A few colleagues at Roche are looking for a PhD-level intern to investigate the potential to combine knowledge graphs with patient-level data for prediction tasks. See the link at the bottom to apply for it.

Knowledge graph for biology

Knowledge graphs represent our knowledge of biological systems in networks, i.e. nodes (biological entities such as DNA, RNA, protein, bacteria, virus, or abstract knowledge entities such as disease, life style, etc.) and edges (how different entities relate to and modulate each other). These networks are notoriously complex, incomplete, and noisy, since our understanding of biology is probably still naive, often uncertain, and not rarely completely wrong. However, people try to do their best to collect their knowledge in these networks and to improve on them, because many people hold the assumption that by factoring in the network, we may better understand our observations of and data collected from biological systems, for instance human disease. Many researchers consider the success story of repurposing Baricitinib against SARS-Cov-2 as an success story, though the proposed MoA, namely AAK1 inhibition instead of the primary MoA of Baricitinib, JAK inhibition, needs to be verified. This is a crucial point, because any biologist may have hypothesized that an JAK inhibitor may be helpful for an over-activated immune system insulted by the virus. But coming to adaptor-associated kinase 1 (AAK1) would have required some deep insight into both human immune system and viral biology, and some luck.

For any given biological question, it is of general interest to learn whether using the network improves our capability of addressing and solving the question compared with the scenario where biological knowledge is used heuristically at best.

Representation learning and graph embedding

Representation learning on graphs tries to find optimal ways to represent, or encode, the graph structure so that it can be used by machine learning models1. Like language models which embed words in high-dimensional spaces, graph representation learning aims at deriving features from the graph structure so that they represent latent organization principles of the graph and are subject to machine learning models.

Traditionally, researchers relied on user-defined heuristics to extract features, for instance degree statistics or kernel methods. Alternatively, we can exploit matrix factorization and stochasticity such as random walk to address the problem. When data amount is huge, nonlinear dimensionality reduction with neural-network-like approaches such as graph convolutional networks and variational autoencoders can work well, though we often have to sacrifice the interpretability of the model. Again, for any given task, it is of interest to learn the optimal strategy (or strategies) by considering both its performance, and perhaps more importantly, the complexity and interpretability of the model.

While some machine-learning researchers speculate that deep graph networks may give us the power of relational inductive bias, which may constitute the combinatorial generalization capacity of human intelligence, other researchers are more cautious about them, and argue for using interpretable models instead, especially for decisions at high stake2.

About the internship

My colleagues are interested in using graph networks and patient-level data for prediction tasks such as label expansion, prediction of patient outcome, and inference and prediction of efficacy and safety profiles of drug combinations.

I think such tasks may be perfect cases to test whether graph learning exceeds various baseline benchmark results, which can derive from (1) data-only interpretable methods, (2) data-only black-box methods, and (3) interpretable methods coupled with heuristic biological knowledge, and (4) back-box methods coupled with heuristic biological knowledge. While comparing all these strategies presents a daunting task, the intern will hopefully make the first move, helping us better understand the strengths and limitations of applying graph-learning approaches for questions in personalized healthcare.

If you are interested, please find more details and apply here. And welcome to spread the words for my colleagues.

Notes

  1. See Representation Learning on Graphs: Methods and Applications by Hamilton, Ying, andmark Leskovec for a review of the techniques. ↩

  2. See Rudin, Cynthia. “Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead.” Nature Machine Intelligence 1, no. 5 (May 2019): 206–15. https://doi.org/10.1038/s42256-019-0048-x. ↩