In 2014, together with our clients, we published algorithms for automated selection and assessment of classification and regression models in a paper in the Journal of Cheminformatics. Our main research goal has always been to find the best way to select and assess these models.

In medical diagnostics our constant research goal has been to find the best way to select gene markers. In 2014, together with our clients, we published our algorithms for automated feature selection in a paper in the Journal of Cheminformatics.

In certain scenarios, possessing an AI expert system capable of making predictions while also acknowledging uncertainty with a 'don't know' response could prove immensely beneficial. Our research endeavours to address two key inquiries: Firstly, how can we effectively delineate 'don't know' predictions in real-world applications? Secondly, what metrics can we employ to gauge the efficacy of such systems in terms of success or failure?

In 2020 Damjan Krstajic published a chapter Non-applicability Domain. The Benefits of Defining “I Don't Know” in Artificial Intelligence in the book Artificial Intelligence in Drug Discovery. Furthermore, in our spare time we are developing aloom R package..

While working with numerous right-censored (survival) datasets, we have observed their fragility and noted how heavily our estimates rely on specific data points and assumptions. Our research seeks to address the following questions: How can we develop optimal predictive models based on survival datasets? How should we effectively report such models? What are the essential criteria for a survival prediction model to be considered suitable for clinical practice?