leading to a more precise predictive model.
“Our objective, together with Mayo Clinic, is to ascertain which medication is likely to be effective for a particular patient,” said Feldman.
The model incorporates one billion parameters, making it ten times larger than AlphaFold and trained on an expansive dataset comprising a trillion tokens. This robustness enhances its predictive accuracy, according to Feldman.
Frequently, patients undergo a frustrating trial-and-error process to determine suitable medications. However, Feldman expresses optimism that this model can streamline that process, specifically targeting rheumatoid arthritis, a condition impacting approximately 1.3 million Americans.
“Although this is still an early stage, our initial results indicate that we can predict the efficacy of drugs for specific patients with substantial accuracy,” he stated.
As a noteworthy milestone, the model demonstrated an accuracy rate of 87% in predicting treatments for arthritis, although the data requires publication and peer review before it can be conclusively validated.
Source
venturebeat.com