(595a) Invited Talk: Interpretable and Explainable Support Vector Machines: Applications in Cancer Therapeutic Design and Immunoengineering

There is serious concern that the continued increase in drug development costs observed in recent decades is unsustainable and will lead to further increases in already exorbitant drug prices, highlighting the need for a new paradigm for therapeutic discovery. To meet these challenges, our group develops tools for therapeutic design to accelerate discovery and enable personalized medicine. Our work integrates molecular and systems modeling, machine learning, and global optimization to develop multi-scale models of biomolecular systems, and then leverages these models for the design of novel therapeutics. Though recent advances in deep learning have transformed the field of bioinformatics, there is still great need for explainable data-driven models that can connect biomolecular descriptors to biomolecular function, and ultimately to health and disease. For instance, explainable machine learning models can be used for understanding the physicochemical driving forces of biomolecular function, to identify novel biomarkers, and serve as guides for the design of novel therapeutics. Recent work in our group has focused on developing methods for training and analyzing explainable or interpretable support vector machines, including novel approaches for feature selection. Explainable machine learning can also be used to “attribute” model predictions to specific characteristics or features of a particular sample, and recent work within our group has also been aimed at developing methods for attribution for support vector machines. These methods for interpretable and explainable machine learning have broad potential applications, and this talk will cover examples from therapeutic design and personalized medicine.