Single-oocyte proteomics of Diploria labyrinthiformis: turbidity vs control

Results at a glance

Counted directly from the DE tables. Significant = adjusted p < 0.05 (Benjamini–Hochberg within each contrast; de_provenance.json), the only rule the DE applied — no fold-change filter. Up / down = sign of the fold change. Positive log2FC = higher in the first-named group of the contrast.

Overview

This report analyses **single Diploria labyrinthiformis coral oocytes (about 500 µm, symbiont-free)**, each prepared in the Kültz laboratory without reduction or alkylation (a stage-tip-based workflow; digestion details to be confirmed with the laboratory) and acquired there on a nanoElute 2–Bruker impact II by data-independent acquisition (DIA). The data arrived as mzML files. Nine oocytes came from a turbidity-exposure group and five from a control group. The study was designed as a proof of concept for single-cell proteomics (SCP) in a non-model coral, and the comparison of turbidity vs control is a first look rather than a powered experiment.

Depth. The search identified 2,809 protein groups across the 12 oocytes analysed, and each oocyte had 1,652 to 2,599 proteins measured directly (at least one precursor observed in that run), with a median of about 2,000. Between 5,821 and 15,418 precursors per oocyte passed the 1% false discovery rate. That depth from a single cell, on an instrument without ion mobility, is the main technical result.

Comparison. One comparison was tested, Turbidity vs Control, using the 9 turbidity oocytes and 3 of the 5 controls. Two control oocytes (BL10_e_b and BL10_e_d) identified only 889 and 1,966 precursors, against 5,821 or more for every other run, so they are left out of the main analysis (see Data Quality Notes). The Results at a glance panel above gives the counts: 15 proteins differ at an adjusted p-value below 0.05, 10 lower and 5 higher in turbidity-exposed oocytes. All 2,809 protein groups were tested.

Annotation. The coral proteome used for the search carries gene-model codes only (for example g6863.t1). Every protein was matched against UniProt SwissProt with DIAMOND, and the best-hit gene name is shown wherever one exists. These names are homology labels, not identifications: the median identity of the best hit is 53% among the identified proteins, so a name such as "CYP10" means "most similar to cytochrome P450 10", not "is CYP10".

Three cautions shape everything below, and make this an exploratory, feasibility result rather than a finding about turbidity. First, the three controls are few, so the comparison has little statistical power and only large changes can reach significance. Second, every turbidity oocyte was injected before every control. The tests for that confound in Data Quality Notes did not detect a run-order effect, but they can only detect large ones, so run order cannot be excluded. Third, the hit list depends on normalisation: DIA-NN's normalisation was unstable in every run, and without it only 1 protein reaches significance.

QC Assessment

Each oocyte had 59% to 92% of the 2,809 protein groups measured directly. The rest are values that DPC-Quant inferred from the detection-probability model. Depth varies about 1.6-fold between oocytes: BL10_e_8 (Turbidity) is the deepest at 2,599 proteins measured, and BL10_e_2 (Turbidity) and BL10_e_c (Control) are the shallowest at about 1,650 to 1,680. Both groups span much of that range, so depth is not obviously split by group among the 12 oocytes analysed. The control group's median is lower (1,818 vs 2,052 proteins measured), with only three controls behind it.

The PCA shows no clean separation between the groups. The control and turbidity hulls overlap on PC1 (18.4% of variance) and PC2 (17.3%). Two controls sit together on the right, and the third falls among the turbidity oocytes. Two turbidity oocytes, BL10_e_12 and BL10_e_8, lie more than twice the median distance from their group's centre. Neither is a low-depth run (BL10_e_8 is the deepest), so they look like biological or preparation variation rather than failed injections. The weak global separation is what a small or subtle treatment effect looks like, and it matches the small number of significant proteins.

Reproducibility, computed on the linear scale from measured values only, is a median CV of 35.6% in turbidity oocytes (2,656 proteins with at least two measured values) and 32.8% in controls (1,954 proteins). That spread is typical of single-cell measurements, where biology and handling both vary from cell to cell.

The p-value distribution for the comparison has a peak near zero over a flat background, which is consistent with a calibrated model and some signal (Appendix: p-value calibration). The run-order control test below, which compares turbidity oocytes with each other, gives no significant proteins, as a calibrated model should when there is no treatment difference.

