clusterMAP – user guide
clusterMAP turns complexome-profiling data into heat maps, clusters similar protein profiles and helps identify co-migrating proteins – candidates for a complex, which still have to be judged by the biology. Your research data stays on your computer: loading, clustering, plotting and exporting happen entirely in your browser – nothing is uploaded unless you explicitly create a shared link.
§Overview
Everything happens in one window. On top there is the toolbar, below it the heat map, on the left the sets (the saved states of your analysis) and at the bottom the panels – profile viewer, mass scale, groups, validation, search – each of which looks at one thing.
Two things are worth knowing before you start:
- Nothing* is uploaded and nothing is computed on a server. The program runs in your browser. A network is needed only for the four things listed in Files and offline use – loading a data set from complexomemap.de, sharing, the load diagnostics and the update check – and three of those happen only when you ask for them.
- Every picture in this guide comes from the built-in demo sets, so you can follow along without a file of your own – most chapters use Simple complexome data, the peaks of Splitting profiles into peaks use Peak Clustering and the comparison of Comparing up to four runs uses Comparison.
→Quick start
- Open the program and choose Open → Load Demo → Simple complexome data. A heat map appears: 140 profiles over 48 fractions.
- Press Cluster with the default settings (AVERAGE linkage, Pearson distance, no normalization). The cluster tree appears on the left and the rows are reordered to match it.
- Look at one protein: click a name in the map. Its curve appears in the profile plot below.
- See what it is: search for a name or an accession (Ctrl+F), or point at a cell to read its value.
- Make it readable: choose a palette in Colours, decide about the Scale chip next to it (per profile or whole map), add a mass scale if the fractions have known masses, and use groups for the complexes you recognise.
- Keep the state: Set
→ New set from current view stores what you did, and
Ctrl+S saves the whole project as a
.cmapfile. - Get it out: the Export menu writes a coloured image or a table – SVG for a poster, Excel with the colours in the cells.
Afterwards replace the demo data with your own table: Open → Open File, and Advanced data import says what the reader expects from it.
Loading your data
Open your own data – or one of the built-in demos – from the Open menu in the toolbar:
| Option | Opens |
|---|---|
| Open File | your own CSV, TXT, Excel or OpenDocument table, or a saved clusterMAP project. |
| Open File for Comparison | another data set alongside the current project, with up to four maps in one comparison (Comparing up to four runs). Only shown while a project is open. |
| Import ComplexomeMap | a data set from complexomemap.de, by its address. |
| Load Demo | one of the three built-in demo data sets below. |
The three demo data sets
| Demo | Best for |
|---|---|
| Simple complexome data | following this guide and exploring the basic workflow · 140 profiles over 48 fractions |
| Comparison | trying a ready-made comparison of several runs · the simple set plus three runs of it |
| Peak Clustering | splitting multi-peak profiles and clustering the resulting peaks · 24 profiles over 36 fractions |
Advanced data import
Choose Open
→ Open File, or drop a file anywhere onto the
window. clusterMAP accepts delimited text files (CSV, TSV and TXT),
Excel and OpenDocument spreadsheets, clusterMAP projects (.cmap) and
NOVA projects (.npf).
If a project is already open, clusterMAP asks whether the new file should replace it or be added for comparison.
Dropping multiple files and replacing projects
A drop answers while you drag: the window gets a dashed frame and the sentence “Drop the file to open it”, and letting go runs exactly the reader Open File would have picked. Only files are taken – dragging text or a link onto the window changes nothing – and a drop is ignored while a dialog is open, so close the dialog first.
The program reads the dropped file first and then asks once, and the question is the useful one: Compare to project puts the dropped data next to the open project (Comparing up to four runs), Replace project opens it as the project and discards what was there, Stay in project changes nothing. Two of the three are only offered when they can do something: Replace needs a file that can be read, Compare needs a file that shares at least one profile with the project – otherwise the button is grey and says on hover why. Only the demo data – and only while no comparison is running – is replaced without a word, because one click brings it back.
Several files at once are taken as one gesture: the first readable file becomes the project, and the files that fit join it as further maps (Comparing up to four runs). A file that cannot join is left out, and the status line says in one line why. A project file that brings several sets asks which of them should take part – one question per file, because it cannot be answered for you.
How clusterMAP reads your table
| Input | How it is handled |
|---|---|
| Empty numeric cell | interpreted as 0 – “nothing measured in this fraction”, which is what an empty cell means in a profiling table. If an empty cell means “missing” rather than zero in your data, resolve it before importing the file. |
NA, NaN or text in a numeric column |
the import fails – the whole file is refused, and nothing
is half-read. The message names the first cell
(row 10, column 12) and how many others are like it: a file
with ten thousand of them is one sentence, not ten thousand. Replace the
cell with a number, or empty it. |
| Separator | comma, semicolon or tab – guessed from the file; the status line names the choice. |
| Decimal separator | . or , – guessed per file. Spreadsheets always count as
.
|
| Spreadsheets | only the first sheet is read – move the table you want to the front before saving, or export it as CSV. |
| Fraction columns | the block of numeric columns. clusterMAP finds it from the known label
columns (UNIQID, SHORTID, GROUP) and
the widest run of numbers in the first line. |
| Row label | the column named UNIQID – or, without one, the first column. A
header-less file takes the ID | NAME pair out of its last text
columns, and the status line says that it did.
Accessions inside the label are recognised and are what
the comparison matches on.
|
| Masses of the fractions | a row directly under the header line whose label reads
mass or mass (kDa) becomes the
mass scale of the file instead of a profile, and the unit
in brackets becomes the unit of the axis. That is the row
add the masses of the fractions as a row writes, so
a table exported here comes back with its masses; the
status line names what was read.
|
| Negative values | accepted and kept – they are not the same as “nothing found”, and are never silently dropped. |
| Duplicate protein IDs or fraction names | kept as they are: nothing is merged, renamed or dropped. Duplicate profiles are reported by the comparison, which cannot tell them apart. |
| Rows with a different number of cells | refused, with the number of the row. |
The sample file contains eight mitochondrial proteins across twelve fractions. It uses a tab-separated format with a header row:
UNIQIDbecomes the row label.GROUPassigns the coloured group strip.- The numbered columns become fractions.
A header row is strongly recommended because it makes the import unambiguous. The file in full:
| UNIQID | GROUP | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Q9LF11 | NADH dehydrogenase 1 alpha subunit 9 | Complex I | 2.5 | 4.1 | 9.8 | 18.4 | 31.2 | 44.7 | 52.1 | 47.3 | 33.8 | 19.2 | 8.4 | 3.1 |
| Q9SJQ4 | NADH dehydrogenase 1 beta subunit 4 | Complex I | 1.1 | 2.0 | 3.4 | 6.9 | 14.2 | 27.5 | 39.8 | 51.2 | 58.4 | 52.7 | 38.1 | 19.4 |
| P93231 | Cytochrome c oxidase subunit 5b-1 | Complex IV | 0.4 | 0.9 | 2.2 | 5.1 | 9.8 | 16.3 | 25.9 | 38.4 | 49.1 | 56.2 | 48.7 | 29.3 |
| Q9ZW33 | Cytochrome c oxidase subunit 6b-2 | Complex IV | 0.2 | 0.6 | 1.4 | 3.2 | 7.7 | 15.1 | 26.4 | 39.2 | 47.5 | 44.1 | 30.6 | 14.8 |
| Q9ZR72 | ATP synthase subunit O, mitochondrial | Complex V | 0.1 | 0.3 | 0.9 | 2.4 | 5.6 | 11.8 | 22.7 | 36.9 | 45.2 | 41.8 | 26.4 | 11.2 |
| Q9SJ12 | ATP synthase subunit d, mitochondrial | Complex V | 0.3 | 0.8 | 2.1 | 4.8 | 9.4 | 17.2 | 28.6 | 41.3 | 50.7 | 47.6 | 34.2 | 16.9 |
| Q9M1Q2 | Inner membrane translocase TIM13 | TIM complex | 0.6 | 1.3 | 2.8 | 5.9 | 11.4 | 20.1 | 31.7 | 42.8 | 44.3 | 36.1 | 22.9 | 10.4 |
| P46274 | Voltage-dependent anion channel protein 1 | Outer membrane | 0.2 | 0.5 | 1.2 | 2.9 | 6.4 | 12.9 | 23.8 | 37.1 | 48.6 | 51.4 | 37.7 | 18.1 |
Download the
sample file (samples/example-profiles.tsv, 1 kB) and open it in the
program.
