The human genome comprises three billion letters of code, and each human has millions of combinations. Although no human can accurately scrutinize all that code, computers can be programmed to do so.
Artificial intelligence (AI) programs are capable of discovering patterns in the genome associated with diseases much more quickly than humans can. They also detect things that humans overlook.
In the future, AI-driven genome readers could perhaps predict the occurrence of diseases from cancer to the common cold. Regrettably, AI’s current popularity surge has resulted in a bottleneck in innovation.
It’s like the Wild West right now. Everyone’s just doing whatever the hell they want.
Peter Koo, Assistant Professor, Cold Spring Harbor Laboratory
Similar to how Frankenstein’s monster was a combination of diverse parts, AI scientists are continuously formulating new algorithms from different sources. Furthermore, it is hard to assess whether their designs will be good or bad. Basically, how can researchers assess “good” and “bad” when handling computations that surpass human capabilities?
That is where GOPHER, the latest creation of Koo lab, enters. GOPHER (an acronym for GenOmic Profile-model compreHensive EvaluatoR) is a new technique that assists scientists in recognizing the most efficient AI programs to examine the genome.
“We created a framework where you can compare the algorithms more systematically,” explains Ziqi Tang, a graduate student in Koo’s laboratory.
GOPHER evaluates AI programs on numerous criteria: how precisely they predict vital features and patterns, their ability to manage background noise, how well they learn the biology of the human genome, and how decipherable their conclusions are.
“AI are these powerful algorithms that are solving questions for us,” says Tang. But, she observes: “One of the major issues with them is that we don’t know how they came up with these answers.”
GOPHER assisted Koo and his team to unearth sections of the AI algorithms that promote performance, reliability, and accuracy. The outcomes help establish the central building blocks for building the most efficient AI algorithms in the future.
“We hope this will help people in the future who are new to the field,” says Shushan Toneyan, another graduate student at the Koo lab.
Visualize feeling ill and being able to establish precisely what is causing the illness at the click of a button. AI could, in the future, transform this science-fiction trope into a real feature found in every physician’s office.
Akin to video-streaming algorithms that learn the preferences of users based on their viewing history, AI programs may detect distinctive features of the human genome that could help bring about personalized treatments and medicine. The Koo team anticipates GOPHER will help enhance such AI algorithms so that we can be sure they are learning the correct things for the correct reasons.
If the algorithm is making predictions for the wrong reasons, they’re not going to be helpful.
Shushan Toneyan, Graduate Student, Koo Lab, Cold Spring Harbor Laboratory
The study received funding from the Simons Center for Quantitative Biology at Cold Spring Harbor Laboratory, National Institutes of Health.
Toneyan, S., et al. (2022) Evaluating deep learning for predicting epigenomic profiles. Nature Machine Intelligence. doi.org/10.1038/s42256-022-00570-9.