Mouse-Human Dictionary

Despite the extensive use of mice in studying human disease, the translation of such studies to the human body is as yet limited, and in many cases, a drug that yielded good results in lab mice does not work in humans. Obtaining the ability to infer biological processes in humans from their murine parallel would diminish that gap, thereby leading to a significant breakthrough in the development of new drugs and in understanding biological processes. A group of scientists, led by Prof. Shai Shen-Orr from the Faculty of Medicine at the Technion, in collaboration with scientists from Stanford University, have recently published a study in Nature Methods that provides a significant advancement towards developing a biological version of a "mouse-human dictionary".

Cells of all living creatures, including mice and humans, continuously produce and break down proteins according to demand. Proteins are large molecules that govern nearly all bodily processes: moving muscles, sending signals from cell to cell, cell division, and even the production of additional proteins – nothing can be done without them. Due to their great importance, a comparison between proteins found in the cell under normal conditions can and those found in it during a disease can inform us about the disease's mechanism of action and what the possible therapeutic approaches may be.

There are many genetic differences between humans and mice, and therefore, even if both would fall ill with the same disease, it would have a different effect on the abundance of different proteins in their cells. Therefore, there is a need to "translate" between the two, for instance, by telling us that elevated levels of protein A in a sick mouse correspond with a situation in which proteins B and C are elevated in humans with the same disease.

Found IN Translation

To address the difficult task of inferring from mouse experiments into humans, the researchers developed a statistical, machine-learning-based model, termed Found IN Translation. The field of machine learning involves the development of algorithms (rules for processing information) that allow the computer to make decisions and predictions. In order for an algorithm to make decisions, it must be "trained", namely, supplied with numerous examples and instructed what they mean, so that the next time it encounters a similar piece of information, it would know what to do with it. For instance, for face-recognition learning algorithms to work, we will first present them with a large number of pictures and classify them according to whether they contain a face or not. After this type of training, the computer would know how to assess, with high certainty, if a new picture contains a face or not.

The current study involved training the algorithm using a huge database that contained thousands of pieces of genetic information regarding healthy and ill mice and humans. After the training, the researchers presented the algorithm with information on the amounts of proteins in the cells of a mouse ill with a disease the algorithm has not encountered before, and it assessed the identity of the proteins expected to be relevant to the same disease in humans, along with the level of certainty of this assessment.

The researchers demonstrated the algorithm's power using one of the proteins they found: it indicated that this protein plays a role in inflammatory bowel diseases, though its connection to these diseases has never been demonstrated before. Testing tissue samples from patients has, indeed, shown significantly elevated levels of this protein in comparison to healthy individuals.

According to the researchers, their method is 20-50% better at identifying proteins involved in human diseases than direct inference from mouse experiments to humans. So we still have not found the "holy grail" that would allow us to fully decipher all the connections between laboratory experiments and patients, but this is an important step in the right direction. In addition, since machine-learning mechanisms are capable of training and improving, we can expect that the more information it is fed, the results it would yield would improve.

Translated by Elee Shimshoni