Uniformly reprocessed yeast RNA-seq and Ribo-seq, with SPELL's microarray compendium. Search for genes that co-express with yours across every dataset, see where a gene or a gene set goes up or down, and compare gene sets.
Datasets where your genes move together get the most weight; every other gene is ranked by how closely it follows them there.
Score: weighted mean z-scored correlation with your genes. Click a gene for its page; + adds it to the query.
Weight: mean z-scored correlation among your genes in that dataset. Click one to see the expression.
Loading.
Every within-study contrast: samples of one strain in a condition against the same strain in standard growth (same lab, same protocol).
Quiet in the dataset's standard growth, at least 4 times its usual level in some samples of a condition (bulk RNA-seq, stranded, datasets with at least 6 samples).
Protein interactions, negative genetic interactions and the strongest co-expression partners of your genes. Every pair of genes in the view is linked when they interact or co-express above the cutoff, so groups that work together show up as dense knots.
The mean change of your genes in every contrast. Click a contrast to find others with a similar genome-wide change.
Expression level and variability in wild type, spread between labs, and genetic interactions, side by side. Matching the expression level removes the effect of level on variability.
Ranked by the median over the class's contrasts (each contrast one vote). Click a row for the gene.
Genes clustered by their profile over the classes; right, the GO process most over-represented. Click a row for its genes.
A check on the atlas: ribosome genes fall in stress and starvation; stress and glucose-repressed genes rise; meiosis genes rise in meiosis.
Data. The yeast recount (release v1.0): every public S. cerevisiae RNA-seq, nascent-RNA, Ribo-seq and 3′-end dataset we could reprocess, counted the same way, TMM-normalized counts per million. Every sample carries a curated label: genotype, strain background, condition class, and whether it is standard growth. The search also covers the microarray datasets of SPELL's compendium (from SGD; 500 rebuilt from their GEO deposits), as deposited (log ratios or log intensities), mapped to the same genes; their samples have names but no condition labels.
Co-expression search (after SPELL, Hibbs et al. 2007). In each dataset, genes are centred and reduced to the dataset's main patterns (up to 8 components; all of them when the dataset has 9 samples or fewer); the correlation of two genes is computed there, Fisher-transformed and z-scored within the dataset. A dataset's weight is the mean z-scored correlation among the query genes (negative weights are zero). Genes are ranked by their weighted mean correlation with the query over the heaviest datasets; a gene measured in less than half of that weight is not ranked. With one query gene, datasets are weighted by how much more the gene varies there than the dataset's typical gene. Test: for 300 GO processes, a query of 10 member genes puts the held-out members in the top 50 at about 17 times chance (median AUROC 0.78 with sequencing and microarrays together; 0.77 sequencing only; 0.76 microarrays only). Weighting all datasets equally does about as well on average (0.78), so the weights mainly show which datasets carry the signal. Keeping 16 components instead of 8 changed nothing (0.772 both).
Conditions. A contrast compares samples of one strain inside one dataset: a condition class against standard growth (or, for time courses with no standard sample, against the first time point), or mutants against the wild type. Values are log2 of the median CPM + 1. A gene set's value in a contrast is the mean over its genes.
Gene facts. Wild-type level and variability (CV) from wild-type samples; spread between labs from wild type in standard growth (between-lab and replicate SD of log2 expression); genetic interactions from Costanzo et al. 2016 (p < 0.05, |ε| > 0.08); poorly characterized = functional characterization index below 0.5.
Read with care. The meiosis class is mostly time courses compared with their own start in sporulation medium. GAL genes rise in meiosis because several studies synchronize meiosis with GAL-promoter systems. Gene × sample values affected by an overlapping gene or a plasmid decoy are left out everywhere.