How MIT & Harvard’s Broad Institute is Accelerating Cellular Program Discovery At Scale

Customer: Dr. Adit Radha, Dr. Yajit Jain, Dr. Caroline Uhler & Dr. Eric Lander — Researchers at the Broad Institute and MIT

Industry: Computational Biology / Genomics

Summary

Researchers at the Broad Institute have been at the forefront of genomic sequencing, pushing past the limits of what’s possible with traditional infrastructure. The vast datasets and complex calculations involved have required specialized compute - previously only available at high prices and with rigid contracts. Mithril has provided the Broad Institute with flexible yet high-performance compute at 30% of the cost, enabling the next breakthrough.

Challenge

The rise of single-cell RNA sequencing in the last decade has opened up a universe of data to biologists and researchers. Tens of trillions of cells, each expressing tens of thousands of genes present incredible challenges and opportunities alike. Hidden within the data are cellular programs: networks of genes that activate together to drive essential functions, such as hemoglobin production or cell division. By identifying and referencing “normal” cellular programs, scientists can pinpoint “altered” cellular programs and develop targeted therapeutics.

Scientists at the Broad - an independent research institute focused on genomics and jointly sponsored by MIT & Harvard - had previously relied on traditional infrastructure to perform the complex matrix calculations required to uncover cellular programs. However, the scope of the Broad team’s research was unprecedented. Rather than focusing on a single cell type or study, the researchers performed an analysis of 460 human cell types using a publicly accessible single cell RNA sequencing dataset containing 30 million curated cells and covering 20,000 thousand genes. While the team had access to in-house GPU compute, each computation run took more than a week to complete, preventing them from quickly iterating and developing their methods. Hyperscaler clouds offered the capacity, but at three times the cost and the subsequent financial maneuvering to optimize those costs was logistically harder to scale.

Solution

Mithril’s spot-market and short-term reservations gave the Broad team the power, flexibility, and price-performance needed to accelerate their research. Over several months, the team ran iterative experiments on millions of samples using Mithril-provisioned A100 GPUs, uncovering novel gene clusters that standard correlation methods had missed. Compared to other platforms, Mithril provided several distinct advantages:

• Significantly lower cost: GPU hours were 70% cheaper than hyperscalers, enabling the Broad to run three times as many experiments while exploring and refining different techniques.
  
  
  

• On-demand scalability: Mithril’s spot market provided rapid, cost-efficient cluster provisioning. With the ability to right-size compute in real time as new learnings emerged. Broad scientists completed the study in under a month, work that would have taken a year using traditional compute.
  
  
  

• No long-term commitments: The team could access the compute in real time without being locked into long-term contracts or risking underutilized resources

“Through Mithril, computations that used to take up to a week to run on traditional infrastructure, can now be done in hours on a single GPU across 30 million cells and 460 cell types. None of this would have been possible at this scale otherwise.”

~ Dr. Adit Radha, MIT Mathematics and Broad Institute

Impact

The combination of GPUs and large public sequencing datasets gives researchers an unprecedented ability to decode the mechanisms underlying human biology. With Mithril’s unique flexibility and cost structure, Broad scientists can continue exploring cellular programs, laying the groundwork for deeper insight and more precise therapeutics.

"Uncovering and cataloging the cellular programs that run in the human body is the next frontier for understanding disease. Mithril's infrastructure has let us uncover these programs across tens of millions of cells and hundreds of cell types in a matter of hours, turning a long-term vision into something we can do today."

~ Dr. Eric Lander, Broad Institute of MIT and Harvard

About Mithril

Founded in 2022 with backing from Lightspeed Ventures and Sequoia Capital, Mithril empowers AI teams to access compute resources, scale rapidly, and focus on research rather than infrastructure. Create your instance here or contact sales at sales@mithril.ai