Figure 1.
Temporal genetic effect patterns.
Schematic view of gene expression patterns (top) and the relevant temporal genetic effects for these genes (bottom). The cartoons demonstrate a non-dynamic genetic effect pattern (A), a dynamic, linear genetic effect pattern (B), and a dynamic, non-linear genetic effect pattern (C). Top: shown are gene expression levels (y-axis) during a response to stimulation (x-axis). Each curve represents measurements in a different homozygous animal strain (segregants), where brown or black indicates whether the genotype of the associated genetic variant is or
, respectively, in each strain. Bottom: shown are genetic effects (that is, the change in gene expression between the
-carrying and
-carrying strains, y-axis) during a response to stimulation (x-axis). (C) Examples of non-linear genetic effect patterns, which are the focus of this study, including (left to right) a single state-transitioning pattern, which may be followed by a sustained new level of genetic effect, a single-pulse (impulse) pattern, and a multiple-pulse (complex) genetic effect pattern.
Figure 2.
A methodology for reconstructing genetic associations and their temporal genetic effect patterns from gene expression and genotyping data. (A) A cartoon example of input data, including the expression of a single gene over time for strains s1–s4 (top panel; shown as in Fig. 1A), and a typical genotyping of (homozygous) strains carrying either the (brown) or
(black) genotype in each genomic position (bottom panel). Correct and incorrect variants (v, u, respectively) are highlighted. (B) Shown are observed effect matrices for each time point from t1 to t4 (red, high-effect size; white, low-effect size). DyVER calculates the observed effects between each pair of strains carrying distinct alleles (strains carrying aa or
in columns and rows, respectively), using a variant u (left) or v (right). (C) Searching for the temporal two-state model that best fits the data. Shown are four cases, for two possible variants u, v, and two possible two-state models. The two states are ‘H’ (light blue) and ‘L’ (white) indicating high and low genetic effect, respectively. DyVER's fit of observed effects (high or low) in two Gaussians and the respective likelihood scores are presented in each case. For each variant, DyVER uses an HMM-based dynamic programming to identify its best-likelihood effect pattern. (D) A Manhattan plot of DyVER scores. Shown are likelihood ratio scores, called DyVER scores (y-axis), quantifying each variant (x-axis) with its selected temporal two-state model (from C). A dashed line indicates the significance threshold, generated using a permutation test.
Figure 3.
Comparative performance analysis on synthetic data.
Shown is the accuracy measure (scatter plots, left) and an example (histograms, right) across compared methods and different synthetic data parameters. Left: The accuracy measure (y-axis) using different patterns of genetic effects (impulse, single state-transitioning (sustained), linear, and complex sub-panels). Results are shown over genes that were measured in different numbers of time points (measures were averaged over effect sizes; x-axis, A), or over genes of different effect sizes (averaged over time points; x-axis B). Plots depict six alternative mapping methods (color coded). Right: Examples of performance (y-axis) using the four different dynamic effect patterns (color coded) across various methods (x-axis) for nine time points (A) or for genetic effect size 0.5 (B). The plots indicate that for non-linear genetic effect patterns, DyVER has an advantage over existing methods.
Figure 4.
A catalogue of dynamic, non-linear genetic effects in gene response following rapamycin treatment in yeast.
(A) Genetic effect profiles (left) and gene expression profiles (right) at six time points following rapamycin treatment (columns) for all genes identified by DyVER. Genetic effect values are the average increase (red) or decrease (purple) in effect size relative to non-stimulated cells (log-scaled). Gene expression values are the average increase (blue) or decrease (green) in gene expression relative to non-stimulated cells (log-scaled). Cis-associated genes are marked in gray (left color bar). Genes are partitioned into seven groups (C1–C7) based on their temporal two-state model (two state cartoons, shown as Fig. 2c, right; four singleton genes are omitted). (B) Four temporal two-state model groups C1, C2, C5, C7 (top to bottom). Left and middle panels: representative genes in each group. Left: gene expression of a representative gene (y-axis, log-scaled) across time points (x-axis). Each curve represents a different segregant, color coded by the best genetic variant found using DyVER (BY/black, RM/brown). Middle: genetic effect profiles of the representative gene, averaged across strains (log-scaled, y-axis) at each time point (x-axis). Right: shown are mean genetic effects (relative to non-stimulated cells, log-scaled; y-axis) and standard deviation (error bars) across time points (x-axis) for a certain group of genes.
Figure 5.
Co-associated genes typically share a similar pattern of genetic effects over time.
(A) Six gene modules (column 1), constructed on the basis of a shared trans-associated genetic variant (a genomic interval; column 2), are listed together with their known causal gene, if available (column 3; †-cis-associated causal gene, references are in parentheses) and the number of associated genes in a module (column 4). Significant enrichments in biological processes are detailed in column 5. Significant enrichments of temporal two-state patterns in each module are presented together with the description of these enriched patterns (columns 6 and 7, respectively). (B−E) Gene expression and genetic effects in modules nos. 1 (left), 3 (middle) and 4 (right). Gene expression (B) and genetic effect (C) of representative genes, as well as genetic effects of an entire module (D); plots are shown as in Fig. 4B. (E) Average gene expression (y-axis) at six time points (x-axis) for the known causal gene of each module. For cis-associated causal genes (modules nos. 3 and 4), brown and black indicate strains carrying the RM and BY alleles, respectively. The plots demonstrate the good match between the timing of abrupt changes in causal genes (E) and the timing of alterations in the observed genetic effects of their associated target genes (D).
Figure 6.
A genetic variant acting on the timing of response of the poor nitrogen-source degradation pathway (module no. 5-II).
(A) The genomic interval underlying module no. 5-II residing in Chr2: 533–562 kb. Shown are DyVER scores (y-axis) across the genomic positions in chromosome 2 (x-axis) for seven associated genes (color coded; the module includes only those six genes that cross the FDR 6% threshold). Positions of two potential causal variants, RPB5 and CNS1, are marked below. (B) Genetic effects, relative to non-stimulated genetic effects (y-axis, log-scaled) for different associated genes from A (color coded) at six time points (x-axis). The plot depicts the short impulse of high genetic effect in all associated genes. (C) Module genes, in the context of the poor nitrogen-source degradation pathway. Enzymes are shown as color-coded rectangles (bold-pink/module genes, pink/associated genes, white/non-associated genes). The pathways show the uptake of poor nitrogen sources (allantoate, allantoin, and GABA) and their degradation into ammonium. (D) A representative gene. Expression profiles (left) and genetic effects (right, y-axis) of DAL80, a gene in module no 5-II, during response to rapamycin (x-axis). Shown as in Fig. 4B but using a marker near the RPB5 gene (marked in A).