UNIQID supplies the row labels, and
GROUP supplies the coloured group strip.The heat map
The heat map is clusterMAP's main workspace. Each row is a protein profile, each column is a fraction, and each cell shows the abundance of that protein in that fraction. The profile names stay fixed on the left while the fractions scroll sideways.
Reading the heat map
| Area | What it shows | Interaction |
|---|---|---|
| Above the cells | fraction numbers and, if a mass scale is available, the corresponding molecular masses. | hover over a column to highlight its fraction and its mass. |
| Left of the cells | the group colour and the profile name. A long name is shortened in the picture; the full name waits in a tooltip. | use the pencil to rename a profile. A double click on a recognised accession opens its database entry, any other name becomes a search. |
| Inside a cell | the abundance of one profile in one fraction. | hover to see the profile, the fraction and the value – twice, if a normalization was used for the clustering: once as stored and once as the clustering saw it. |
Opening external links. A double click on an AGI code
(AT1G77120.1) opens TAIR, a double click on a UniProt
accession opens UniProt, and any other profile name is used as a
search term. The same actions are in the right-click menu of a name.
Colours and scaling
Three separate controls determine how the heat map is displayed: the Colours button chooses the palette, Scaling determines how values are mapped onto that palette, and Reverse (⇄) changes the order of the fraction columns.
Colours
Colours changes the palette used for the heat map. It does not change the values or their scaling.
All palettes but one run from low to high values. The lowest value of the range is painted with the first colour of the palette and the highest with the last – there is no tone set aside for “nothing”. A band of its own for “not detected” is a decision about your data, so it belongs in the palette: Edit colour scheme… lets you put a stop at 0. Viridis (perceptual) and Magma (perceptual) are perceptually uniform, which makes them the safest choices for quantitative figures and generally more robust for colour-vision deficiencies. Greyscale (print) is meant for black-and-white output. The other palettes are familiar or stylistic alternatives.
The exception is Ice & Fire (diverging), and it is the one for tables whose values can be negative: a log-ratio, a z-score, a background-corrected intensity, a column that was normalised around the mean. Its middle stands on zero – cold blue to the left, warm red to the right – so “no change” is one colour and equal amounts above and below it are mirror images. With non-negative data it wastes half of itself, which is why it is not the default: choose it when zero means something. The same rule can be given to your own scheme with zero in the middle in the editor.
Custom colour schemes and sharing (advanced)
Edit colour scheme…, the last entry of the palette list, opens the editor. The preview is live – the editor is a panel, not a window, so the map behind it shows what you are doing.
| Action | Result |
|---|---|
| drag a colour stop | changes its position. |
| click the gradient bar | adds a new stop. |
| enter a hex value | sets an exact colour. |
| zero in the middle | turns the palette into a diverging one: the middle stop stands for 0, the left half carries the negative values and the right half the positive ones. Set it when your data can be negative – without it, the light middle of a diverging palette would sit wherever the smallest value happens to be. |
| + Add colour | adds a stop to the list. |
| ⇄ Reverse | reverses the complete palette. |
| Apply | uses the scheme in the current project. |
| Save (in browser) | adds it to the palette list in this browser. |
| Delete scheme | removes a scheme that was saved in this browser. |
| copy the export/import string | creates an exact, shareable representation of the palette. |
| paste a valid string | replaces the editor contents with that palette. |
| Esc (or the ×) | closes the editor without applying the changes. |
[0:#103468] [0.4:#488cc2] [1:#f7fbff] is a palette with three
anchors, the middle one at 40 % of the scale. A line that cannot be read is never
half applied: the field turns red and says which entry it did not understand.
For reproducible documentation, copy the export/import string and store it with the analysis settings – it is the palette, exact and complete.
Scaling
Scaling determines which values share the same colour range. This affects what can – and cannot – be compared visually.
| Mode | Colour range | Best used for |
|---|---|---|
| Per profile · shared across maps | one range for each profile, shared across all comparison maps. | comparing the shape and position of the same profile between maps. |
| Per profile · separate per map | a separate range for every profile in every map. | maximum contrast when inspecting patterns within individual maps. |
| Whole map · shared across maps | one range for all profiles across all maps. | comparing displayed magnitudes across profiles and maps, provided the input values are comparable. |
| Whole map · separate per map | one range for all profiles within each map. | comparing profiles within a map, but not magnitudes between maps. |
The two separate per map modes exist only while a comparison is open – a single map has nothing to share with.
Hover over a cell to see its numeric value. The Scale chip shows the active scaling mode.
Reverse fraction order
Reverse changes only the displayed order of the fraction columns. Use it when a gel or gradient was processed in the opposite direction.
Selecting profiles and fractions
Most actions in clusterMAP apply to the current selection: to profiles (the rows of the map) or to fractions (its columns). Profiles and fractions are selected independently: selecting a profile does not change the selected fractions, and vice versa.
| Target | Select one | Select a range | Add or remove one |
|---|---|---|---|
| Profile | click its name or any of its cells. | Shift + click another profile. | Ctrl + click a profile. |
| Fraction | click its column header. | Shift + click another fraction. | Ctrl + click a fraction. |
Press Esc to clear both selections.
This is useful for fractions that contain little useful separation, such as the ends of a gradient (Hiding profiles and isolating groups).
Selection shortcuts
| Shortcut | Action |
|---|---|
| Ctrl+A | select all visible items on the last-used axis. |
| Ctrl+I | invert the selection on the last-used axis. |
| Ctrl+Shift+A | select all profiles in the group. |
| Esc | clear both selections. |
| Ctrl+D | disable the selected profiles or fractions. |
| Ctrl+E | enable them again. |
The last-used axis is the one the last click worked on: a click on a profile name or a cell makes it the profiles, a click on a column header makes it the fractions. The status line names it beside the two counts – 3 profile(s) selected · 4 fraction(s) selected · active: profiles – so nobody has to remember which axis the keys will take.
Keyboard navigation
| Shortcut | Action |
|---|---|
| ↑ / ↓ | move the profile selection by one row. |
| Shift+↑ / Shift+↓ | extend or reduce the selection. |
| Alt+↑ / Alt+↓ | move the selected rows. |
Renaming and moving profiles
Use the pencil beside a profile name to rename it. The name is also the profile's identifier throughout the project and in comparisons, so renaming is unavailable while a comparison is open.
Drag a profile name to reorder its row. During a comparison, the row moves in every map.
The right-click menu
Right-click a profile, fraction or group row to open the actions available for it and the current selection.
- Profile actions: disable it, open its identifier in a database or start a search for it, move it to a group, or add it as a reference profile.
- Selection actions: select all, invert the selection, or create a set from the selection.
- Fraction actions: disable the fraction, select fractions, or define its molecular mass (Mass scale).
