Why does AI struggle?
Inefficient training data
No robust underlying model
The way we train AI is fundamentally flawed (MIT technical report, 2020)
No Free Lunch Theorems for Optimization
Networks beyond simple differences
•
Conventional statistical analysis
◦
DEG 분석
•
Regulatory network를 도입
◦
다만, condition-specific regulatory network이 필요하다. (Individual의 차이, cell type의 차이를 반영)
▪
Analyze network topology and structure
▪
compare network topologies
▪
compare structure and expression
◦
Use individual networks as biomarkers!
•
netZoo: An integrated platform
Can we solve the “GWAS Puzzle?”
•
Rare variants = Dust?
•
eQTL analysis
◦
Which SNPs are correlated with the degree of gene expression
◦
Most people concentrate on cis-acting SNPs
eQTL Networks
•
Standard eQTL analysis
•
SNP와 gene의 association을 나타내는 bipartite graph를 구성한다.
◦
Network의 특성: 하나의 유전자와만 연관된 SNP가 많고, 여러 유전자와 연관된 SNP는 매우 적다 (hubs) (~scale free?)
◦
근데 Disease와 연관된 SNP는 hub SNP가 아니다! Hubs are GWAS desert!
•
Network comunity modularity에 기여하는 정도를 score로 하여 각 SNP에 할당. (core score)
◦
Significant SNPs 의 core score는 non-significant SNP보다 median이 20.3배 높더라.
→ In most instances it isn’t a single gene controlling a single traint, but a family of genetic variants that influence a process
References
•
Defining the role of common variation in the genomic and biological architecture of adult human height
◦
250,000 individuals, GWAS reveals 697 explain 20% of height, ~10,000 SNPs explain 30% of height → Individual variants have very small effect size
•
A saturated map of common genetic variants associated with human height
◦
12,111 SNPs account for 40%!
Can we model gene regulatory processes?
Integrative Network Inference: PANDA → Infers gene regulatory network
•
Template은 motif 기반의 TF-TG network
•
Co-expression is evidence for regulation → Expression data tkdyd
•
PPI is evidence for regulation → Interaction data 사용
References
Understanding tissue-specific gene regulation, Cell Reports
Reconstructing gene regulatory network
Single-sample Networks (LIONESS)
•
Sample i 를 제외하고 network를 만들었을 때 차이가 나는 network의 일부가 sample i의 contribution 이라는 아이디어
•
Sexual Dimorphism in Colorectal Cancer → Gene regulatory network이 drug metabolism 관련 pathway에서 차이가 나더라
References
Estimating Sample-specific regulatory network, iScience
Sex differences ~ , Cell Reports
Does this tell us something that co-expression does not?
Differential expression vs Differential co-expression
References
Gene targeting in Disease Neteworks, Front Gen
How do SNPs and TFs alter gene regulation?
EGRET: Extending the model to include genotype, Gen Res
CAD, CD → Differential regulatory activity에 의해 발생한다는 증거