Fitness for differential analysis. The 12 oocytes analysed are fit for an exploratory comparison. Depth is good for single cells and the groups have similar spread. The limits are the three controls and the run order, both discussed below.

Key Findings Per Comparison

Turbidity vs Control

Fifteen of 2,809 proteins are significant (adjusted p < 0.05, Benjamini–Hochberg). The non-significant proteins are centred at +0.18 log2, close to zero, and the up:down ratio is 5:10, so the result is not dominated by a global shift. All 15 also exceed the 2-fold reference line; with three controls, only large changes can reach significance, and the fold changes of hits found at this power are likely to be overestimates. Several hits sit just under the threshold (four at adjusted p 0.046), so the list is fragile at its lower end.

Lower in turbidity-exposed oocytes (ranked by adjusted p; names are best-hit homology labels):

Higher in turbidity-exposed oocytes:

Four of the 15 are presence calls: PTGR3-like was not detected in any turbidity oocyte, and RAD23A-like, the perlucin-like protein and g11088.t1 were not detected in any control, after match-between-runs. For these the difference is detected vs not detected, and the size of the fold change comes from the missing-value model, not from a measurement. Several others are partly inferred with the turbidity side thin (CYP10-like detected in 2 of 9 turbidity oocytes, HTD2-like in 1 of 9, MCEE-like in 3 of 9), so their fold changes should not be quoted. Five hits rest on a single peptide (ATPAF1-like, HTD2-like, MCEE-like, the perlucin-like protein and g11088.t1) and are leads only. The tubulin alpha-1-like hit is a close relative of a coral tubulin whose shared peptides were assigned to a contaminant entry (Data Quality Notes), so its fold change is less certain.

The violin panel shows the pattern directly: for most of the "lower in turbidity" proteins the control oocytes are all measured (filled points), while many turbidity values are hollow, meaning inferred. The difference is real in the sense that these proteins were seen consistently in the controls and seldom in turbidity oocytes. How large it is cannot be read off the fold change.

Sensitivity analysis with all 14 oocytes. Re-running the comparison with the two weak controls included gives 226 significant proteins, 222 higher and 4 lower in turbidity oocytes:

Almost all of these go one way, which points to a global intensity or depth difference rather than a treatment effect. The two weak controls carry far fewer measured proteins, the model infers low values for the rest, and the control mean falls for hundreds of proteins at once. Before normalisation the control group sat 0.70 log2 below the turbidity group in this all-14 design, with 82% of proteins shifted the same way. Six of the 15 main-analysis hits remain significant with all 14 oocytes. The main analysis, without the two failed injections, is the one to interpret. A smaller depth difference remains in it (median 1,818 proteins measured per control oocyte vs 2,052 per turbidity oocyte), which would push results toward "higher in turbidity": the 10 hits lower in turbidity run against that bias, and the 3 presence calls higher in turbidity are the ones most exposed to it.

Cross-Comparison Biomarkers

Only one comparison was tested, so there are no proteins shared across comparisons. The heatmap shows how the 15 significant proteins behave across the 12 oocytes:

The controls form a consistent block for the "lower in turbidity" proteins, and the turbidity oocytes vary more among themselves. About a quarter of the cells shown are inferred values, so a uniform block can partly reflect the model's estimates for proteins not seen in that oocyte; the violin panel above marks which values were measured.

High-Confidence Findings

By the tier rule (significant and measured in every oocyte of both groups), four proteins carry measured ratios rather than detection events:

• GCN1-like (g10403.t1; log2FC −1.35, adj.P 0.019; 25 peptides). GCN1 senses stalled and colliding ribosomes and activates the integrated stress response through GCN2. Lower GCN1-like protein in turbidity-exposed oocytes is a measured change, and the best-measured hit in the set: it rests on 25 peptides, detected in every oocyte.
• RTN4IP1-like (g5801.t1; −1.55, adj.P 0.036). A mitochondrial oxidoreductase linked to mitochondrial respiratory function.
• CYP27A1-like, sterol 26-hydroxylase (g15648.t1; −2.11, adj.P 0.046; 12 peptides). A mitochondrial cytochrome P450 of sterol metabolism.
• TLL2-like, tolloid-like 2 (g14309.t1;g14309.t2; +1.70, adj.P 0.046). Tolloid-family metalloproteases process extracellular-matrix precursors and modulate BMP signalling in development.

