ProbGen24 Day3

Brian Clarke

DeepRVAT uses deep learning to find non-linear associations for rare variants. Improves phenotype prediction.

https://doi.org/10.1101/2023.07.12.548506

Roka Borbely

A simple biophysical model of transcriptional regulation.

cis-regulatory element. Occupancy of a binding site: increases concentration, reduces mismatch number

Binding site at every position.

Total occupancy: additive contribution of all BSs

Time vs fitness.

Theoretical limit: low TF concentration, sharp activation threshold.

Yun S. Song

Nonparametric density estimation using deep learning.

Protein language models: mask out some position and try to predict the masked amino acid. Very effective. What sequence should be used in training the model? Can we apply the same model to DNA? What part of the genome shall we use to train? In fact, using the whole genome does not work.

Genomic Pre-trained Network (GPN).

Why should it work? Functional information occurs multiple times in the genome.

Challenges

• Highly variable levels of conservation along the genome.
• Curse of repeats. Repeatitive elements, overly represented.

Arabidopsis Thaliana, simpler - 20% repetitive compared to 50% in humans. PNAS 2023, Benegas et al.

Model could predict genomic context (appeared as separate clusters in UMAP). It learned context-dependent probability distribution. GPN understands pathogenicity of different variant types. Ranking of variant types is correlated with pathogenicity. Low GPN scores related to enrichment of GWAS hits.

DNA LLM for Humans

NVIDIA tried using 128 GPUs for 28 days. But failed.

Nucleotide Transformer. GPN-MSA performs significantly better than HyenaDNA and NVIDIA model.

Matthew Aguirre

GRN graph network construction

Bollobas et al 2003. Extend that algorithm, construct sparse graph with n nodes, put into k groups. Sparsity \(p\). Group affinity \(w\). In-degree uniformity \(\delta_{\textrm{in}}\). After constructing the graph, use difference equation to simulate expression of genes: terms for synthesis, degradation and noise. Find steady-state expression for genes.

Perturb a gene: Measure effects. In-silico knockout. Perturbation effects in many structurally diverse GRNs. Compare synthetic GRNs with Perturb-Seq, Replogle et al, Cell 2022.

Realistic synthetic GRNs. Correlation tracks co-regulation due to sparsity and hub regulation. Co-regulation is still hypothesis driven.