- Group actions: select, deselect or show only the profiles in that group (Profile groups).
move to group and the group entry open a submenu – the › on their right says so. While the menu is open, an underlined letter runs an entry, and the groups behind move to group carry digits instead: 1, 2, 3 … 9, 0, with new group… taking the next one.
Search
Open 🔍 Search above the name column and enter part of a protein name, accession or group. Selecting a result moves to that profile in the heat map.
Search actions can add all results to the current selection, create a set containing them or assign them to an existing or new group.
Search is case-insensitive and matches substrings. Multiple terms are
combined with AND. Use quotation marks for a phrase, ! to
exclude a term,
id: to search only accessions and is: to search profile state.
| What one types | What it finds |
|---|---|
nadh subunit |
profiles containing both terms, such as Q9LF11 |
nadh !beta |
NADH profiles that do not contain beta; in this example, the alpha
subunit remains |
"complex I" |
the exact phrase, including the space. Without quotation marks,
complex I is treated as two substring terms and may also match
Complex III or Complex IV
|
id:Q9LF11 |
accession Q9LF11 only; matching text in the protein name or group is
ignored |
is:disabled |
profiles currently disabled in the analysis |
is:selected |
profiles currently selected; !is:selected finds all others |
complex !ii, excluding profiles that also contain
ii.Profile viewer
Below the heat map the selected profiles are drawn as curves over the fractions. Clicking a profile name in the map is usually the fastest way to look at it.
Reference profiles
Add as reference pins the current selection to the plot: the reference curve stays there while you click through other profiles, so shapes can be compared directly. Member(s) in the info line counts what is pinned, and Clear reference empties the list. That list belongs to the project – save it and it is still there tomorrow.
The options in the toolbar
- Show mass puts the mass on the x axis instead of the fraction number (grey while no masses are known – the mass scale panel is where they come from).
- Legend names the plotted profiles in the plot.
- Relative scales every curve between its own smallest and largest value, so shapes can be compared even when the amounts differ wildly.
- Line dots marks every measured value with a dot.
- The button at the far right, , lets the plot use the whole window width; the heat map keeps its own size.
Analysis sets
A set is a state of the analysis, not a copy of the data: it remembers which clustering was run on which profiles, the row order, the settings and which profiles are disabled. Sets are shown in the left column, and a set made from a set is indented under it.
| Entry | Use it when |
|---|---|
| New set from current view | you want to keep what you see – the clustering, the order and the profiles that are disabled. |
| New set from selection | you want exactly the profiles you selected, and nothing else. |
| New set from IDs… | you have a list of accessions or names from somewhere else. |
| Split the selection into components… | you marked profiles and want to see where their fitted parts would sit in this map. It opens the Split peaks panel – settings and preview at hand – where Insert into this map does the fit. A crutch for the eye, not a result (the plain-words reason). |
| Remove the inserted components | you want exactly those rows out of the map again (nothing else moves). |
The right-click menu of a set card carries two entries: rename analysis and open as a new project in a new window. The second one hands the set to a second window of this browser, where it is a project of its own - with only the profiles and fractions that set has enabled, as a data set of its own. Labels, accessions, masses and groups travel along, nothing is lost (the window on screen keeps everything), and a set that was never clustered arrives unclustered, so the Cluster button is one click away. In a peak set the entry is greyed out: there the rows are fitted peaks, and the run they came from stays with the project.
Splitting a selection into components – the crutch
Mark profiles in the heat map (click their names, or Ctrl+A) and use Split the selection into components… in the Set menu – or the same entry in the right-click menu of the heat map. It opens the Split peaks panel, where you can work on the settings and watch the preview of the selected profile: the marks are the detected maxima, the solid curves the peaks that are kept. Then the action Insert into this map fits the marked profile(s) and inserts the components into this map as rows, directly behind their profile. Labels, accessions, groups and colours are inherited, so a component stays in the same band as its protein, and its name carries the same peak suffix a peak set uses – pk-3porin P1/3 F22.5 – which the name column sets off in bold and red.
This is not a result, it is a crutch. The inserted rows are the fitted Gaussian curves – model values, not measurements – and they now sit in the same matrix as the measured profiles: the same signal is counted twice, and a distance between a component and a profile says as much about the fit as about the gel. Nothing there says whether a band really was two. The panel says the same thing directly above that action, and the FAQ answers it in plain words. Use it to see where the parts would sit, then decide with your own eyes; for anything you want to show, use Create peak set instead – there every row is a fit, and the distance is position and width. There is no second dialog afterwards: what the run did is in the status line, and Remove the inserted components takes exactly those rows out again.
Because the rows of the data set change: stored clusterings are dropped (the settings stay, and Cluster computes them again – now with the components), and an open comparison is closed.
Select a set to switch to it, or rename the active set from its card. Sets are saved with the project and included in shared links.
Profile groups
Groups give the profiles a colour and a name (usually a complex): complex I, respirasome, unassigned. The colour is drawn as a band behind the profile name in the heat map, and the group name is part of the row label – which also means that a search finds it.
- Create and edit groups in the Groups panel, then assign selected profiles from either the profile or group context menu. The group menu also provides actions to delete a group, isolate it or select all of its members.
- The list at the end of a row moves a whole group: pick a target and every profile of the group goes there, the group itself disappears. The default group keeps its name and cannot be dropped, because unassigned profiles belong to it.
- Import and Export read and write a
small table with the columns
id,groupandcolor– the same file can go back and forth between sessions.
Hiding profiles and isolating groups
Sometimes a profile destroys the picture – an outlier, a contaminant, a protein that was only found in one fraction. Instead of deleting it, disable it – switched off, but still in the data: it is dimmed and left out of the next clustering run.
A fraction can be disabled the same way, and that is what a gel front needs: right-click its column header and choose disable this fraction (Ctrl+D does the same for the marked lanes). A dimmed lane stays in the picture and in the data set, the distance matrix of the next run simply does not see it – so the clustering no longer has to explain both ends of the gradient. Ctrl+E enables it again.
A whole group can be hidden the same way: the groups panel has the entry show only this group. That is a view filter, not a change of the data – no profile is disabled, and Ctrl+Shift+H brings the rest back.
| Shortcut | What it does |
|---|---|
| Ctrl+D | disable the marked profiles or fractions |
| Ctrl+E | enable them again |
| Ctrl+H | show or hide the disabled profiles |
| Ctrl+Shift+H | leave the group isolation again |
| Ctrl+Shift+D | make a new set out of the selection |
Clustering
Clustering sorts the proteins so that similar profiles end up next to each other, and draws the result as a tree next to the heat map. Everything it needs is in the menu behind Cluster in the toolbar.
The bar holds a second way in, right before Cluster: Split peaks does not cluster the measured profile of a protein but its peaks: a protein that migrates in two populations becomes two objects, and each of them can be clustered with the peaks of other proteins. That is Splitting profiles into peaks, the next one – the rest of this chapter is about clustering whole profiles, which is what the program does by default.
Quick start
Open Cluster, choose the settings and run the clustering. The rows are reordered and a tree appears beside the heat map; the status line records the linkage, distance and duration. For a defensible starting point, choose ↺ Reset to default.
Recommended starting point
For the profiles of a complexome profiling experiment: Pearson Correlation Distance with AVERAGE linkage and no normalization. That is exactly what ↺ Reset to default restores, and it is what the program starts with.
- Pearson compares the shape of two profiles around their own mean. A faint and a strong version of the same profile belong together – and that is what these profiles are about: where a protein migrates, not how much of it was measured.
- AVERAGE (UPGMA) is the balanced standard: SINGLE chains and COMPLETE cuts real groups apart, and both need an argument. The mean over all pairs needs none.
