Olga Kalinina

Biomedical research presents specific challenges for data science, and in particular for applications of artificial intelligence. Whereas the world has witnessed tremendous success of AI in generative applications, where large and black box models can produce unique content or assist the user in many other ways, biomedical applications require models that are reliable, robust, and explainable. Moreover, in these applications, the training datasets are limited in scale and often quality. From the bioinformatics perspective, this presents very specific challenges in how one should approach developing AI models. One particular important example are models for predicting effect of genetic variants, which have been of paramount importance in the field for over 20 years.

In my talk, I will address some specific challenges arising when applying AI in these scenarios. One challenge is related to data leakage, an illicit spill-over of information during the models’ training process that undermines their robustness and generalizability to unseen data. I will present a novel training scheme developed in my lab that allows to address and minimize data leakage. Towards variant effect prediction, I will present a novel and fully explainable approach to predicting functional effect of genetic variants that is free from data leakage and goes beyond the state of the art. I will demonstrate how features derived from analysis of protein three-dimensional structure significantly contribute to the success of the prediction.