week1 AI bridges gap towards better medicine

I'm Chengqi Xu, a graduate student at Elemento Lab, which resides in the Englander Institute of Precision Medicine. Since my clinical mentor Dr. Manish Shah went to ASCO meeting, this week I was mainly involved in participating in an intercampus meeting hosted at the Belfer building. Immersed in the two-day talks, I am inspired by lots of excellent works that deploy big data analytics and machine learning techniques to tackle hard biology questions ranging from processing h&e image staining, and drug repurposing, to inequality in pain. There are lots of exciting topics I would recap in this blog.

Dr. Fei Wang talked about the future of health informatics using federated learning, which is a generalised learning strategy that can be applied to any machine learning model. In their recent work, they collaborated with Mont Sinai scientists, to predict the mortality rate of covid-19 in five different hospitals at Mont Sinai. Federated learning is featured for it can securely store information locally to protect privacy, where it updates the parameters to the centralised model. A new decentralised model allows avoiding of communication between local nodes and the central hub. He also mentioned the previous release of the integrative biomedical knowledge graph, which integrates data from different publicly available data sources. The efforts to build such a curated and standardised hub are very useful for biomedical researchers to train/test new machine learning models.

Dr. Emma Pierson gave us a fascinating talk on using an AI approach to discriminate knee pain in the underserved populations sourced from radiographs. They showed the algorithm’s ability to reduce unexplained disparities stems from in the racial and socioeconomic diversity of the training set. In her another recent study, the Leskvoc team showed the mobility can have a real impact on infection rates. They have constructed an undirected bipartite graph to model mobility network, which can infer joint distribution from CBG (people) to POI (places) at time t. The computational prediction of the disadvantaged group who are not being able to reduce their mobility helps the policy maker aware of the disparities, and makes a more effective decision.

Dr. Olivier Elemento highlighted the topic that can we use AI to best combine drugs. His lab has being contributed to the drug community throughout years' efforts toward understanding the complexity behind anticancer therapeutic strategies. I really enjoy reading the paper he mentioned in his talk, such as Cory's preprint on finding synergy drugs that might share more chemical similarities, and also Wei's paper which established a kinetic  BCR signalling framework in DLBCL, especially the highly detailed way to directly model the PPI using ODEs. He also mentioned my recent work with Dr Heng Pan to create a comprehensive knowledge-graph-based model package to predict synergy. I hope I could learn and get involved in the automated high-throughput drug screening pipeline at Sandra and Edward Meyer Cancer Center, to further test the in silico predictions in the in-vitro experimental setting.