Learning Representations on Biological Data with Weakly Supervised Learning

Explore a comprehensive lecture on learning representations of biological data using weakly supervised learning techniques. Delve into high-throughput perturbational experiments and their role in identifying biological relationships, characterizing mechanisms of action, and constructing biological networks. Examine the impact of dataset transformations on cosine similarity distributions and discover Perturbational Metric Learning (PeML), a novel weakly-supervised method leveraging replicate data to enhance representation learning. Investigate how PeML improves the recovery of known biological relationships and enables more effective downstream analysis tasks. Gain insights into the exciting frontier of representation learning with weak supervision and self-supervision in computational biology through this in-depth presentation by Ian Smith from Valence Labs.

- Intro
- Background
- Similarity analysis
- Structure of biological data
- Cosine similarity
- Results
- Perturbational metric learning
- Contrastive learning
- PeML
- Results
- Future directions
- Conclusions
- Q+A