None of the four shows a significant trend with injection order inside the turbidity group (see Data Quality Notes), and none is a presence call.

Biological Interpretation

The proteins lower in turbidity-exposed oocytes have best hits in mitochondrial and lipid or sterol metabolism: two cytochrome P450-like proteins (CYP10-like and CYP27A1-like), an HTD2-like protein (HTD2 acts in mitochondrial fatty-acid synthesis), a methylmalonyl-CoA epimerase-like protein, an ATP-synthase assembly factor-like protein (ATPAF1-like) and RTN4IP1-like. One hypothesis is that oocytes from the turbidity group carry less of this machinery. Coral oocytes are symbiont-free, so any link to turbidity would run through maternal provisioning: less light, less symbiont photosynthesis in the parent colony, and less energy or lipid invested in each egg. This is a hypothesis for the proposal, not a conclusion. It rests on about ten homology labels, five of which are presence calls or mostly inferred on the turbidity side, and abundant mitochondrial proteins (for example ATP synthase, VDAC and the phosphate carrier) remain among the most abundant proteins in every oocyte, so there is no sign of a global mitochondrial loss. Labels near 30–40% identity, such as CYP10-like and the caffeoyl-CoA O-methyltransferase-like protein (a plant-type name), say little about the coral protein's actual function.

The proteins higher in turbidity oocytes are a mix: a perlucin-like C-type lectin (perlucins were first described from mollusc shell matrix; what a perlucin-like protein does in a coral oocyte is unknown), tolloid-like 2 (tolloid-family proteases process extracellular-matrix precursors and modulate developmental signalling), RAD23A-like (DNA repair and proteasomal delivery of ubiquitinated proteins) and tubulin alpha-1-like. Three of these are presence calls, so they are leads to confirm rather than findings.

Confidence. Low. The comparison rests on three controls, all injected after the turbidity oocytes, the hit list depends on the normalisation used, and many hits rest on detection or a single peptide rather than measured ratios. Its value for the proposal is as a demonstration that single-oocyte proteomics resolves about 2,000 proteins per cell and can surface candidate biology, not as evidence of a turbidity response.

The dark proteome

The proposal defines dark = functionally not annotated. Every identified protein was searched against SwissProt (DIAMOND, sensitive mode), and those without a SwissProt hit were searched against all of NCBI nr, keeping hits to other species only.

A protein with a functionally named nr hit counts as annotated in every row. Proteins with only a weak SwissProt hit were not searched against nr, so the stricter rows may slightly overstate the dark share. About 40% identity is a commonly used threshold for transferring function by homology. On that rule, about a quarter of the proteins measured in a single coral oocyte (698 of 2,809) have at best a low-confidence functional annotation: no characterised relative close enough to say what they do. Almost every protein has some distant relative: only 67 lack even a weak SwissProt hit or a functionally named nr hit. Counts are of protein groups, and a group can hold several gene models, so the number of distinct dark genes may be somewhat higher. Of those 67, 66 have nr homologs that are themselves uncharacterised, 53 of them only in other cnidarians, and one has no homolog in nr at all.

Dark proteins are not confined to trace levels: 14 of the 100 most abundant proteins are dark under the 40% rule (3 under the any-hit rule). None of the 20 most abundant is dark under either rule. One significant protein, g3779.t1, is dark under every rule. It has homologs only in other cnidarians, all uncharacterised, and it is lower in turbidity-exposed oocytes (log2FC −2.49, measured in 8 of 9 turbidity and 3 of 3 control oocytes). A cnidarian-specific protein of unknown function that responds to an environmental exposure is the kind of candidate the proposal's functional-inference aims are built around.

Coral peptides in the Core's search archive (FRAN)

FRAN is the UC Davis Proteomics Core's archive of the peptides identified in its own past searches, thousands of searches of many species. Every one of the 17,213 coral peptides identified here (1% FDR, contaminant entries excluded) was looked up in it by exact sequence. This coral search is not yet in FRAN, so every match comes from other projects.