- No normalization is not laziness here: a Pearson clustering comes out unchanged under all four normalizations, so none of them could change this tree. "None" states that instead of hiding it in a setting.
- It is the combination the original NOVA used and the common choice in complexome profiling, so your result stays comparable to what others show.
A starting point, not a law: as soon as the amount of a protein is part of the question, the distance has to change – the tables below say which one then.
How the settings work together
The four settings are not independent. They are applied in a fixed order, and each one works on what the previous one left behind:
- normalization rewrites every profile – a scaling per profile, nothing else. It is used for this run, not for the picture.
- distance function measures how far two profiles are apart, on the normalized values. The distance is therefore chosen for the numbers the normalization left behind, not for the ones in the cells.
- linkage defines how the distance between two clusters is measured once profiles start merging – this is the merge order itself.
- optimize leaf order turns the finished tree so that similar branches sit next to each other. It changes the drawing, not the clustering.
Two details follow from that order. A profile or fraction you disabled is not part of the values at all: the normalization never sees it and no distance is measured over it either (Selecting profiles and fractions). And the heat map keeps showing the imported values – the normalization lives in the clustering, and the cell tooltip shows both.
Which settings should I use?
Change one setting at a time and look at what it does to the tree – with the tables below at hand:
- Nothing in particular? Stay with the recommended start: Pearson, AVERAGE, no normalization.
- The amount matters? Euclidean or Manhattan Distance, and then a normalization – Maximum or Area – because with these distances the size of a profile is part of the distance.
- Compact groups of similar size? WARD with Squared Euclidean Distance, and with nothing else – see the note under the table.
- How much does a single protein bridge two groups? SINGLE: the chaining it is famous for is exactly what makes the bridges visible.
Linkage methods
The linkage decides what "the distance between two clusters" means once more than one profile sits in a cluster. It is the setting that changes the picture most.
| Linkage | How it measures a cluster | Useful when |
|---|---|---|
| SINGLE | the closest pair of members (nearest neighbour) | you want to see the bridges: one similar pair is enough to join two clusters, so long chains and one large group are the normal outcome. |
| COMPLETE | the farthest pair of members (farthest neighbour) | you want compact clusters of similar size: every member has to be close to every other one – which can also cut a real group apart. |
| AVERAGE | the mean over all pairs (UPGMA) | you want the balanced answer – the recommended start, and the default of the program. |
| CENTROID | the distance between the two cluster centres | you want clusters of natural size. Beware of inversions: a merge can sit below a later one, and then the tree looks folded. |
| WARD | the pair whose union costs the least variance | you want compact groups of similar size – and only together with Squared Euclidean Distance. |
Distance functions
The distance says what "similar" means for two profiles. It is measured on the values the normalization left behind, so it decides which differences count at all.
| Distance | What it compares | Useful when |
|---|---|---|
| Pearson Correlation Distance | the shape of two profiles around their own mean. Insensitive to the overall size: multiplying one profile by any positive factor leaves the distance unchanged. | the migration behaviour is the question – the standard for complexome profiles, and the recommended start. |
| Pearson Correlation Distance (uncentered) | the same without subtracting the mean – the angle between the two profiles, so the zero point of the scale counts with it. | the zero of the scale means something: a fraction without signal should count as "nothing there", not as a shape of its own. |
| Euclidean Distance | the straight-line difference, value by value. The size of a profile is part of the distance, so a faint and a strong profile can be far apart although their shapes agree. | the amount of a protein matters, or the profiles are already on one scale because of how the experiment was made. |
| Squared Euclidean Distance | the same, squared – large differences weigh more than proportionally. This is the criterion WARD minimises. | you want WARD: the variance criterion exists in no other distance. |
| Manhattan Distance | the sum of the differences, value by value – less sensitive to a single extreme fraction than Euclidean. | a single outlier – one fraction with a spike – must not dominate the distance. |
Normalization
A normalization scales every profile on its own before the distances are measured: the same numbers, brought onto a comparable scale – no logarithm, no quantile, no imputation.
| Normalization | What it does to one profile | Useful when |
|---|---|---|
| None | nothing is changed | the profiles are comparable as they are – the recommended start with Pearson, where the normalization cannot change the tree anyway. |
| Range Normalization | the minimum becomes 0 and the maximum 1, per profile | every profile should be on the same range, no matter how strong it is. |
| Maximum Normalization | divides by the largest absolute value: that value becomes 1 (or −1) | one strong peak dominates a profile and should not decide the distance. |
| Area Normalization | divides by the sum of the absolute values: the area becomes 1 | the total amount differs – loading, sample size – and only the distribution over the fractions should count. |
| Unit Vector Normalization | divides by the Euclidean length: the length becomes 1 | only the direction should count – the closest thing to Pearson without subtracting the mean. |
- With a Pearson Correlation Distance a normalization changes nothing at all: the correlation is invariant under exactly these transformations, so the same tree comes out with and without it. Do not spend time comparing the four options here – compare the distances instead.
- With Euclidean, Squared Euclidean or Manhattan Distance the same normalizations change the tree completely, because the size of a profile is part of the distance. There, Maximum or Area Normalization is the usual way to compare profiles whose overall amount differs for reasons that have nothing to do with the complex.
- Two edges are worth knowing. An all-zero profile stays all-zero under every normalization – nothing is divided by zero, and no NaN can get into the matrix – and such a profile is left out of the clustering, which the status line reports as "empty profile(s) filtered". And "sum" always means the sum of the absolute values: with negative values a profile does not end up summing to 1.
The settings belong to the next run; a set that is already clustered keeps the settings it was made with.
Tree display
After the run the tree stands left of the map, and the two share the row order – a branch always points at the profiles it contains:
One heading of the right-click menu on the tree is worth knowing: setting a minimum branch length (flat branches). In a data set with many tiny differences the tree collapses at the right; this stretches the distance axis so that flat branches become visible. It changes the drawing only – the tooltip of the heading says so, and the clustering stays as it is.
How wide a branch is drawn is the accumulated distance from the root, so the horizontal scale of the picture follows the distance you chose: with Squared Euclidean Distance that depth is a sum of squared values, not of lengths.
Splitting profiles into peaks
One protein does not always appear in one place. A profile can carry several peaks – two populations of the same protein that migrate differently, for instance one assembled into a complex and one free. Split peaks takes such a profile apart: the program fits the peaks of every profile, keeps the ones that pass its quality checks and makes each of them an object of its own – able to cluster with the peaks of other proteins, while the link to the measured profile it came from is never lost.
Why split profiles?
Clustering treats the complete profile as one vector: the whole curve over all fractions, one row, one entry in the distance matrix. Its peaks all contribute to its distance to the other proteins – but they cannot be assigned to different branches independently. A protein with two populations therefore stands in one row, and the resulting position in the tree reflects the profile as a whole.
Two-minute version
The Peak Clustering demo (Loading your data: Open → Load Demo → Peak Clustering) is 24 profiles × 36 fractions: four populations of five partners around the fractions 6.5, 11.5, 22.5 and 27.5, and four proteins that carry a band beside their own population – one of them with three peaks. Three steps:
- Open Split peaks in the toolbar – the entry of its own, right before Cluster – and press Create peak set. The run fits every profile, applies the quality gate, derives the peak rows and clusters them with AVERAGE on peak position and width; the status line reports that in one sentence (“… clustered with AVERAGE on peak position and width”). While the run works, the button itself fills from the left and its second line counts the profiles it has done.
- Review the report that opens: 29 peaks from 24 proteins, the five numbers of the run, and the profiles that need attention.
- Press Open here to work with the set - the data set the peaks were fitted in then steps aside. It is already clustered and brings its tree: there is no unclustered peak map in between and no second Cluster click. If the report shows something you do not like, simply close it (× or Escape): the map stays as it is, and the run can be repeated with other settings.
