Panel on Interpretability in the Biological Sciences
summary
Genome-scale work treats a three billion base input where true positive binding sites number 2,000–20,000 positive sites per factor, yet neural nets trained on in vivo profiles correlate with in vitro binding energies after marginalization; calibrating with several thousand synthetic sequences lets models generate millions of in silico examples. Stability improves by predicting single-base resolution profiles instead of binary labels and by modeling assay biases as an extra modality; plans move toward chromosome-scale models and CRISPR plus population variation to learn cis–trans interactions.
AI-written summary · about
chapters (12)
- 0:00Interpretability challenges
- 2:06Data and model constraints
- 4:02Bridging physics and learning
- 8:00Neuroscience vs. genetics
- 10:33Experimental design and causality
- 13:41Data generation and calibration
- 17:27Representation discovery
- 21:14Foundational models
- 24:28Scientific reporting bias
- 29:02Robustness and bias in models
- 32:58Scaling and complex diseases
- 37:06Future methodology needs
related talks
36mFormatting Biological Big Data to Enable (Personalized) Systems Pharmacology
01h 13mIntroduction to Regulatory Genomics and Epigenomics I: Intro to the Biology of Gene Regulation
01h 07mIntroduction to Biological Network Analysis II: Protein-Protein Interaction Networks: From Graphs to
01h 19mAnalyzing Genetic Networks Using Neural Networks
10mGenesis Final Round: Nanyang Biologics
08mQ n A with Dr Denis Bauer
39mThe Translation Gap
28mSome Current Computational Challenges in Bioinformatics