Written by: Matthew Busekroos | Produced by: Nate Caldwell
Originally from Lynnwood, Washington, Jason Yim was a research engineer at DeepMind and obtained his B.S. from Johns Hopkins University in computer science and applied mathematics prior to continuing his education at MIT’s Computer Science and Artificial Intelligence Laboratory. Yim’s research aims to develop machine learning methods in scientific domains, such as biology and chemistry. Yim credits the thriving biotech scene in Cambridge as an appealing factor to furthering his education here. He said the scene provides some of the best opportunities to train and do research in machine learning and biology.
Yim said the encouragement to collaborate across departments and with colleagues was appealing to him. He added that the ecosystem in Cambridge fosters the exploration of ambitious and cutting-edge ideas of combining AI and the sciences.
Additionally, Yim said he was drawn to the strengths of MIT in AI and ML from the theory to the applications while working on ML for biology at DeepMind. While there, he frequently read papers by Professors Tommi Jaakkola and Regina Barzilay.
“Their work had uncommon traits of being well principled in ML and the domain they focused on,” Yim said. “From language to healthcare to chemistry and now biology, it just seemed they knew very well how to bridge ML with high impact applications. They were at the top of my list of professors I wanted to be advised by.”
As advisors, Yim said, the trait that pops to mind about Tommi and Regina is their high standard of what research should be.
“At the beginning of my PhD, I would bring a lot of half-baked ideas to Tommi and Regina but they stressed quality and not quantity,” he said. “They want students to think deeply about the problem and not aim for incremental research or churning out papers for every conference.”
Yim said they strongly encouraged him to think about technical novelty and impact.
“This combination is often contradictory: ML research can be technically novel but not useful in unsolved applications,” he said. “It’s very challenging to satisfy both novelty and impact, but it can lead to new breakthroughs and influence how the field progresses. As a PhD student, there’s this temptation to go for low hanging fruits and publish lots of papers but I’m trying to resist and follow Tommi and Regina.”
Yim’s current research is on machine learning for protein design. In the past, Yim said he worked on diffusion models for generating novel and diverse protein structures. Part of this work was done in collaboration with the Institute for Protein Design at the University of Washington where they developed RFdiffusion, a state-of-the-art method for de novo protein design. The work was recently published in Nature and has led to breakthroughs in designing new proteins that can be experimentally validated. In tandem, Yim worked on FrameDiff in collaboration with researchers at Columbia University, University of Oxford, and CNRS, which can be thought of as a lightweight version of RFdiffusion that runs faster but at a cost of generation quality.
“FrameDiff is nice because it’s lightweight and easy for researchers to train for their particular applications,” he said. “I’m working on developing a follow-up that is both faster with improved generation quality and incorporates other data sources such as protein sequences and experimental measurements. I’ve started to work on active learning and adaptive experimental design which is how to incorporate ML into the workflow of biologists. I am part of multiple collaborations at MIT where we are using ML with wet-lab experiments.”
Yim said protein engineering is a crucial technique in biotechnology. He added that it can lead to developing new vaccines, enzymes, and reliable gene editing. Pharma companies and startups are interested in using ML to advance protein engineering. Yim’s research is developing new methods that companies and other labs can use in their research.
“I’m driven by the potential benefits of AI to revolutionize medicine and environmental causes,” he said. “I had family members I’ve lost to cancer. It would be my dream to find a cure to cancer or solve climate change.”
While Yim’s plans after CSAIL are still unclear, he said he is somewhere between doing a startup or going into academia. His dream job would be to run his own research group.
For more information on Jason Yim, check out his personal website: http://people.csail.mit.edu/jyim/.