Algorithms for Inferring Phenotypes from Ancient DNA

on YouTube ↗43m

" summary

A deep learning pipeline predicts 3D Hi-C contact maps from 1 Mb of DNA and applies Akita deep learning model to Neanderthal, Denisovan and diverse modern genomes, then compares maps to find 167 divergent loci where archaic and modern folding differ. In silico mutagenesis highlights drivers—often CTCF motif disruptions—and 24% of divergent loci appear introgressed into some modern Eurasians; predictions generalize to chimp with a ~10–20% accuracy drop and experimental validation is underway.

AI-written summary · about

" chapters (22)

1. 0:00Intro
2. 1:08How does genetic variation influence phenotypic variation?
3. 2:02Ancient phenotypes are hard to obtain
4. 3:19Migration, replacement, and admixture were common
5. 5:13Reveals sequence of ancestors and extinct relatives
6. 6:12How do modern humans differ from Neanderthals and Denisovans?
7. 7:11Polygenic risk scores (PRS) predict phenotypes from GWAS
8. 8:11PRS accuracy is low for many traits
9. 10:21Machine learning is enabling prediction of molecular phenotypes from DNA sequence
10. 11:17Genome folding influences gene regulation
11. 13:13Deep learning models can accurately predict 3D genome folding from sequence
12. 14:12Gene regulatory changes drive human evolution
13. 15:24Leverage algorithms for predicting 3D contact maps from sequence
14. 16:02Combining old genomes with new methods
15. 19:18Akita remains accurate across species
16. 24:20Humans and Neanderthals are more 3D similar than expected from sequence divergence
17. 24:38Are there Neanderthal-specific 3D contact patterns?
18. 25:45167 genomic windows with human-archaic 3D divergence
19. 26:30Reconstructing the 3D genome folding of archaic hominins
20. 27:06In silico mutagenesis identifies a CTCF motif disruption that contributes to the 3D divergence
21. 34:57The human family tree is complicated
22. 35:34Did Eurasians inherit Neanderthal 3D folding patterns?

" related talks

• 38mInterpreting Deep Learning Models Of Functional Genomics Data To Decode Regulatory Sequence...Anshul Kundaje · Interpretable Machine Learning in Natural and Social Sciences (2022)
• 57mBig Data and New Models Needed to Study DNA Variation in Evolution and CancerDavid Haussler · Symposium on Visions of the Theory of Computing 2013
• 05mDecoding the Human Genome with Deep LearningOlga Troyanskaya · O'Reilly Artificial Intelligence Conference 2019 - New York, New York
• 37mDynamics of 3D Enhancer-Gene Associations Across Diverse Human Cell Types and TissuesAnshul Kundaje · Network Biology 2016
• 30m23, Python and Me: Using Machine Learning in Python to Analyze Consumer Genomics DataNathan Brouwer · PyOhio 2024
• 12mHow data science can make your genome actionableRiley Ennis · Collision 2017
• 31mHacking the Human Genome Using Clojure and Similarity SearchArnoldo Jose Muller-Molina · Clojure/conj 2011
• 43mMutation and Site-specific Selection Driving B Cell Affinity Maturation is Consistent Across...Erick Matsen · New Directions in Probabilistic Models of Evolution