At MIT, researchers have taken a big leap in making scientific experiments faster, more reliable, and less expensive—especially when studying complex systems like genetics or cancer. Rather than sticking to decades-old trial-and-error routines that struggle to keep up with today’s scientific questions, this new strategy introduces a refreshing way to handle huge numbers of treatment combinations, a challenge that’s long frustrated scientists everywhere.
Imagine being a scientist trying to figure out how different combinations of gene therapies might affect cancer growth. Traditionally, you’d be facing a process that’s both time-consuming and daunting—there are billions of possible combinations. Testing each one? Not a chance. If you only try a handful, your results could be skewed. There’s always a risk that some important interactions might slip by unnoticed.
To shake up this tedious process, MIT’s team built a probabilistic framework that changes the game completely. Instead of deciding ahead of time which combinations to test, scientists now assign treatments randomly, but in a way that still takes into account the dosage levels they care about. Each cell, for example, is exposed to various treatments in parallel, guided by carefully designed probabilities. This removes a lot of guesswork, reduces bias, and paints a fuller picture of how different treatments might interact with each other.
But there’s another important question: how can researchers find the right dosage for each treatment in order to understand their impact most precisely? MIT’s team came up with a novel answer. Think of dosage like tossing a weighted coin—a higher dosage means a “heads” is more likely, so that treatment is given more often; a lower dosage, and it’s less frequent. Over time, the experiment adjusts these “probabilities” based on feedback from earlier rounds. Each tweak edges the experiment closer to the most effective mixture and concentration, relying on data rather than gut feeling.
What’s really exciting is that researchers don’t have to settle for a “one and done” approach. As new results come in, they can fine-tune their strategy, making each round of experiments smarter than the last. This is especially useful when resources are tight or the data is noisy—in other words, real-world conditions that scientists face every day.
After running a series of simulations, the team found their new method consistently beat out traditional approaches when it came to predicting outcomes, particularly during experiments that ran in several phases. In the words of co-lead Jiaqi Zhang, there’s hope that this approach will open the door to answering some of the biggest questions in biology.
This new framework could transform how biological systems are explored, potentially leading to breakthroughs in how we treat genetic diseases, cancer, and more. The researchers aren’t stopping here—they’re looking to further refine their model, addressing challenges such as the ways different samples might influence each other or how selection biases creep into experiments. Real-world tests are next up, as they look to put their findings to the ultimate test.
This work was led by Jiaqi Zhang and Divya Shyamal, with Caroline Uhler as senior author, and drew support from MIT, Apple, various federal agencies, and several other backers. Their study made its debut at the International Conference on Machine Learning and could soon mark a turning point in how research is conducted around the world.
This website uses cookies.