The tree then groups the bands, and every multi-band protein stands in several of them – one protein, several populations, several sets of partners. That is what the whole feature is about, and it is visible in three clicks.
Detection and fitting settings
Split peaks previews the selected profile and reports how many peaks the current settings would produce. Create peak set is available when at least one detected peak passes the quality gate; otherwise the panel states why no peak set can be created.
The preview answers the question the numbers alone cannot: what do they mean here? It shows the measured curve of the selected profile, the maxima these very settings find (the orange triangles, each with a line down to the axis), and the fit that follows from them – components that pass the checks as solid blue curves, the ones that do not as dashed red ones. All of it shares one scale, because the parts are read against their sum. Under the picture stands what it found, in the two halves that matter: how many maxima were detected and how many peaks were kept.
| Setting | What it decides | Reach for it when |
|---|---|---|
| prominence % | how far a maximum has to stand out from the higher of its two saddles, relative to the largest value of the profile. | peaks are lost or found. It is not the height: a tall peak on the shoulder of a taller one does not stand out, a small one on a flat run does. |
| min distance | the closest two peak centres may be, in fractions. | two maxima closer than this become one peak – the more prominent one wins. |
| max peaks | at most so many peaks per profile; the fit compares models with fewer peaks against models with more and takes the one the data supports (never simply the one with the highest R²). | a profile is split into too many small components. |
| baseline | whether one constant is fitted under the components – it changes the model, not the search. | the profile sits on a background that is not zero. |
| smoothing | a window for the search only (the fit always works on the measured values), in measured fractions. It is centred, so only odd lengths describe one – the field counts in steps of two, and 1 means no smoothing at all; a number typed anyway is taken as the window it really averages (an even 4 becomes 5). In the lab book the un-smoothed search is written as smoothing 1 (off). | noise produces maxima that are not peaks. |
| σ min | the narrowest width a component may have, in fractions (FWHM ≈ 2.35 × σ). | the fit puts narrow spikes onto noise instead of one broad band – the flag width at bound in the report says so. |
| σ max | the widest width a component may have, in fractions. | a broad band is cut into several components because none of them may be wide enough. |
| min points | how many measured fractions within the peak's FWHM must support a component. | a peak carried by one or two fractions should not become a row. |
Reviewing the quality report
Create peak set fits every enabled profile, applies the quality gate and clusters the accepted peaks with AVERAGE on position and width. Progress is shown in the status line and the run can be stopped from the same button. When it finishes, the quality report opens and the resulting set records the clustering method.
The five numbers mean five different things and are deliberately not mixed:
| Number | What it counts |
|---|---|
| Profiles analysed | the enabled profiles that were handed to the fit – a profile that is disabled is not among them (Selecting profiles and fractions). |
| Profiles fitted | the ones the fit reached a solution for. The difference between the first two numbers is what could not be fitted at all. |
| Peaks kept | the accepted components – these are the rows of the peak set. |
| Components rejected | components that were fitted but did not pass the quality gate. They are fitted and paid for, but they are not part of the set. |
| Profiles not fitted | profiles without a single usable component (a constant profile, for instance). The reason stands with them in the table below. |
A percentage has no place here – “fitted” and “kept” are not the same thing, and one number cannot say both.
The table of Needs attention is always open – it is the part of the report a reader came for. Every profile the run doubts or could not fit stands there as one row, and the first column says which of the two it is:
- Review (small yellow mark) – usable, but with a reservation: either the run itself doubts the profile (review), or one of its components carries a flag. The Result column then says how many of its components were rejected, and the column Reason / flags names them – here pk-spike · 1 component rejected · R² 0.9999 · few points · width at bound · isolated.
- Not fitted (small red mark) – nothing of the profile could be used, No peaks created, with the reason the run wrote down beside it: pk-flat · constant profile.
In Needs attention, View fit opens the measured profile and its fitted components; rejected components are dashed. If no component could be fitted, View profile opens the original profile instead. Clean profiles and their retained peaks and R² values are listed in the collapsed section below.
The Quality flags section lists each flag, the affected profiles and whether a component was kept or rejected. A flag describes one fitted component, whereas R² describes the fit of the complete profile; other peaks from the same profile may still be retained. The ? beside the heading explains this distinction.
Sections with no entries are omitted. In the clean Peak Clustering demo, all 24 profiles are fitted and no flag is raised, so Needs attention reports that every profile passed and the flags section is absent. The report can be reopened from Export → Peaks → Peak run report….
| Flag | What it says about a peak |
|---|---|
| edge | its centre sits at the edge of the analysed range – the peak may continue outside it, so its height and area are lower bounds. |
| truncated section | the peak reaches over the border of a section (a disabled fraction in the middle), so part of it was not analysed. |
| few points | fewer measured points support it than the setting asks for – a peak carried by one or two fractions. |
| width at bound | its width sits on the limit of the allowed range: the fit ran out of room and the number is not a measurement of the peak. |
| low prominence | the maximum it started from did not stand out enough against the fit's own uncertainty. |
| overlapping | another component of the same fit lies as close and is as wide – the two cannot be told apart. |
| isolated | only one measured fraction carries it. |
| not converged | the fit stopped at its iteration cap – the numbers are the last state, not a solution. |
Working with a peak set
accession P2/3 F12.4 - the protein, this peak,
how many peaks that profile was split into and its centre in fractions. A profile that
was fitted as one peak needs no number: its row is called by the protein
alone.- The set is clustered with AVERAGE on position and
width only:
d = √(Δcentre² + Δsigma²). The height of a peak – how much protein was there – is reported (area and share) but it does not group anything, so a faint and a strong band that migrate together end up as neighbours. The method stands in the status line after the run and in the entry of the set in Sets (AVERAGE · 29 peaks). - A second measure is offered above the ordinary distances as soon as the data set has a mass scale: Peak mass + relative width (log) reads the same peaks relative on that scale, so a band at 50 kDa and one at 55 kDa are as close as bands at 500 and 550 kDa. It needs the calibration of the fractions (the masses travel with the peak set), it never invents a mass beyond the calibrated fractions, and it says so instead of clustering: a peak outside the scale makes the run stop with the number of peaks concerned. The calibration can be done before or after the peaks are fitted: completing or correcting the mass scale of a peak set reaches its peaks, and the peaks themselves – centre, width, area, fit quality – are not fitted again.
- Each peak stays an object of its own. Two peaks of the same protein are free to land next to each other, which is exactly what “this protein has two populations” looks like in the tree.
- Everything from the other chapters works: the map, the tree, the profile viewer, the search (searching an accession finds all peaks of that protein), the sets, the export. Clustering this set again with one of the standard distances described in Clustering (a correlation, say) is a comparison of two methods – it is not evidence, and the neighbours it produces mean something different.
A peak set is a project, so Saving and exporting and Sharing a project apply. What is special about it:
| Entry | What you get |
|---|---|
| Download project | the set as one .cmap file. It carries the whole run – settings, flags
and the measured points of every profile – so the details window and the report
still work after reopening it. |
| Export ▾ → Peak table (CSV) | one line per peak (centre, width, FWHM, height, area, share, the R² of its profile, its flags) and a second file with the profiles that could not be fitted. |
| Export ▾ → Peak table (XLSX) | the same as a workbook: one sheet with the peaks, one with the unfitted profiles, and every cell carrying its kind – numbers are numbers there, which is what a spreadsheet needs. |
| NOVA project (.npf) | greyed out in a peak set: the .npf format of the original NOVA cannot
express peaks (the tooltip says so). Save a peak set as .cmap. |
| Validation panel | keeps working on a peak set and says in front of its numbers what they are about: the peak matrix, not the proteins it was fitted from. |
The special cases: second tab, sharing, two peak sets side by side
- The second tab is the ordinary application with the peak set as its project, so report and set can be read side by side – and saved as a file or shared from there like any other project. If the browser blocks the window, nothing is left behind: the entry prepared for it is taken back and the status line says what happened.
