genomics - functional assays II computation

MAVE

multiplexed assay of variant effect

i think this is the same as deep mutational scans

needs sophisticated models to predict effect of variants in proteins that havent yet been assayed, to account for noise and epistasis

problems with mpra measurements

ratio of counts (RNA/DNA) can be unstable

multiple barcodes might be tagging the same element

transfection of complex libraries across biological replicates can introduce significant variation

problems w barcode rna/dna ratios

ratio of 2 noisy measurements

when read counts are low, small differences in reads lead to very large changes in ratio

errors across measurements of multiple barcodes can be propagated to estimate variance, but common approaches dont do that

RNA = a x DNA

a = transcription rate

assuming linear relationship between dna and rna molecules

plug a into dna and rna linear regression model → produces lower variance in the estimate of transcriptional activity. increases power to detect functional impacts of alternative alleles

what can a genotype - phenotype map look like?

additive (all positions in a sequence contribute independently)

neighbor (interaction terms between neighboring positions)

pairwise (each position interacts w every other position)

black box (roll your own relationship function lol)

biophysical (model ∆∆Gs) w neural network

sort seq expression values

discrete data - limited by num of sorted bins

barcode rna-seq

continuous data

global epistasis

mutations may act non-additively on phenotype, but this may be due to additive effects on some underlying trait

local epistasis

specific pairwise interactions between residues

modeling takeways

mave output is sequencing data, not a direct measure of the property of interest

modeling translates your sequencing data into a measure of the function of interest

a quantitative sequence to function model enables you to predict the effects of unmeasured variants and to learn epistasis

DeepSTARR

deep learning + starr seq

can be used to predict enhancer activity based on dna sequence, and then design synthetic cres

biophysical models of variant effect

capturing biophysical effects underlying the phenotype will lead to better predictions of combinations of variants

big mutant library, include single mutations and many double mutants → assay variant effect in 2 ways (protein binding and protein stability/abundance) → train a neural network model w biophysical ∆G values → ∆G values produce biophysically-based, interpretable

predictions → determine free energy changes → quantify how mutations tune binding affinity

protein fragment complementation (PCA)

bindingPCA: binding = yeast survives

abundancePCA: reporter for how much of the protein is in the cell

recommendations for using MAVEs for variant interpretation

assay should have sufficient dynamic range to robustly separate damaging variants from benign variants

choose an assay that captures effects that are relevant (what variants might not be detected in the assay?)

deposit data in repositories, disclose all experimental and statistical methods, make code available

provide measures of reproducibility and error estimates

validate results w some single variant tests and include known variants as controls