Machine learning competition seeks best models for mapping human proteins
Biomedical researchers at KTH Royal Institute of Technology have opened up an online machine learning competition that will award USD 37,000 to be split among the creators of the best algorithms for classifying protein expression in images of human cells.
The competition will run for about 100 days
Images visualizing proteins in cells are commonly used for biomedical research, and these cells could hold the keys for the next breakthrough in medicine. However, thanks to advances in high-throughput microscopy, these images are generated at a far greater pace than what can be manually evaluated.
“There’s a great need for an enhanced automation of biomedical image analysis to accelerate the understanding of human cells and disease,” says KTH researcher Emma Lundberg, who directs the Cell Atlas project, part of the Human Protein Atlas at the Science for Life joint research center in Stockholm.
The Human Protein Atlas Image Challenge (https://www.kaggle.com/c/human-protein-atlas-image-classification/) calls on participants to train machine learning models to classify the patterns of protein expression within images of human cells. Creators of the best algorithms will split $37,000 in cash provided by Leica Microsystems, and an NVIDIA Quadro GV100 GPU.
Proteins are “the doers” in the human cell, executing many functions that together enable life, Lundberg says. Recent studies have challenged the classical cell biology view of one protein-one location. Rather, they suggest that half of all human proteins are localized to multiple cellular compartments, including many key drug target proteins.
Historically, classification efforts have been limited to single patterns in one or a few cell types, but in order to fully understand the complexity of the human cell, models are needed that can classify mixed patterns across a range of different human cells.
In a recent publication, Lundberg and the Cell Atlas team demonstrated the promise of both citizen science and artificial intelligence in describing the location of human proteins in images, however current results have yet to approach expert-level annotations (Sullivan et al, Nature Biotechnology, Oct 2018).
In this competition, Kagglers will push this idea further to develop models capable of classifying mixed patterns of proteins in microscope images. The Human Protein Atlas will use these models to accurately characterize a protein's location – or locations – from an image in high-throughput, and integrate this with their smart-microscopy system.
“I look forward to seeing the innovative and diverse approaches the community brings to solving the problem of protein localization from images,” Devin Sullivan says.
This approach has the potential to improve studies of protein and cell function, and increase our ability to understand human biology and disease, Lundberg says.
“We’re excited to get the help of Kagglers to develop a model that outperforms experts,” she says.