MoSBi: Algorithm identifies disease subtypes

Doctors have always used symptoms, imaging, and laboratory data to define and diagnose diseases, but at times it is simply not enough: while patients may have the same illness, it may exhibit different changes at the molecular level. A team from the Technical University of Munich has developed the so-called MoSBi algorithm and makes it available to researchers to identify molecular differences.

In this MEDICA-tradefair.com interview, Dr. Josch Konstantin Pauling and Tim Rose talk about stratified medicine and the MoSBi algorithm and explain how researchers can use the web-based tool to benefit their own work.

Dr. Pauling, Mr. Rose, what is stratified medicine?

Dr. Josch Konstantin Pauling & Tim Rose: Stratification in medicine means breaking down a disease into different subtypes based on various causes or processes at the molecular level. While these can involve the same symptoms, they may entail different risks or require other treatment options.

What are examples of diseases with subtypes?

Pauling & Rose: Cancer is perhaps the most notable example of this. A multitude of mutations can lead to changes in tissue, which can develop into a tumor. Tumor growth and possible metastasis formation usually result in additional molecular changes. Doctors examine the specific changes in the tumor and adjust the patient’s treatment accordingly.

How does the MoSBi approach work?

Pauling & Rose: MoSBi (Molecular Signatures using Biclustering) is a bioinformatics method for the analysis of large-scale molecular data sets, as it pertains to clinical research, for example. It belongs to the class of unsupervised machine learning algorithms and automatically looks for patterns in the data. In our case, it identifies patient groups with similar molecular signatures. The targeted identification of characteristic molecular signatures that best describe the similarities between patients is a special feature in this setting. MoSBi is a so-called “ensemble” approach. This means the predictions from different algorithms are combined to produce more robust results. We also developed a scalable network-based visualization that illustrates complex relationships and connections and supports interpretation.

Having said that, it is important to underscore that this is an algorithm used in research. The predictions must be interpreted with caution and can serve as a reference point for hypotheses and further research. Stratification in clinical practice requires far greater expertise and validation.

How can researchers use the MoSBi web tool to their benefit? What added value does this create?

Pauling & Rose: Bioinformatics algorithms are often only made available as libraries for specific programming languages – such as R or Python. While this facilitates further use by other researchers, it requires them to have good programming skills, and a detailed understanding of data analysis. We developed the web-based tool to enable all researchers with the respective data and science research question to use MoSBi independently without the need to collaborate with experts in bioinformatics. The web tool facilitates all functions of the MoSBi algorithm, subsequently enabling the visualization of the results.

Clinical research frequently involves highly sensitive data. The MoSBi web-based tool is an open-source software that can be run on any computer, meaning users do not have to use the web tool that runs on our institute’s servers.

What prompted the idea for the tool?

Pauling & Rose: The idea came up while we were involved in a study that required us to understand the heterogeneity of the patient cohort. We focused on a group of algorithms that could technically accomplish this, but we did not know which algorithm we should use and how we had to fine-tune the respective initial parameters. That is why we developed the MoSBi approach, which can automatically combine the results of multiple algorithms. The visualization aspect was needed because we required an intuitive view of the results to make them easier to interpret.

What exactly did you examine in the study pertaining to the MoSBi approach? What was your conclusion?

Pauling & Rose: The MoSBi publication focuses on bioinformatics. Bioinformatics method development requires thorough evaluation. We subsequently used public, molecular data from published studies and simulated data to address certain data characteristics. It allowed us to assess the strengths and weaknesses of MoSBi compared to other methods and to develop application scenarios. Of course, the method is also described in detail and defined mathematically in the study.

In another publication on nonalcoholic fatty liver disease, we used MoSBi to identify subsets based on lipid signatures for disease progression. We were able to show that the method can also be used in practical settings to deliver important insights.

What is the potential of algorithms or machine learning in medicine in the future?

Pauling & Rose: Machine learning can facilitate more accurate diagnoses or treatment options by combining and analyzing vast amounts of data. One can also systematically include the patient’s medical history, for example. It is primarily intended to be a clinical decision support tool for doctors. The legal framework outlining the digitization of patient records is obviously a key aspect in all this, enabling standardization in their use and contribution to research, if patients are on board.