- Sharing is the Share menu, where a link is what one asks for – the same as for any other project, with the same reminder that a link carries your data to the server (Sharing a project).
- Two peak sets can be put side by side: the maps then line up peaks – the same protein and a mass window around the centre, assigned one to one by the smallest total mass distance, never by accession (Comparing up to four runs). The status line says how many pairs were made, how many of them are ambiguous and how many peaks stayed unmatched in each map, and the tooltip names the four reasons behind them. Peaks that only one run has stay in the picture as empty rows on the other side.
- A peak set beside a set of profiles is refused with the reason instead of being matched wrongly.
Inspecting one fitted peak
Right-click a row of a peak set and choose peak details… (the key k does the same) to see the fit a single peak came from:
Below the picture stand the numbers of the peak (centre, width as sigma and FWHM, height, area and its share of the profile), its flags, and the quality of the profile it belongs to – including its R². That R² is not the peak's: R² belongs to a whole fit, and a peak inside a badly fitted profile is not thereby a bad peak. A peak the fit itself does not trust is still shown: it carries flags instead of being hidden.
With a mass scale (Mass scale) one more line stands between them: the
mass of the centre and the two bounds of the width, each mapped through
the calibration itself – mass 123.0 kDa · FWHM 102 – 148 kDa (factor 1.45).
A peak is symmetric in fractions, not in mass, so the two bounds are not a symmetric
“± kDa” around the centre; they are shown low to high, also when the mass axis falls.
Without a calibration the line is absent rather than empty. Every finding keeps a line of
its own, so nothing stands outside the window.
Mass scale
Masses turn the fraction numbers into a biochemical axis: the position a protein runs at
becomes an estimated native mass. They can be typed in per fraction, or
fitted from a few points; the mass scale panel is the table
behind it. A number may be written with a point or a comma – 0.4 and
0,4 are the same mass, in this panel and in every other number field of the
application. A fraction whose name is too long for the narrow column in front of its
field is shown by its number there – the same way the column headers of
the heat map do it, and the full name is in the tooltip of the cell.
kDa, and the field is there to be changed when the gels were run in
something else.
Every cell in the grid is an input field, and it shows that as soon as the pointer is over it - the pale cell turns white, the field inside it gets its outline, so it is clear where a number can be typed before the click. Clicking into it puts the cursor there and adds the green edge; the estimated masses stand in grey italic as placeholders, a mass you typed stands in black and bold.
Right-click a fraction header and choose define mass to enter a marker mass for that fraction. For a selected fraction, the same command is also available from a cell in the heat map. Selecting a header without opening its menu marks the fraction instead.
Marker masses and a calibration model
Enter known marker masses in the corresponding fraction fields. These measured values remain fixed, while masses calculated from the calibration appear in grey italics. Entering a value for an estimated fraction turns it into a measured marker.
The calibration model says how the estimated masses are computed from those markers, and Fit calibration fits it and writes the estimates into every fraction without a marker mass - the same numbers the cells preview, and with them the ruler above the heat map, the profile plot and every export. A mass you typed in is never overwritten: an estimate fills a gap, it does not correct a measurement.
- Log-linear (recommended) - the default. Migration in a gel
is approximately linear in the logarithm of the mass, so this fits
log(mass)against the fraction position. With exactly two markers the line through both of them is the calibration; with three or more it is the best line through all of them, and then the status line also reports R². - Local log-linear - a straight line between each pair of neighbouring markers, so every marker is hit exactly. Useful when migration deviates locally from one single straight calibration line.
- Monotonic spline - a smooth curve through all markers that never overshoots (a monotone cubic, not a plain spline). It needs at least four markers; before that the entry says so and cannot be picked.
- Advanced - Linear mass is the historical fit of the mass itself (not the default, and rarely the right model for a gel); None means the markers are the calibration and nothing is estimated.
The model is never switched behind your back: if the markers do not suit the one you picked, the panel says what is missing instead of quietly choosing another.
Taking the masses to the next experiment
Copy masses puts the masses you typed in on the
clipboard, Paste masses reads them back - one line per
fraction, column<TAB>mass:
# clusterMAP marker masses (unit kDa) # column<TAB>mass - the column is the 1-based position in the fraction order 5 120.5 33 780
The line addresses a fraction by its column number, never by its label. That is what makes the list travel: a gel has its standard in column 5 and a band of interest in column 33 run after run, while the labels - and the numbers in the cells - change with whatever quantified them (raw intensities here, iBAF or ratios there). Estimates produced by the calibration are not copied: they belong to that calibration, not to the lane. Pasting sets the masses for the columns the text names (a fraction it does not mention keeps its mass), makes them markers, and reports what it could not use: unknown separators, a decimal comma, columns outside the data set and lines that are not numbers. Tab, semicolon and spaces all separate the two numbers; a comma is read as a separator when the line also contains a dot (the way a spreadsheet hands a table over), otherwise as the decimal mark. The unit in the header is reported when it differs from the project's - the numbers themselves are taken as they are. Both buttons confirm their work the same way the ƒx does: the button flashes green for a second and a message at the bottom of the window says how many masses went out or came in.
If a window with a field opens instead: the browser would not hand the clipboard over. Reading it needs a secure address (https), a permission the browser asks for once per site, and a window that is in front - Chrome and Firefox refuse the clipboard for a page in a background tab. The window says which of the three it was, and takes the text by hand: paste it into the field, press Take over.
The result box is the short answer - it fitted, how well, on how many markers. The exact equation is detail information behind ƒx (hover: the equation and the straight line it was fitted from, which is where R² is measured; click: the equation on the clipboard - the box flashes green for a second and a message at the bottom names what was copied), and the calibration graph carries it in its footer as well. The line under the table says the rest: how many fractions are estimated, and that these are apparent native masses derived from the calibration – not measured molecular weights.
A marker mass is a value in the data set, so it is saved and shared with the project: a link shows the same axis. The model travels with it as well - except in NOVA's own file format, which knows only one global fit; an export says so when the model cannot be expressed there.
The calibration can be completed later. If the data set in front of you is a peak set (Splitting profiles into peaks), the fractions there are the ones the peaks were fitted on, and a mass scale you fill in, correct or remove reaches the peaks: the mass of a peak in the details window, the mass measure of the clustering, the mass window of a comparison. The peaks themselves are not fitted again - a calibration says what a fraction position means, it does not change the fit. What it does change is every result that was measured on the old axis: a clustering on the mass scale is computed again from the new masses (the message says so), and if the calibration is removed, the mass measure says why it cannot run instead of quietly measuring something else.
Do the markers fit?
The calibration graph updates while marker masses are edited. It shows the fitted curve, measured markers as dark dots and estimated values as pale dots, making outliers and typing errors easier to spot.
The mass axis is drawn logarithmically, because that is the space in which a good calibration is a straight line - a gel curve on a linear axis is a hyperbola, and a good one is hard to tell from a bad one. Outside the outermost markers the curve is dashed: there the model extends its outer interval instead of interpolating between two standards, and a number read off there deserves less trust.
