Bartender model predicts how drugs affect each person

“Our model integrates many very different data for the first time. One can characterise the model as a kind of bartender, because it can mix data ingredients into a cocktail we can learn from as a whole and not just the ingredients separately.”

Professor Søren Brunak explains their new study, where he and colleagues have published a new model to combine many different health data.

By combining health data, the model provides unique insight into how a person would respond biologically to treatment with a specific drug or whether dietary changes or changes in exercise habits would have a greater effect on biomarkers of disease and health.

An oncologist may consider four or five anticancer treatments. The oncologist can then test all the drugs digitally on a person and determine how each drug would affect relevant biomarkers for cancer before deciding on treatment. Again, provided that the many different data types have been generated.

In addition to providing insight into the effect of a drug, the model can also predict the effect of drug–drug combinations.

The model combines not only genomic data, metabolic data or questionnaire surveys but are increasingly trying to integrate different deep data for each person. Within biopharmaceutical research, we have increasingly moved away from using just one data type at a time, explains Søren Brunak, Head of Research, Novo Nordisk Foundation Center for Protein Research, University of Copenhagen.

“Numerous processes are linked in a living being, and determining how a person’s genes would affect metabolites if the person has a certain type of lifestyle is important,” he says.

The model is based on a deep learning framework that integrates data from almost 800 people with type 2 diabetes and contains 12 types of data related to genomics, lipidomics, metabolomics, proteomics, maps of the gut microbiome, questionnaire surveys on diet, exercise, smoking habits and medication use and many clinical data types: for example, on blood glucose, blood insulin and organ scans.

The research has been published in Nature Biotechnology.

Based on people with type 2 diabetes

The deep-learning model is based on data from people with newly diagnosed type 2 diabetes. These data originated from the public–private partnership DIRECT, which has been operating for more than 12 years, undertaking one of the most comprehensive mappings of disease biology that analysed their lipid and metabolite profiles, genomes, gut microbiomes etc. in great depth.

Based on a previously developed deep-learning framework, the model collected data and learned from them, so that it can now suppress redundant data for each person and predict, for example, how a certain drug would biologically affect any other person.

“The model initially learned to understand data, and then we developed a framework so that we can ask questions. The new aspect is that the generative model integrates data into one big cocktail, and we can then ask questions related to the data and get answers that are otherwise impossible to obtain. For example, we can ask what would happen if you gave a person a drug they have not taken before. We can determine how this would affect the proteins, lipids or metabolites we may be interested in increasing or reducing,” explains another leading researcher behind the study, Simon Rasmussen, Associate Professor and Group Leader from the Novo Nordisk Foundation Center for Protein Research.

Advancing knowledge on the effects of drugs

Søren Brunak says that the model cannot yet be used clinically because this requires large quantities of expensive omics data on each person. But since revealing patient-level disease characteristics is likely to become more cost-effective, the effects of medication in a clinical setting can be predicted, thereby identifying what would be most effective for a specific person in a specific situation and the effects of alternative drugs at the molecular biological level.

This does not just apply to people with diabetes but also to people with other diseases, including cancer.

For example, an oncologist may consider four or five anticancer treatments. The oncologist can then test all the drugs digitally on a person and determine how each drug would affect relevant biomarkers for cancer before deciding on treatment. Again, provided that the many different data types have been generated.

Another potential is within research, with researchers from both academia and industry determining much more easily how a drug would affect people with certain biological characteristics at the molecular level.

“The model can be used to understand how an individual drug works or to indicate biomarkers for disease or drug effects or new targets for treatment,” says Søren Brunak.

Functions like a person’s brain

Simon Rasmussen says that the model will initially be used to show that unsupervised deep learning can be used to integrate multimodal data.

He compares it to the human brain, which can also integrate visual impressions, sound, taste, smell and other sensory input.

The brain integrates all the data into one cocktail and then extracts the information needed. The model does the same type of data integration.

The quantity of data in the data sets that comprise the basis of the model is so huge and complex that traditional statistical models do not work very well, and deep learning is needed.

“People cannot create sensible statistics from 10,000 input parameters, but a deep neural network can. The model can thereby also assist doctors, who cannot possibly assess all conceivable combinations of lifestyle, metabolite profiles and drug treatment to recommend what the most effective drug would be for each person,” explains Simon Rasmussen.

Providing information on more than just drugs

The model is configured such that the researchers can activate exactly the functions they need.

Instead of asking the model about the effects of a given drug, researchers can also ask how a person starting to exercise, eating more vegetables or stopping smoking would affect all other parameters.

Tobacco smoke negatively affects many people’s health, although a few smokers live for more than 100 years.

“We live in the omics age, in which our differences can be described in great detail. If we were all identical and responded identically to the same diet or the same medicine it would be much easier, but we are not. People react differently, and this model can predict how individual people would biologically react to changes in various parameters,” concludes Søren Brunak.

Read the full study here.