A new AI testing system could help unlock secrets of the human genome
Artificial intelligence (AI) is an innovative tool that can be trained to make predictions and solve problems quickly and with accuracy. However, the reasoning behind the output, or information sent out after the AI software receives input from datasets, is not yet clearly understood.
Researchers have been trying to comprehend the way AI produces information and what rules and regulations the AI follows, or creates, as it processes data.
Understanding how AI creates its predictions
Peter Koo, a computational scientist and assistant professor at Cold Spring Harbor Laboratory (CSHL) in the U.S. wanted to figure out how AI generates its answers with precision. Instead of knowing and accepting that the correct or precise output is produced, he wanted to comprehend how the AI was creating its answers.
“If you learn general rules about the math instead of memorizing the equations, you know how to solve those equations. So rather than just memorizing those equations, we hope that these models are learning to solve it and now we can give it any equation and it will solve it,” Koo said.
Using the novel system to understand AI prognostications about the human genome
Koo and his team created a method to find patterns within comprehensive AI-generated answers and predictions. His research allowed him and his team to figure out which AI algorithms work best for creating a prediction when studying the human genome.
The human genome is the entire set of DNA found in each cell. It contains 23 pairs of chromosomes in the cell’s nucleus, along with small chromosomes in the cell’s mitochondria.
The human genome altogether contains three billion letters of code, with each individual having millions of variations. Although humans are not able to quickly sift through all the codes for DNA, artificial intelligence systems have the ability to do so. AI can also catch different factors that a person might miss while going through genomes.
The innovative process that uses AI for answers has created a phase where AI is processing the data, but it is difficult for researchers to figure out how it’s being processed. Koo also mentioned the difficulty in determining which new algorithms, a form of machine learning that tells AI how to 'learn', will work best since AI creates computations beyond human capacity.
The new method
Koo and his research team created a new computational machine learning process called the GenOmic Profile-model compreHensive EvaluatoR, also known as GOPHER. It is a novel method that allows researchers to determine the best AI program to utilize for analyzing the human genome. The process lets researchers assess various algorithms more methodically, without being a random selection.
The criteria GOPHER uses to analyze algorithms
The evaluator judges AI programs based on four criteria, which are how well the AI software learns the biology of human genome, how accurately AI predicts important patterns and features, the ability to handle background noise, and how interpretable the decisions are.
The system answers the question about how the AI generates certain predictions. “AI are these powerful algorithms that are solving questions for us,” says Ziqi Tang, a graduate student in Koo’s laboratory. “One of the major issues with them is that we don’t know how they came up with these answers.”
The most effective use of AI algorithms in the future
GOPHER helped Koo and his research team to unearth AI algorithms and understand how AI produces reliable and accurate answers. The system identifies the main essentials for creating the most efficient AI algorithms. The researchers believe that GOPHER will help optimize AI algorithms so that people can trust that they are learning the right and necessary things from the AI output. “If the algorithm is making predictions for the wrong reasons, they’re not going to be helpful”, said Shushan Toneyan, another graduate student at the Koo lab.
The results from the study were published yesterday, Dec. 5, in the journal Nature Machine Intelligence.
Koo and his team hope that AI can one day predict the prevalence of diseases by reading the human genome codes. In order to read the human genome and comprehend it, GOPHER can assess the algorithm being used by AI to create predictions.