Consider looking at millions upon millions of genetic mutations. With CRISPR gene-editing technology a select few of these mutations might have therapeutic potential, but discovering and then validating which would involve a considerable amount of lab work, and cost. But what if it was possible to achieve this virtually, using artificial intelligence (AI)?
Researchers at the Cold Spring Harbor Laboratory (CSHL), headed by assistant professor Peter Koo, PhD, and his team, have developed an AI-powered virtual laboratory, cis-regulatory element model explanations (CREME), that allows geneticists to run thousands of virtual experiments with the click of a button. Using CREME, scientists can use the tool to begin identifying and understanding key regions of the genome.
Koo and colleagues reported on development of CREME, in Nature Genetics, in a paper titled “Interpreting cis-regulatory interactions from large-scale deep neural networks,” in which they stated “CREME can provide interpretations across multiple scales of genomic organization, from cis-regulatory elements to fine-mapped functional sequence elements within them, offering high-resolution insights into the regulatory architecture of the genome.”
“The rise of large-scale, sequence-based deep neural networks (DNNs) for predicting gene expression has introduced challenges in their evaluation and interpretation,” the authors wrote. Current evaluations align DNN predictions with experimental data, and while these approaches provide insights into generalization, they may offer only limited insights into their decision-making process, the team continued. “… the extensive sequence size of large-scale DNNs presents a challenge when evaluating their predictions and interpreting learned patterns.”
Current methods for evaluating large-scale models have relied on assessing the alignment between predictions and existing experimental perturbation assays, such as CRISPR interference (CRISPRi) technology, the authors further noted. CREME draws inspiration from CRISPRi, a genetic perturbation technique based on CRISPR, which allows biologists to turn down the activity of specific genes in a cell. CREME is almost akin to an AI version of CRISPRi, and lets scientists make similar changes in the virtual genome and predict their effects on gene activity. “Here we present cis-regulatory element model explanations (CREME), an in silico perturbation toolkit that interprets the rules of gene regulation learned by a genomic DNN,” the team commented. “CREME provides a suite of in silico experiments for unbiased interpretations of large-scale sequence-based DNNs, enabling CRE-level analysis similar to CRISPRi perturbations.”
Koo added, “In reality, CRISPRi is incredibly challenging to perform in the laboratory. And you’re limited by the number of perturbations and the scale. But since we’re doing all our perturbations [virtually], we can push the boundaries. And the scale of experiments that we performed is unprecedented—hundreds of thousands of perturbation experiments.”
Koo and his team tested CREME on another AI-powered DNN genome analysis tool called Enformer. They wanted to know how Enformer’s algorithm makes predictions about the genome. Koo says questions like that are central to his work.
“We have these big, powerful models,” Koo said. “They’re quite compelling at taking DNA sequences and predicting gene expression. But we don’t really have any good ways of trying to understand what these models are learning. Presumably, they’re making accurate predictions because they’ve learned a lot of the rules about gene regulation, but we don’t actually know what their predictions are based off of.”
With CREME, Koo’s team uncovered a series of genetic rules that Enformer learned while analyzing the genome. That insight may one day prove invaluable for drug discovery. The investigators stated, “CREME provides a powerful toolkit for translating the predictions of genomic DNNs into mechanistic insights of gene regulation … Applying CREME to Enformer, a state-of-the-art DNN, we identify cis-regulatory elements that enhance or silence gene expression and characterize their complex interactions.” Koo added, “Understanding the rules of gene regulation gives you more options for tuning gene expression levels in precise and predictable ways.”
With further fine-tuning, CREME may soon set geneticists on the path to discovering new therapeutic targets. Perhaps most impactfully, it may even give scientists who do not have access to a real laboratory the power to make these breakthroughs. “CREME provides a road map to improving perturbation experiments to better characterize cis-regulatory mechanisms,” the team concluded, noting that any insights gained via DNN interpretation should be “treated as hypotheses and validated by laboratory experiments.”
