Artificial intelligence has taken the world by storm. In biology, AI tools called deep neural networks (DNNs) have proven invaluable for predicting the results of genomic experiments. Their usefulness has these tools poised to set the stage for efficient, AI-guided research and potentially lifesaving discoveries—if scientists can work out the kinks.
"Right now, there are a lot of different AI tools where you'll give an input, and they'll give an output, but we don't have a good way of assessing the certainty, or how confident they are, in their answers." explains Cold Spring Harbor Laboratory (CSHL) Associate Professor Peter Koo . "They all come out in the same format, whether you're using a large language model or DNNs used in genomics and other fields of biology."
It's one of the greatest challenges today's researchers face. Now, Koo, former CSHL postdoc Jessica Zhou , and graduate student Kaeli Rizzo have devised a potential solution—DEGU (Distilling Ensembles for Genomic Uncertainty-aware models). DNNs trained using DEGU are more efficient and more accurate in their predictions than those learning via standard methods.
"When we want to make claims in biology, we don't want to rely on a single model," Koo explains. "For example, we might train 10 models, and we'll get predictions from each. Typically, we'd then use a method called deep ensemble learning to see where they agree and disagree. But handling 10 models and ensembles is challenging, especially as the model sizes grow. That's where DEGU comes in."
DEGU is built on a previously developed method called "deep ensemble distribution distillation," which focuses on learning a DNN's overall distribution of predictions rather than its individual estimates. Regardless of how many models are used, DEGU distills the resulting ensembles down to one much more manageable tool. Koo, Zhou, and Rizzo found that models trained using this distillation process provided better predictions—and better explanations for those predictions— than those without it. They also required less power.
"Instead of needing to analyze 10 models at once, you're working with a single model one-tenth the size with the same predictive capabilities," Rizzo explains. "And because you're only working with one model, it's easier to understand what's driving its predictions and uncertainty."
The Koo lab is now working to improve DEGU's efficiency and make it more accessible to researchers worldwide.
"Lab experiments are expensive," Rizzo says. "If we can make our models as reliable as possible for downstream applications, scientists will spend less time chasing predictions that a model isn't even that confident about." Fewer wild goose chases and better bases for strong hypotheses—that's the goal. Now, we'll see how DEGU delivers.
