When Sachit Saksena was studying computational oncology in college, he often found himself advocating for the power of machine learning. “There was kind of this skepticism from the traditional modeling people” he said, who use knowledge of biological behaviors to design models. “There’s this notion that you don’t get generalizable models by trial and error,” using algorithms to hunt for lucky patterns in reams of data.
But biological data were exploding in scale, and Saksena realized that biologists didn’t have much of a choice in the matter. By the time he was doing his Ph.D. at MIT, researchers were drowning in snapshots of complex biological processes from single-cell sequencing. Traditional modelers might have tried to stitch them together into a time series using differential equations — “but you couldn’t do that with this type of data, because it was so high-dimensional,” said Saksena. So he developed a machine learning method that simulated cellular differentiation in real time
Today, he’s extending that work for one of the stealth labs under Flagship Pioneering, working to apply data-driven machine learning methods to remove some of the guesswork from drug discovery. While he’s still a strong advocate for machine learning, he’s also aware of its pitfalls — especially when it comes to the patient at the end of the drug discovery pipeline. “A lot of technology can sometimes add cost,” said Saksena. “I hope machine learning doesn’t go that route, and we really use these things to make things better and bring down costs and get medicines to people faster.”
— Katie Palmer