Peptides from dark coral proteins are about twelve times rarer in the archive than peptides from annotated proteins, and none of the 186 peptides from the fully dark proteins has been seen in any other Core search. The peptides that recur belong to conserved machinery: 1,064 of them occur in 100 or more searches, mostly of mammalian samples. At the protein level, 740 of the 2,809 quantified protein groups share at least one peptide with another Core search, against 28 of the 698 dark protein groups. A cross-species sequence archive therefore separates a single cell's conserved, recognisable proteome from sequence that has never been observed before, which is a practical way to flag dark-proteome candidates for follow-up. The archive's organism records are not reliable enough to break the matches down by species precisely.

How This Analysis Works

LC-MS/MS. Each oocyte's proteins were cut into peptides by enzymes, like cutting a long sentence into words. The peptides flowed through a thin column that sorts them by how "sticky" they are, so they leave it over about an hour. As they left, the mass spectrometer weighed them (MS1), then broke them apart and weighed the pieces (MS2). The pattern of pieces reads out each peptide's sequence.

Data-independent acquisition (DIA). Older methods pick the loudest peptides one at a time and miss the quiet ones. DIA instead fragments everything in a series of mass windows on every cycle, like photographing the whole room rather than following one person. Nothing is skipped, which matters when each sample is one irreplaceable cell, but the resulting spectra are mixtures that need specialised software to untangle.

DIA-NN untangles them. It predicts, from the coral protein sequences, what every possible peptide should look like, and uses neural networks to score which predicted peptides are really in the data. Here it found 5,821 to 15,418 peptides (precursors) per oocyte at a 1% false discovery rate.

limpa and limma. A protein is often measured in some oocytes and missing in others, especially at single-cell input. Rather than filling the gaps with guesses or throwing those proteins away, limpa models how likely a peptide is to be seen at a given abundance (the detection-probability curve), and uses that to estimate protein amounts with honest uncertainty. limma then tests each protein, borrowing information across all 2,809 proteins to steady the variance estimates, which is especially valuable with three controls.

Key terms, with this dataset:

• log2 fold change (logFC). +1 means twice as much in turbidity oocytes, −1 half as much. GCN1-like has logFC −1.35, measured in every oocyte: roughly 2.5 times lower in turbidity oocytes.
• p-value. If a protein truly did not change, the chance of seeing a difference this big from noise alone.
• Adjusted p-value (false discovery rate). Testing 2,809 proteins is like flipping 2,809 coins: some will land heads ten times running by luck. The Benjamini–Hochberg correction raises the bar so that, among the 15 proteins called significant, the expected share of false positives is at most 5%, under one protein on average.
• Volcano plot. Each dot is a protein: left–right is how much it changed, up is how confident the test is. Big, confident changes sit at the top corners, which gives the volcano shape.
• Coefficient of variation (CV). Standard deviation divided by the mean, from measured values only. The median CV here is 33–36%, normal for single cells.
• Normalisation. Each oocyte contributes a slightly different amount of protein to the instrument. Normalisation rescales the runs so that differences in loading are not mistaken for biology. Here DIA-NN's cross-run normalisation was used.

