Our engine for biomarker discovery platform is a "Cooperative Co-evolutionary fuzzy modelling algorithm" (Fuzzy CoCo). Fuzzy CoCo combines two techniques: Fuzzy Logic and Artificial Evolution.
Fuzzy logic transforms complex models into simple, accurate and human-interpretable rules. Rules such as "if gene A is up regulated and gene B is down regulated, then the probability of disease X is high" are less ambiguous.
Artificial evolution creates populations of candidate models by mimicking natural evolution. An evolutionary algorithm selects the best models (for example, those that better classify "healthy" vs "non-healthy") and the worst models "die". The fittest models survive and blend with the others survivors to generate better models in the next generation.
Candidate models are further processed with our patented filtering algorithm (WO2017199067) and Tolerance Assesment module. This validates real-world applicability of the models and tolerance to data perturbance/corruption Figure below shows evolution of all models and illustrates an example of BOSS's automated model selection process
We provide clients with models (including multiple biomarkers) that best discriminate their endpoint for future prognosis. Results can be exported in any format.
Clients can also use our SaaS product, BOSS Explorer, to explore the results and identify models (containing interesting biomarkers) based on their own selection criteria. We also provide training for this. Figure below shows 4 candidate models selected by a user and their corresponding performance metrics.
The BOSS platform addresses real world applicability of models at several stages: - During artificial evolution, preference is given to models that generalize better. - Thousands of models are generated on bootstrapped data; thus model evaluation happens on both seen and unseen data. - Multi-objective selection employs both standard metrics and internal metrics. - The original data is perturbed in different ways (i.e. artificially, but realistically, "corrupted") to identify the "break point" of a model.