Data Driven "Pathway" Analysis with ADAGE
summary
An unsupervised denoising autoencoder approach called ADAGE mines noisy public expression compendia to recover pathway- and gene‑level signals without prior annotations; an ensemble version eADAGE improves stability across runs. Applied to a 1,000‑sample Pseudomonas dataset (Pseudomonas aeruginosa), models capture roughly 80–100 KEGG pathways, need ~500 samples for pathway resolution, and give ~45% nearest‑neighbor function prediction versus ~5% random. Node changes map to biology and validate by perturbation (anr knockout RNA‑seq shows denitrification and quorum sensing), and networks derive from gene weight correlations.
AI-written summary · about
chapters (20)
- 0:00Intro
- 2:05What is ADAGE
- 3:20Why ADAGE
- 4:10Data Overview
- 5:50High Weight Genes
- 7:00Protein Function Prediction
- 9:25Neural Networks
- 10:20Model Characteristics
- 10:55Performance Characteristics
- 12:00Networks
- 12:40Similarity
- 14:15Correlation
- 16:00Neural Network Model
- 16:25Ensemble Approach
- 17:25Deep Architecture
- 20:25Capturing Pathways
- 21:00Our Goal
- 21:35Pathway Style Analysis
- 25:30Research Parasite Awards
- 26:00Avoid the Situation
related talks
01h 19mAnalyzing Genetic Networks Using Neural Networks
27mA Statistical, Reference-Free Algorithm Subsumes Myriad Problems in Genome Science
01h 06mIntroduction to Biological Network Analysis III: Identifying Network Modules
27mAutomated Discovery of Cancer Types from Genes
38mInterpreting Deep Learning Models Of Functional Genomics Data To Decode Regulatory Sequence...
30mLeveraging Molecular Data for Drug Discovery
29mMachine Learning Techniques for Omics Data Analysis
01h 08mUnmeasured Confounding and More Recent Developments/Challenges in Causal Discovery