Validation
Validation does not answer “is this clustering good?” – nothing can. It answers a narrower and more useful question: if the tree is cut into k clusters, how well are they separated? The panel computes the indices for the cuts of the current clustering, and the curve shows how one of them behaves as k changes. A maximum (or minimum) in that curve is a hint at where the data divides most clearly – not a decision.
| Index | What it measures | Better |
|---|---|---|
| Dunn | the smallest distance between two clusters against the largest diameter inside a cluster | higher |
| Silhouettes | per profile, how much closer it sits to its own cluster than to the nearest other one – averaged over all profiles | higher |
| C-Index | the distances inside the clusters against the best and the worst they could be | lower |
| Davies-Bouldin | the worst ratio of spread to separation over all clusters | lower |
| Q Modularity | how much more of the profile similarity falls inside the clusters than chance would give | higher |
Three settings belong to the calculation, and all three have to be written down if the number is to be reproducible:
- index – which of the five criteria is drawn; one curve at a time, so comparing runs means comparing measurements of the same kind.
- distance – the indices have a distance of their own, Euclidean by default, and it is not automatically the one the clustering used. They are also computed on the stored values, without the normalization of the clustering. That is deliberate – the indices judge the clustering as it was drawn – but it means a number from this panel belongs together with these two settings.
- max points – how many cuts of the tree are evaluated; the curve then has one point per k. Lower it if the computation takes too long on a large data set.
Validation is a tool, not a verdict: it says which cut of the tree separates the data best according to one criterion. The biology still decides.
Comparing up to four runs
Two runs of the same sample – before and after a treatment, wild type and mutant – can be put side by side. Open → Open File for Comparison asks for the second file; Load Demo → Comparison does the same without a file: it loads the demo data and puts three runs of it beside the project, one of them with three sets to pick from its header. Open File for Comparison can be used again – every further file adds one more map, up to four in one comparison, the project's own map included, and a fifth is refused with a note in the status line. A file dropped onto the window is offered the same choice (Advanced data import).
Every map carries its own header above it, and every map has a letter: A) is the project – the reference, whose rows and their order carry the others –, B) to D) follow in the order the maps stand. Next to the letter stands the name the map was read from (cut short; the whole name is in the tooltip), and then what belongs to this map: its set – the control says Set (98), the number of profiles in the set that is shown, and opens the list of sets with the name and the number of each –, the arrows ‹ › that move the map one place to the left or right (dragging its header does the same), and the small × that takes it out again. The project's header carries Exit comparison, which closes them all at once. A comparison is one picture, not four: taking one map out leaves the rest standing, and taking out the last one closes the comparison.
Dropping a map on the reference – the A) card – asks a question instead: Use this map as the reference? Saying yes swaps the two roles: that map becomes A), and the map that was the reference stands in the place it left, as a map like the others. From then on the new map's profiles and their order carry the comparison, its row names and its groups are the ones you see on the left, and its sets are the sets of the project. The map that was the reference keeps its data set, but not its groupings, its subsets and its cluster tree – and there is no way back, which is why the question offers Download project right beside Yes, replace the reference.
Compare → compact hides profiles that are absent from the reference project, leaving only rows with a counterpart on the left. Turning it off restores the hidden profiles; no data is removed.
All maps of a comparison share everything: the scroll position, the selection and one colour scale – the same value has the same colour in every one of them – and they are all the same width: the tree, the profile names and the buttons keep their room, and what is left of the window is divided equally among the maps. The profile names stand once on the left, because every map shows the same rows.
Fraction selections are shared across every map in a comparison: selecting a fraction in one map marks the corresponding lane in all of them.
The profile plot below belongs to the project – and inside a comparison it also shows every other run: wherever such a run has the same profile, its curve is drawn in the same colour, behind it, thinner and faded, one run after the other getting fainter. The legend names each of them with “(faded)” appended, and your own curve stays the loud one.
Profiles can be dragged by their names here just as anywhere else – also with Alt+↑/↓. A profile that only another run has exists nowhere in your data, so it keeps its place at the bottom of the list.
The colours of all the maps come from one scale by default (the shared across maps half of the Scale chip, Colours and scaling): the smallest and the largest value of all the runs together – that is what makes one colour mean one amount in every map, and it is why a protein can be read across the maps at all. It has a second side worth knowing: a run of which less was loaded – or that was normalised differently – looks paler all over, although its pattern is exactly the same, and one extreme value in one run pushes the other maps towards the quiet end of the palette. So look at the overall brightness of the maps before you take it for biology. If one run is systematically lower, either switch the chip to each map (every map then gets the range of its own values: the run shows its full contrast, but the colours of the maps cannot be compared with each other any more), or compare the numbers (cell tooltip, table export) rather than the shades, or bring the runs onto a comparable scale beforehand – normalization in the clustering settings changes what is clustered, not the values in the cells.
And the maps can become one data set:
Compare →
Combine maps… asks which maps take part, what the values of a
cell become – mean, median,
max, min or
sum – and what a gap means: a run that has no value there
either does not count for that cell, or counts as a zero. On top of that a profile can
be required to be present in several of the maps, which is how “only what two runs saw”
is said. The sentence under the fields always says what the result looks like, and the
buttons at its foot decide where it goes:
Open result as shared map (filled blue, the first one)
puts it on the server and opens it in a window of its own,
Download result writes it into a .cmap file, and
Replace project with result puts it in place of the project,
which then is the combined data set and can be clustered like any other.
The result is named for what it is: Combined result: mean values, with its one set combined data: mean – this data set was not measured, it was computed.
After upload, the combined map can be opened directly or revisited from its private link.
Matched is by name: first the full label, then the accession, then the ID at the start of
the label; fractions are matched by their names, or by position when the names carry the
run (as in iBAQ 48_S2-A1 against iBAQ 48_S1-A1).
Two peak sets get a rule of their own. When the project and a map are
both peak sets (Splitting profiles into peaks), the rows are fitted peaks –
and neither the label nor the accession may decide which peak of one run is which peak
of the other: every peak of a protein carries that protein's accession, so matching by
it would collapse the peaks of one protein onto a single row. Peaks are matched on
their calibrated mass instead: two peaks of the same protein are
candidates when max(ma, mb) / min(ma, mb)
≤ 1 + t, with Peak matching →
mass tolerance (%) as the number (10 % to begin with). The
condition is symmetric – swapping the two runs changes nothing, so it is not “± t %” –
and it is a starting value for the trial, not a validated limit. Peak
widths are shown as information and never required: two runs may resolve the
same population differently.
Within a protein the assignment is decided for the whole set of its peaks: as many pairs as the window allows, and among those the smallest total mass distance. What does not find a partner stays unmatched and says why – four different findings, never one label for all of them: Missing mass calibration (the run has no usable calibration), No such protein in the other measurement, No peak within the tolerance (the protein is there, but nothing of it is inside the window) and Taken by the one-to-one assignment (a candidate was there, the partner went to another peak). A pair is marked ambiguous when leaving it out would give a pairing of the same size that is very nearly as good – the data does not decide which peak belongs to which. Ambiguous pairs are a subset of the matches, never counted as unmatched, and how near “very nearly” is stands in the tooltip.
The status line reports matched pairs, how many of them are ambiguous, and the unmatched peaks of each map; further statistics are available from the Compare button. A match means “assigned under these settings”, not “biologically confirmed as the same complex”.
Without a mass scale, no peaks are matched by mass: they stay
unmatched with the reason Missing mass calibration and the rest
of the comparison works as usual. There is no automatic fallback to the fraction
axis – but the axis can be chosen deliberately: match peaks by
offers Mass (calibrated) and
Fractions (bands that overlap, the rule of earlier versions),
and one comparison uses one of the two, never both. The mode and the number travel
with the comparison in the .cmap file, and changing them re-assigns the
peaks without fitting anything again.
A peak set and a set of profiles cannot be compared: a row is a fitted peak on one side and a measured profile on the other, so there is nothing to line up. That comparison is refused with the reason – compare two peak sets or two profile sets.
