Donnelly Centre PhD student wins 2023 Borealis AI Fellowship

A PhD student at the Donnelly Centre for Cellular and Biomolecular Research, Anastasia Razdaibiedina, has been awarded the prestigious Borealis AI Fellowship for 2023-2024.

Borealis AI is a research institute at the Royal Bank of Canada working to forge the future of AI for finance. The institute offers one-year fellowships of $10,000 to graduate students conducting research in machine learning or artificial intelligence at Canadian universities. The aim of the Borealis AI Fellowships program is to foster a diverse next generation of AI researchers in Canada.

Razdaibiedina is one of 10 researchers to win the 2023 Borealis AI Fellowship – and one of two from the University of Toronto. Razdaibiedina is co-advised by Brenda Andrews, university professor of molecular genetics, and Charles Boone, professor of molecular genetics. Jimmy Ba, assistant professor of computer science and faculty member at the Vector Institute, is a key collaborator on her thesis work.

Razdaibiedina is currently a fifth-year PhD student in the Computational Biology track. Her research spans representation learning and foundation models, including the application of these areas to biological imaging datasets. She anticipates graduating in 2024, with the Borealis AI Fellowship helping her to reach the finish line.

Razdaibiedina developed an interest in computational methods and biology while pursuing her Bachelor’s degree in Applied Mathematics at the Taras Shevchenko National University of Kyiv in Ukraine. She was drawn to U of T by its international reputation – particularly for its graduate programs focusing on AI and biology.

The fellowship will support Razdaibiedina’s research to build a generalist method for predicting protein function in human cells from fluorescent microscopy data. What sets this method apart from others currently in use is that it is built on a generalist visual transformer model and can characterize proteins from a wide range of image and cell types without model retraining or fine-tuned adjustments.

The project will build on previous research conducted by Razdaibiedina to discover functions of unknown proteins using deep learning and to improve the multi-task performance of deep learning models.

Earlier in her PhD, Razdaibiedina developed a deep convolutional neural network, called PIFiA, that learns to group proteins with similar biological roles – without requiring the input of human-created labels during training. She also developed an architecture for multi-task and continual learning that will allow the new model to be trained on millions of images.

Razdaibiedina’s goal is to make her deep learning method publicly available by next year through an open-source Python library and an interactive website. The open-source library could help biomedical researchers make strides in improving medicine by elucidating how proteins respond to drug treatments in preclinical trials.