Flow Advanced Analysis
Dimensionality reduction, clustering, and the specialized tools for cell cycle, proliferation, kinetics, and marker intensity. It all runs in your browser and consumes no compute usage. Clustering and dimensionality reduction can also run on a server, one file at a time, and that's the only part of this page that uses any.
Dimensionality Reduction
Dim. Reduction in the sidebar's Analysis section collapses a high-dimensional panel into two axes you can look at. Three methods:
| Method | What it is |
|---|---|
| UMAP | Uniform Manifold Approximation and Projection. It keeps local clusters intact and preserves how they sit relative to each other, which makes it the usual first thing to try. |
| t-SNE | Classic t-SNE, with perplexity and learning rate |
| PCA | Principal components, which is linear, fast, and interpretable |
Pick at least two parameters. Each method has its own settings, and the defaults adapt to how many events you have, so you can usually press run and look at the result before deciding what to tune. Subsampling is on by default for large files, and the panel says how many events are actually being used.
When it finishes, Conspecta adds a scatter plot on the new axes. Those axes then appear in the Computed section of any axis menu, so you can put them on other plots or gate on them like anything else.
t-SNE and PCA results coexist, so you can compare two embeddings of the same data side by side.
Clustering
Clustering finds populations without you drawing them. Five algorithms:
| Algorithm | Notes |
|---|---|
| K-means | You choose how many clusters |
| DBSCAN | Density-based, so it finds clusters of unequal size and leaves noise unassigned |
| SOM clustering | Self-organizing map with metaclustering, in the style of FlowSOM |
| Graph clustering | Neighbor-graph community detection, in the style of PhenoGraph |
| Density | Grid-based peak finding on two parameters |
The SOM and graph methods are implementations in that style rather than ports of the R packages, and Conspecta says so in the dialog box. Report them as what they are.
How many parameters a run can take depends on the algorithm. K-means, SOM clustering, and graph clustering accept up to 64 parameters, so a full high-parameter panel fits in one run. DBSCAN stays at eight, and that's a real limit rather than a speed setting: past eight dimensions, distances between events stop separating populations in a way you can trust. Density clustering works on the two axes you pick.
All selects your fluorescence channels and leaves scatter and time out, since cell size and acquisition order say nothing about your markers. If you select more than eight parameters and then switch to DBSCAN, nothing is dropped. The dialog box keeps your selection, says how many channels to remove, and holds the run until you're back under the limit.
The dialog box has two steps. The first is the run: algorithm, what it covers, parameters, settings. Once the run finishes it moves to the second step, which shows how many clusters were found, how large each is as a share of the total, and lets you rename them. The heading names the run you're looking at. The back arrow returns to the setup form with your choices as you left them, so you can change one setting and run again; the new answer replaces the old one. Create N Gates turns the clusters you tick into real polygon gates on the projection you choose, at which point they behave like gates you drew yourself and appear in every statistic.
The results step carries the same export actions as the other statistics screens. Export CSV downloads the clusters and their sizes, or for a run across several files, each file's share of each cluster. Send to Data Table puts the same numbers into a data table you own in the project. Names you gave the clusters come through in both. A run across several files also offers Add to figure, which charts each file's share of each cluster onto a figure. A single-file run doesn't offer it, because its cluster sizes compare nothing.
A scatter plot colored by cluster is added automatically, titled with the algorithm and the channels it clustered, and clicking its title bar (View run) opens the run that made it. Only one run's colors are on screen at a time, so a new run takes the colors from an earlier plot rather than quietly repainting it, but the earlier plot keeps its own record and its title bar reads Not colored until you run it again. A result is only drawn against the file it was computed from, so if you switch files the plot says it has no data rather than coloring the wrong cells. Applying or removing compensation clears the clustering, since the assignments describe values the events no longer have.
Close the dialog box and reopen it and you're back on your last run, with the same result, the same cluster names, and the settings that produced it waiting behind the back arrow. The newest run's cluster assignments are saved with the analysis too, so after a reload its plot comes back colored with no re-run, and members with view-only access see the colors as well. Not colored still appears where the saved colors would be wrong or missing: a plot whose run was replaced by a newer one, any clustering after you apply or remove compensation (the saved assignments describe the values the events had when it ran), and runs saved before Conspecta kept assignments.
