Daniela Kalafatovic
Associate Professor, University of Rijeka
AI/Machine Learning Worshop
Abstract
The discovery of active peptides, such as antimicrobial, antiviral, catalytic, or self-assembling is challenging, due to the vast search space and limited understanding of how peptide sequences correlate with desired activities and/or functions. The exponential growth of the peptide permutation space with increasing sequence length makes it difficult to explore all possibilities efficiently. To avoid expensive and time-consuming guesswork and experimental failure, we combine machine learning (ML)-based
predictions with genetic algorithm-based optimizations to accelerate peptide discovery. ML can find patterns or regularities in data, build mathematical models based on the theory of statistics and make up for the lack of knowledge.
We curated a dataset of experimentally validated self-assembling peptides and combined aggregation propensity data from molecular dynamics (MD) simulations with a sequential properties representation scheme to train a neural network classifier able to identify peptides with self-assembly propensity. The ML classifier achieved 81.9% accuracy, outperforming the current state-of-the-art models. Furthermore, we developed a flexible and adaptive generative model based on the ML-driven genetic algorithm that allows for a directed search of the sequence space by promoting self-assembly propensity. This enabled the discovery of sequences in unexplored regions of the peptide space, with low similarity to the dataset, which were validated using MD simulations and experiments [1]. This approach not only improves the efficiency of the search but also contributes new knowledge to expand existing peptide datasets, driving future advancements in the field.
[1] Njirjak, M., Žužić, L., Babić, M., Janković, P., Otović, E., Kalafatovic, D., Mauša, G. (2024) Nat. Mach. Intell., 6, 1487–1500.
Bio
Daniela Kalafatovic is Associate Professor at the University of Rijeka, where she was Head of Medicinal Chemistry Division from 2021 to 2024. She received her Ph.D. degree in chemistry from the University of Strathclyde in 2015. After the PhD, she was a Research Associate at the Advanced Science Research Centre, City University New York,
Nanoscience initiative. In 2016, she joined the Institute of Research in Biomedicine in Barcelona as a Marie-Curie cofund postdoctoral fellow. She began her independent career as Assistant Professor in 2019 at the University of Rijeka. She is the holder of a major national project for young researchers, the Croatian Science Foundation starting grant entitled “Design of short catalytic peptides and peptide assemblies” (DeShPet, UIP-2019-04-7999) and since 2024 she is Action Chair of the COST Action
“Searching for nanostructured or pore forming peptides for therapy” (CA23111) that has more than 330 participants from over 35 countries. She is also leading several University founded projects. She published more than 25 articles with the recent Nature Machine Intelligence worth mentioning.