From mouse to man: New machine-learning model reduces number of ‘wasted experiments’

07-Dec-2018 - Last updated on 10-Dec-2018 at 13:02 GMT

CytoReason's machine learning model translates the results of new mouse experiments into the equivalent human condition. (Image: CytoReason/Technion)

The new “groundbreaking” model can translate the results of mouse experiments into the equivalent human condition, outperforming traditional methods of extrapolation by up to 50%, say researchers.

Scientists from CytoReason and the Systems Immunology and Precision Medicine Lab at the Technion Faculty of Medicine developed the mouse to human model, which was recently described in Nature Methods.

The model – Found In Translation or FIT – is a machine-learning platform for human immune system cell-level simulation. It was tested on mouse models of 28 different human diseases.

According to the researchers, the platform “outperformed direct cross-species extrapolation from mouse results, increasing the overlap of differentially expressed genes by 20-50% in pre-identifiable disease conditions.” It also uncovered novel disease-associated genes.

Specifically, the researchers discovered a role for Interleukin Enhancer Binding Factor 3 (ILF3) in Inflammatory Bowel Disease (IBD). Neither IBD mouse models nor human datasets previously demonstrated the role of ILF3 in IBD. This discovery “bears testament to the exciting possibilities of this approach,” said CytoReason Chief Scientist Shai Shen-Orr.

“Drug developers need to be able to bridge the gap between preclinical mouse data and clinical outcomes – hence the need to develop methodologies to help do this,” Shen-Orr told us.

The platform works by computing a per-gene human effect size for each dataset, learning a gene-level statistical model of mouse-to-human relationships by modeling and re-sampling the data, before predicting the human effect size by computing the mean of the estimated effect sizes resulting from the re-sampling, according to the researchers.

“In short, what lies behind this model is the machine learning approach that powers the ability to translate outcomes from one thing to another – from cell to gene, from tissue to blood, from disease A to disease B, and now, from mouse to man,” Shen-Orr explained.

Left: Direct translation of mouse results to human biology is not explicitly informed of species differences. Right: The FIT model leverages matched mouse-human gene expression data from the public domain to inform the human-relevant interpretation of a new mouse experiment. By learning prior mouse-human similarities and differences, FIT reduces time spent on false leads and highlights masked signals that may otherwise be missed. (Image: CytoReason/Technion)

While the mouse to human model is still in its infancy, the machine-learning platform on which it is built is already commercialized via CytoReason’s Cell Centered Model. The company is collaborating with several pharmaceutical companies on a variety of programs.

Shen-Orr said, “This model demonstrates the power of our growing and rich data sets, feeding our proprietary machine learning technologies and unique methodologies, to more accurately understand context, and to generate insights and hypotheses that would have been a matter of serendipity before.”

He calls the new model a groundbreaking application, one which, in the future, will benefit various applications across the discovery and development continuum.

The model is not going to replace mouse experiments, but it will make them “more targeted, more accurate, more informative and thus more effective,” consequently reducing the number of “wasted experiments” and, potentially, the number of post-mice human trial failures, he added.

“Furthermore, it can uncover vital information – disease genes, signals, and pathways that you might not see with traditional extrapolation.”

While additional work is required at both an academic and commercial level, Shen-Orr said at its current stage the technology is already augmenting the company’s existing translational capabilities.

“Every new data set we feed into the machine learning infrastructure we have built takes us another step towards re-defining our understanding of the immune system at a system level,” he said. “And another step along the way to having the most powerful and accurate methodologies for discovering new targets, new pathways, and new uses for the drugs that make so much difference to the patients that need them.”

Successful drug development requires asking and answering the right questions at the right time, added Shen-Orr, noting the new model and the broader applications of the machine learning approach on which it is based, improves researcher’s ability to do this.

“No data exists in a vacuum and it really only comes to life when put into context,” he added. “To provide that context you need a learning environment that grows, enriches and becomes ever more accurate with every new dataset of almost any type.”