Methods (publication-ready) & Reproducibility

Single Diploria labyrinthiformis oocytes were prepared in the Kültz laboratory without reduction or alkylation (stage-tip-based workflow; digestion details to be confirmed) and acquired there on a nanoElute 2 coupled to a Bruker impact II quadrupole time-of-flight mass spectrometer (micrOTOFcontrol 6.3) by data-independent acquisition, with MS1 scans over m/z 100–2,000 and 42 isolation windows of 21 m/z (1 m/z overlap) covering m/z 389.5–1,230.5, at a cycle time of 0.83 s over 33–36 min of acquisition per run. Data were converted to mzML with ProteoWizard (Bruker reader 3.0.21141). LC conditions and collision energies are not recorded in the mzML files. Spectra were searched with DIA-NN 2.7.0 (Demichev et al., Nature Methods, 2020) in library-free mode against a spectral library predicted from the laboratory's D. labyrinthiformis proteome (28,405 gene models) appended with 381 common contaminant sequences. Search settings were in silico cleavage after K and R with up to one missed cleavage, peptides of 7–30 residues, precursor charge 2–4, precursor m/z 389–1231 (measured from the isolation windows), fragment m/z 200–1800, N-terminal methionine excision, no fixed cysteine modification and no variable modifications. MS1 and MS2 mass accuracy were fixed at 15 ppm; DIA-NN's scan window radius was measured on three representative runs and fixed at 9 spectra. Match-between-runs (--reanalyse) was enabled, precursors were filtered at 1% FDR, and contaminant precursors were excluded from quantification. Protein quantities were estimated with DPC-Quant in limpa 1.2.5 from DIA-NN's normalised precursor intensities, with missing values modelled through the detection-probability curve rather than imputed (Li, Cobbold & Smyth, bioRxiv 2025, 10.1101/2025.04.28.651125; Li & Smyth, Bioinformatics 2023). Differential abundance was tested with limma 3.66.0 (empirical Bayes moderation) in R 4.5.3, with Benjamini–Hochberg correction; proteins with adjusted p < 0.05 were called significant, with no fold-change filter. Two control oocytes with fewer than 2,000 precursors identified were excluded. Protein groups were annotated by DIAMOND 2.1.7 blastp (--more-sensitive, E ≤ 1e-5, best hit) against UniProt SwissProt; proteins without a SwissProt hit were searched against NCBI nr at default and more-sensitive settings, and taxonomy was assigned from NCBI nodes.

A full reproducibility bundle accompanies this analysis (output/reproducibility/): which skill produced it and how it was installed, a pinned conda environment, R sessionInfo, exact tool versions, input and output checksums, and a runnable reproduce.sh. The analysis was produced by the UC Davis Proteomics Core pipeline Claude skill and can be reproduced from that bundle (see reproducibility/REPRODUCE.md). The statistical analysis on its own is tables/reproducibility_log.R, plain R with every value written out.

Instrument & Acquisition

Samples were acquired in the Kültz laboratory on a Bruker impact II quadrupole time-of-flight mass spectrometer (no ion mobility; nanoelectrospray source; acquisition software micrOTOFcontrol 6.3) coupled to a nanoElute 2. The mzML files record data-independent acquisition with an MS1 scan over m/z 100–2,000 followed by 42 isolation windows of 21 m/z, overlapping by 1 m/z and covering m/z 389.5–1,230.5, repeated every 0.83 s; runs lasted 1,983 or 2,163 s (about 33 or 36 min). The files were converted from the Bruker format with ProteoWizard (Bruker reader 3.0.21141). The LC column, gradient and collision energies are not recorded in the mzML files and should be taken from the Kültz laboratory's methods.

Next steps

1. Follow up first (measured in both groups): GCN1-like (g10403.t1), RTN4IP1-like (g5801.t1), CYP27A1-like (g15648.t1), TLL2-like (g14309.t1;g14309.t2).
2. Check the counts (partly inferred): g3779.t1 (dark; measured in 8/9 turbidity and 3/3 control oocytes), OMT5-like (7/9 · 3/3), ATPAF1-like (6/9 · 3/3), tubulin alpha-1 (9/9 · 2/3); then MCEE-like (3/9 · 3/3), CYP10-like (2/9 · 3/3) and HTD2-like (1/9 · 3/3), whose turbidity side is mostly inferred.
3. Confirm presence (detection events): PTGR3-like (never measured in turbidity oocytes); RAD23A-like, the perlucin-like protein and g11088.t1 (never measured in controls). Confirm with a targeted method before quoting a fold change.

Whatever its tier, a hit measured in fewer than half the runs of every group is a lead only. For the next experiment: more control oocytes, injection order randomised across groups, and lower enzyme amounts per cell.

See README.html at the top of the results folder for where every file is and how to reuse them.

Top proteins per contrast

The 20 proteins with the smallest adjusted p per contrast, from the DE tables (the full lists are the DE_*.csv files). Positive log2FC = higher in the first-named group of the contrast. A row flagged Ig chain (the antibody's own chains) or contaminant is not the sample's biology.

Appendix: p-value calibration