Goran Mausa
Associate professor, University of Rijeka, Faculty of Engineering
AI/Machine Learning Workshop
Abstract
Machine learning is changing the fundamental concepts of discovery in all scientific disciplines, and peptide chemistry is no exception. From classifiers that predict peptide function, to generative models that try to navigate a vast search space of peptide sequences, machine learning is very successful in knowledge discovery from experimentally verified data. In this process, intelligent algorithms are not merely competing with human experts, but they are mainly used to complement our understanding of complex phenomena.
The lack of transparency of many machine learning models and the complex mathematical principles that govern them are the reason why some researchers still do not completely accept their use. Although computer scientists made an effort to allow for a rather simple use of machine learning models, it is essential to understand the importance of data preparation and pre-processing, and the basic concepts of training, validating and testing of these models.
This talk aims at breaking common misconceptions which hinder the application of machine learning to peptide chemistry, explaining its capabilities and limitations and providing examples of good practices. The following challenges will be addressed: (i) the reliability and ability to generalize knowledge, as a paramount concern of any prediction model, (ii) the means of transforming the data into machine-readable format and selecting the essential features, (iii) the amount of data needed to properly train a machine learning model and (iv) the generative capabilities of neural network-based models that overcome the intractability of exhaustive search.
Bio
Goran MauĊĦa is a computer scientist and an associate professor at the Faculty of Engineering, University of Rijeka in Croatia and head of Computer Science department. His research is focused on developing reliable and accurate prediction models based on machine learning, algorithmic solutions for optimization and generative AI models with application primarily in the field of peptides, but also environmental engineering, software engineering and robotics. His work about the development of a generative AI model for reshaping the discovery of self-assembling peptides was recently published in Nature Machine Intelligence. Goran participated in 20 research projects, supervised over 50 defended master and bachelor theses and 1 PhD thesis, and received the national annual award for research excellence in 2022.