Click the plot's title bar, or reopen the dialog box, and you land on the run's record: who ran it and when, how long it took, whether it computed in your browser or on the server, the exact settings, and the clusters with the names you gave them. Every run the analysis keeps is listed under Runs on this analysis, so any of them is one click away, and New run opens a fresh setup form. Re-run recomputes the run you are looking at with its own settings and brings the colors back to that same plot. The back arrow opens the setup form with those settings filled in, and if you change them the run button lets you choose whether the result updates that plot or lands on a new one. It updates the plot unless you say otherwise.
Running Across Several Files
Run a reduction or clustering on one file and then on another, and the two results have nothing to do with each other: the axes are unrelated, and cluster 1 in one file is not cluster 1 in the next. To compare samples you have to analyze them together.
Run on in both dialog boxes does that. Leave it on This file for the file you're looking at, or choose Working set to cover every loaded file at once. Conspecta takes an equal number of events from each file and runs the algorithm once over all of them, so every file lands on the same axes and inside the same cluster boundaries. Nothing is merged and no combined file appears in the sidebar, because the pooled data only lives for the length of the run. A working-set run happens in your browser on Local, free, or on Conspecta's machines when you pick Server. That's the same choice a single-file run offers, and the server draws the same events a local run would. See Flow Cytometry Compute for what a server run uses.
Events per file is the draw from each file rather than a total. That matters when one tube ran four times as long as the others: without a per-file cap it would dominate the map and the result would describe that tube more than the experiment.
Only parameters present in every file in the set can be used, so the parameter list narrows when you switch to the working set and Conspecta names the file that's missing a channel. This is deliberate. Filling in a channel a file doesn't have would put all of that file's events at the same value, and they'd separate into a tight cluster that looks like a real population.
For clustering, the results add a table of what share of each file's events landed in each cluster, which is the comparison pooling exists to make. Because every file went through one run, a cluster means the same population in every row.
Create N Gates still draws on the file you're looking at, and outlines each cluster around that file's own events in it, so the boundary matches the plot in front of you rather than the pool. A cluster none of this file's events landed in has nothing to enclose here, so it gets no gate.
Density clustering has no scope option: it reads one file's own density landscape, so there's nothing for pooling to add.
Specialized Tools
These live in the Cell Analysis menu under Analysis. Each one shows a green dot once it has a saved result. A saved result belongs to the analysis: it reopens with it, and the card names the channel and file it came from and how long ago it ran. Changing compensation clears it, since the numbers describe values the events no longer have.
Cell Cycle fits a DNA histogram into G0/G1, S, and G2/M, finding the DNA channel by name where it can (DAPI, PI, Hoechst, DRAQ, 7-AAD, SYTO 9, SYTOX Green). It fits Gaussian peaks to the histogram and reports the phase percentages and peak positions. It doesn't create gates. Save Results keeps the numbers against the analysis.
Proliferation works on a dye-dilution channel (CFSE, CellTrace, CPD), splits the distribution into generations, and reports the division index, the proliferation index, and the percentage of cells that divided, weighting each generation by the precursors it came from.
Kinetics looks for instability across acquisition time and can gate out the unstable windows.
Quality Control does the same job with the emphasis on excluding bad time windows, and is covered in Flow files and templates.
Applying a kinetics or quality-control time gate re-parents your tree. Conspecta creates a time gate and moves every top-level gate and plot under it, so every population downstream is filtered by it. That's what makes it work, but it's a structural change worth expecting.
MFI Comparison
MFI Comparison compares two populations marker by marker. Pick a control population and a sample population, tick which markers you care about, and you get a table of the median fluorescence intensity in each, the fold change, the log₂ fold change, and a p-value from a Kolmogorov-Smirnov test, with significant rows in bold. A bar beside each row shows the direction and size of the shift.
It needs at least two gates before it will open, since it compares populations. Where the control's median is zero or below, the fold change reads as not applicable rather than inventing an infinity.
Export CSV takes the whole table.
Where It All Runs
These run in your browser, on your machine. That means no queue, no usage, and no data leaving your session, but it also means a very large file is limited by your laptop. Subsampling is the lever, and each dialog box tells you what it's using.
Clustering and Dim. Reduction are the exception. Both carry a Compute Location setting with Local and Server, and Server hands the run to Conspecta's machines for the runs your laptop can't get through, whether that's this file alone or the whole working set. That consumes compute usage: 1 for most algorithms, 5 for DBSCAN and Graph clustering, whatever the run covers. The dialog box shows the number before you run. Density clustering stays local. Flow cytometry compute has the table and the arithmetic behind it.
Project members with view-only access can read saved results but can't run these tools.