A machine learning model for mouse-to-human inference

A machine learning model for mouse-to-human inference

CytoReason, developer of the world's first machine learning platform for human immune system cell-level simulation, today announces publication of a groundbreaking new model for translating data from mouse models to human disease. Described in Nature Methods, the mouse to human model (Found In Translation or FIT), proved its ability to more accurately and effectively extrapolate results from the mouse-based research that is a vital, and necessary, part of every drug discovery and development program.

Much of what we know about disease is rooted in research done using mouse models (mice bred specifically to have the characteristics of the disease being studied). Furthermore, every single new drug will have first had to demonstrate some level of safety and efficacy on mice. But mice are not humans, and cross-species differences have consistently been a major stumbling block to translating lab-based research into something that will be meaningful for patients and clinicians.

Until now, knowledge of species differences has not been systematically incorporated into the interpretation of animal models. In trying to overcome this huge issue, scientists from CytoReason and the Systems Immunology and Precision Medicine Lab at the Technion Faculty of Medicine, developed their mouse to human model. This new model builds on CytoReason's mission of driving biological insights that transform drug discovery and development, lifting analysis of datasets out of an isolated vacuum and applying the full context of existing knowledge, in a similar way to CytoReason's Cell-Centered Models of the immune system do in terms of identifying and understanding gene/cell/cytokine relationships.

Tested on mouse models of 28 different human diseases, the mouse to human model outperformed direct cross-species extrapolation from mouse results, increasing the overlap of differentially expressed genes by 20-50% in pre-identifiable disease conditions. It uncovered novel disease-associated genes, highlighted signals that may otherwise have been missed and reduced false leads, with no experimental cost.

"This is a massive advance. Mice models are a necessary but flawed method of trying to understand what might happen in a human in any given situation,' said the Chief Scientist at CytoReason. "We have shown that we can significantly increase the accuracy of what we learn from these models. This changes the entire dynamic in terms of confidence in making the right decisions for the next steps in a given drug development program - doing mouse-kind and mankind a major service in the fight against disease."

"The mouse to human model clearly demonstrated its ability to uncover novel disease-associated genes. It predicted a role for Interleukin Enhancer Binding Factor 3 (ILF3) in the colon of Inflammatory Bowel Disease (IBD) patients compared to healthy people, even though ILF3 was not seen in either IBD mouse model or human datasets," said the lead author. "We also didn't see any ILF3 associated to IBD in any past research. We did, however, see a significant increase in ILF3 in the colons of IBD patients versus healthy patients in laboratory tests, validating this as a real and novel finding of great significance."

At the heart of the process is the pairing of human and mouse model datasets with a human-disease dataset of comparable conditions to produce cross-species pairings. For each cross-species pairing in each species the difference was calculated between disease and control samples, which was then used to study how different human and mouse genes express under similar conditions and fed into the mouse to human model.

To translate mouse model data to human model data, the mouse to human model follows three steps:

  1. It computes a per-gene human effect size for each dataset

  2. Learns a gene-level statistical model of mouse-to-human relationships by modelling and re-sampling the data (bootstrapping)

  3. Predicts human effect size by computing the mean of the estimated effect sizes resulting from the re-sampling

Genes with high absolute effect-size predictions are more likely to be associated with the human condition of interest.

"This is a real breakthrough. Many drugs that appear to be effective in mice go on to fail in clinical development. This technology, part of our growing portfolio of translation capabilities, will help bridge the gap between pre-clinical results and clinical outcomes", said CytoReason's CEO. "It is a demonstration of the power of our growing and rich data sets, feeding our proprietary machine learning technologies and unique methodologies, to more accurately understand context. This enables the transference of understanding from one element of research to another, in order to improve overall drug development and clinical outcomes."