starplast.screens
Published CRISPR screens and mass-spectrometry abundance, normalized to one gene per row.
Everything here is read from datasets/crispr_screens/ and toxo_stage_atlas/data/proteomics/; see
datasets/crispr_screens/SOURCES.md for provenance. Four screens and two proteomes, each with a very
different scope, and the scope is the thing to keep in view:
- GRA17 synthetic-lethal (PMC10409377) -- the only genome-wide one. Phenotype scores for RH and RHdelta-gra17 by passage, so the difference is the synthetic-lethality signal.
- GRA12 strains/subspecies (PMC12003902) -- two targeted pooled screens of ~235 genes. They are not replicates: their in-vivo L2FCs correlate at r = 0.41, so they are kept separate.
- in vivo CRISPR platform (PMC6722137) -- targeted, 35-147 genes. Its peritoneum/lung/liver/spleen
composite scores are almost certainly already in the tree as
fit_invivo_*, so this is corroboration rather than new signal, and is loaded under its own names so the two can be compared instead of merged. - host-transcription effectors (PMC12033024) -- a pooled single-cell screen scoring each effector by how much it perturbs host transcription (Hotelling T2).
The screens do not share a sign convention or a scale. That is established for the seven screens already in the node table, and nothing here changes it: always rank-normalize and check direction before pooling anything with anything.
Mass spectrometry is thin. The only abundance data in this tree is two Pru immunoprecipitation experiments covering 424 and 594 proteins -- under 8% of the proteome, and enrichment rather than a deep whole-cell proteome. It is loaded because it is real, and flagged because it is not coverage.
Return one row per gene, one column per published screen measurement.
resolve maps a raw accession to a current gene id (see identity.GeneIndex); without it, papers
that cite superseded accessions contribute nothing.
Per-effector host-response signatures from the dual perturb-seq differential expression.
The published table contains 33,531 host genes for each significant parasite effector. Putting thirty-three thousand mostly redundant columns into the node table would make the assay dominate by width, so the centered log-fold-change matrix is represented by its reproducible principal components plus the number and norm of substantial host effects. The parasite target still goes through the identity layer; symbols and the historical numeric target are not guessed as ME49 ids.
Median log2 iBAQ across replicates, from the Pru IP experiments.
Coverage is small and it is enrichment, not a deep proteome -- see the module docstring. Reported as a measured abundance where present and left missing elsewhere, never imputed to zero: absent from a 594-protein IP says almost nothing about a gene's true abundance.
Calcium-induced thermal-shift score per protein.
Verified against the calcium sensors: CAM1 and CAM2 sit at the 98th percentile and CAM3 at the 83rd. A protein whose melting curve does not move when calcium is added is not calcium-binding, so those three had to be near the top or the column would be measuring something else.
Note for anyone comparing this against the paper: its headline conclusion is about PP1, and PP1 is unremarkable HERE. That claim comes from the zaprinast time course in the same paper, which is a different experiment; this is the calcium mineCETSA sheet.
What co-purifies with the cyst wall, across every bait in the study.
Two numbers per protein: the strongest normalised spectral count it reached with any bait, and how many baits saw it at all. The second is the more honest of the two -- a protein found by thirteen independent pulldowns is in the cyst wall in a way that a single strong hit is not.
Most rows are human: the pulldowns were done on infected cultures and the table lists everything identified. Only rows naming a Toxoplasma accession are kept, which is 57 of 265.
Per-gene score from the genome-wide CRISPR screen for defence against oxidative stress.
Read from the authors' own Screening score sheet rather than recomputed from the guide counts
beside it: they published the score, so it is theirs to define.
Negative is required -- a gene whose disruption is depleted under oxidative challenge. The check that says the sign is right is catalase at -6.15, which is the extreme of the whole screen and is the enzyme that disposes of hydrogen peroxide.
Fitness cost of disrupting each gene over eight passages in the reporter strain.
A targeted library of nucleic-acid binding genes, so it covers 262 of them and not the genome. That is its honest extent, and for those genes it answers something the genome-wide RH screens cannot: whether essentiality holds in another background.
Verified by agreeing with them where it should. Against fit_invitro_hff it comes out at
rho = +0.62 over 130 shared genes -- close enough that the direction and the join are right, far
enough that the column is not a copy of one already here.
Per-gene differentiation phenotype from the reporter screen, as a log2 ratio.
Guide counts are summed per gene before the ratio is taken, which is the standard readout and also the only honest one at this depth: a single guide's count is noisy enough that a per-guide ratio averaged afterwards is dominated by whichever guide happened to be sampled least.
Counts are scaled to a common library size first. Two arms are averaged. Genes seen in neither arm are absent rather than zero -- a gene no guide covered is not a gene with no phenotype.
This is COMPUTED here rather than read from the authors' table, because the deposit publishes counts and not the ratio. It is the comparison their design names, and the column is called a ratio rather than a phenotype so nobody mistakes it for something they reported.