Exit comparison – or taking out the last map with its × – ends it; the project stays as it was.
A comparison is saved with the project: Download
project writes it into the .cmap file, and
Create private link puts it into the shared snapshot – the
two are the same file, so whoever opens it gets the maps back, not just the project
(Saving and exporting, Sharing a project). Opening it runs
every map through the same checks as adding one by hand, and the
status line says once what came back. The only export that does
not carry the comparison is NOVA project (.npf), which
holds a single project – the status line says so after the export.
A figure of a comparison holds every map: the export draws them
side by side, in the order of their headers and on the aligned rows – the project's map on
the left with the names, the other ones beside it, all in the shared colours of the
window and each with its own header (Saving and exporting). The
tables follow the same rule: .xlsx and .ods get one sheet per
map, CSV and TXT one block per map, each introduced by
# and the name of the map – and the first sheet (the first
block) holds all of them side by side, one block of columns after the
other, so the runs can be read against each other.
Saving and exporting
The Export menu has two kinds of output: the project itself, and a picture or table for a figure.
.cmap is this program's project file, .npf is the
original's format. The Peaks block is there in every project
and greyed out until one holds a peak set.- Download project writes everything as one
.cmapfile: the data, the sets, groups, masses, colours – and, when one is open, the comparison with all of its maps (Comparing up to four runs). Opening it brings the session back. - NOVA project (.npf) writes a file that the original NOVA software opens – for handing data to someone who still works there. It holds the project and nothing else: a comparison is not part of it, and the status line says so when one was open.
Two of the data options concern the table itself: include group tag adds the group of every profile as a second column (the same tag the group export writes, which is what lets the table be imported again with its colours), and add the masses of the fractions as a row writes the mass of every fraction as a line of its own under the header – the ones that were typed in, and in the gaps the values of the calibration, like the ruler above the heat map. Read back, that line becomes the mass scale of the file instead of a profile, with the unit it names (Opening a file). It exists only once the scale knows masses (Mass scale); without them the option is greyed out and says so.
Keyboard shortcuts
Manual ↗ in the toolbar opens this guide in its own tab – the arrow says so, the same one the links of the help dialog carry – and the ? next to it opens this dialog: the link to this guide at its top, below it the shortcuts you reach for while working. The keys everybody knows by heart and the panel switches used once per session sit behind the entry More shortcuts in the same dialog. Both lists are below, in the same order.
| Keys | Action |
|---|---|
| Ctrl+A | select all – profiles, or fractions, whichever you marked last |
| Ctrl+I | invert the selection |
| Ctrl+Shift+A | select all members of the group |
| Ctrl+D / Ctrl+E | disable / enable the marked profiles or fractions |
| Del | the same as Ctrl+D – disable the marked profiles or fractions, whichever axis you marked last |
| Ctrl+Shift+D | new set from the selection |
| Ctrl+H | show or hide the disabled profiles |
| Ctrl+Shift+H | leave the group isolation |
More shortcuts
One click deeper in the dialog (figure 37 shows it opened):
| Keys | Action |
|---|---|
| Esc | clear the selection |
| Ctrl+G | profile viewer |
| Ctrl+V | validation panel |
| Ctrl+M | mass scale panel |
| Ctrl+F | search |
| Ctrl+S | save the project |
| Ctrl+O | open a file |
| ↓ / ↑ | move the selection |
| Shift+↓ / Shift+↑ | extend the selection; the opposite arrow takes the last step back |
| Alt+↓ / Alt+↑ | move the selected rows – the keyboard way of dragging a name |
Files and offline use
The two project formats
.cmap– this program's project: data, sets, groups, masses, colours and the scale. This is the file to keep..npf– the project of the NOVA software. It can be read and written; the two programs can open each other's file. It does not carry everything this program knows: a peak run – its settings, its flags and the lab-book entry – lives in the.cmaponly, and the original software knows no peak analysis at all, so a peak set travels as what it is: a data set of its own. A comparison of several runs and the colour view are not part of an.npfeither. The export says what it had to leave behind.
What needs the network, and what never does
Reading, computing and drawing your data happens entirely in your browser – the numbers never leave the machine, and nothing is loaded from third-party servers either: no fonts, no scripts, no analytics, no telemetry about your data. Four things talk to a server, and three of them only when you ask:
| What | When | What goes out |
|---|---|---|
| Import ComplexomeMap | you paste an address | the address, to complexomemap.de |
| Create private link / open a link / delete | you press the button | your project data |
| Load diagnostics | every time the program starts from the server | a timestamp, used to log how far the server clock is off – no project data, no identifier |
| Update check | same | a request for the current version of the program files |
After the first visit the program is stored in the browser and starts without a connection; offline, everything keeps working except those four.
Questions and answers
Is it wrong to put fitted peaks into my map and cluster them? For anything you want to show or publish, yes. Two kinds of numbers end up in one matrix: the measured values of the gel, and the model values of a fitted curve. Three things follow from that. (1) The measured profile is still there beside its components, so the same signal is counted two or three times - one protein can end up dominating the picture. (2) A fitted curve is perfectly smooth and knows nothing about noise, so the distance between it and a real profile says as much about the fit as about the gel. (3) Whether a band was really two is an assumption of yours (prominence, minimum distance, the width limits), and after the split you can no longer see which number came from the assumption - you read your own guess back out of the cluster order. That is why the sound version is Split peaks → Create peak set: there every row is a fit, clustered on position and width (or on the calibrated mass), and no measurement is mixed into it. Insert into this map exists because looking at a split inside your own map is still useful - as a crutch for the eye, not as a result, and Remove the inserted components takes those rows out again.
My table opens with the wrong profile labels. During import,
clusterMAP chooses the source column for the profile labels: UNIQID when
present, otherwise the first column. In a header-less file, it builds the label from the
ID | NAME pair in the last text columns (the
status line says when it did). This import mapping cannot be
changed afterwards; if the wrong column was used, correct the source file or its column
header and open it again. Individual profile labels can still be renamed with the
pencil beside them while no comparison is open
(Selecting profiles and fractions). Comparisons match profiles
first by their full label, then by a recognised accession, and finally by the leading
ID.
Why does it say estimated next to my masses? Values the calibration produced are marked - they came from your marker masses and the chosen calibration model, not from the gel. Type the value in directly if it was measured; the panel then keeps it as a marker and never overwrites it.
The comparison shows white rows. Those profiles exist only in one of the runs. That is the point of the comparison – the tooltip names it, and the Compare button counts them. Tick compact there and the white rows that your project does not have disappear.
A clustering looks different every time. It should not: the algorithm is deterministic for the same data and settings. Check the linkage, distance and normalization of the set (they are shown in its tooltip), and whether profiles are disabled.
Where did my shared project go? The upload sits on the server until somebody uses the deletion link. The list under Share keeps both links for this browser; without them the upload cannot be removed.
Does my data leave the computer? Only if you press Create private link: that is the one button that sends your project data out, because a link has to carry it. Everything else happens in the browser – a peak set that goes to a second tab stays here as well (Files and offline use).
There is a ⚠ in the status line after my clustering. The linkage and the distance do not go together. WARD and CENTROID are only defined on Euclidean distances; and WARD wants the squared Euclidean one, because its criterion is a variance. With a correlation distance, with Manhattan, or with plain Euclidean under WARD, the result is not what the names promise (Clustering). The run was not stopped, because a project from the original NOVA has to keep showing what its author saw – but the numbers cannot be read as “minimal variance” any more.
Is the mass axis a molecular weight? It is an apparent native mass estimated from your calibration, not an independent measurement – see Mass scale.

