How to Cluster Peptides

Prev Next

Clustering groups the peptides of a report so you can see which families of analogs exist and how their properties differ. The groups come either from structural similarity or from the values of a dataset column, such as a series or a batch. You set up clustering on the Clusters tab of the Active SAR Slide settings, and you explore the clusters in the Cluster Explorer, whose Summary Table is the main view.

Learn more

For how the grouping is computed, what the quality figures and the Δ stars mean, and why the UMAP plot is not a picture of the clusters, see Understanding Peptide Clustering.

Step 1: Configure and run clustering

Open the settings drawer, choose Active SAR Slide, and open the Clusters tab. In the Clustering Configuration section, turn on Enable clustering.

The Clustering Configuration section of the Clusters tab, with Enable clustering on, Clusters from set to Structural similarity k-medoids, Auto-detect optimal k on, the Component Weights sliders and the Run Clustering button

Choose where the clusters come from in Clusters from:

  • Structural similarity k-medoids groups peptides that look alike. Auto-detect optimal k (number of clusters) is on by default: Ideation tries 2 to 20 clusters and keeps the number that gives the best grouping. Turn it off to set Number of clusters yourself.
  • An Endpoint option, one per column of the dataset, makes one cluster per value of that column. For example, Endpoint: Family on a column holding Linear, Disulfide and Lactam staple gives three clusters. Peptides without a value go to a (no value) cluster. Only columns with at most 10 different values are offered, text or numbers.

Then set the Component Weights: how much the Backbone Monomers, Branches, Bridges and Chemical Chains count in the similarity. A component the dataset does not have is greyed out and marked (none in dataset). When the clusters come from a column, the weights do not change the clusters. They still choose each cluster's representative peptide, and they drive the quality figures, the UMAP plot, the Comparison Matrix and the nearest-neighbour lines.

Click Run Clustering. The button shows how many peptides will be clustered: clustering always uses every peptide of the dataset, whatever your filters. When it finishes, the Clustering Results section lists the clusters, each with its size and its centroid, the peptide that best represents it. Click a cluster to filter the viewer on it. When a cluster is selected, Set as reference peptide makes its centroid the reference.

The Clustering Results section showing the clusters with their sizes and centroid peptides, and the Explore Clusters button

Step 2: Open the Cluster Explorer

Click Explore Clusters to open the full-screen Cluster Explorer. It has three parts:

  • The header, with one button per view: Summary Table, UMAP View, Comparison Matrix, Property Distributions, Parallel Coordinates and Silhouette Analysis.
  • The drawer on the right, with a Clustering section and a Clusters section. Hide or show it with the drawer icon in the header, and drag its left edge to resize it.
  • The footer, which says what the viewer will show, with the export button and Apply To Viewer.

The Clustering section shows how the clusters were made, for example Structural similarity · k-medoids · auto k = 8. Click it to change the settings and re-run without leaving the Explorer: Run applies the new settings, Cancel drops them.

The Cluster Explorer after a run from its Clustering section with Clusters from set to Endpoint: Family: one cluster per family, a (no value) cluster, and the Structural quality column

The Clusters section lists All peptides and every cluster. Click a cluster to select it. Its details open under it: its medoid (the representative peptide), its silhouette quality, its outliers, its average monomer count, and the main endpoint with its Δ. The selected cluster is the one that Apply To Viewer and Export Cluster use, and the one the UMAP view draws neighbour lines for.

Step 3: Read the Summary table

The Summary Table shows one row per cluster. Expand a cluster to see its peptides, with the medoid starred. Use the arrows in the Cluster / Name header to expand or collapse all clusters. The reference peptide's row is tinted, as in the peptide table.

The Cluster Explorer Summary table with the Cluster statistics, Endpoints and Sequence pattern column groups, the Find peptide box and the Legend button, and the drawer on the right

The columns come in three groups. Click the arrow before a group's title to collapse the group to one column.

Column group What it tells you
Cluster statistics Peptides (how many), Quality (the silhouette: Good, Fair or Poor, with the score), Outliers (members that fit poorly), and the average number of Monomers, Bridges and Chem chains per peptide. A cluster of one peptide shows N/A quality. For clusters made from a column, the column reads Structural quality: it says how alike the members are, not whether the grouping is good.
Endpoints One column per endpoint the slide shows, aggregated as the header says (for example mean). Under each numeric value, Δ is how much the cluster differs from the other peptides. Stars mark a difference that is unlikely to be chance (★, ★★, ★★★). Hover a Δ for both values, the peptide counts and the test result. Use the chart icon in an endpoint header to draw the column as a histogram, a boxplot or a bar chart.
Sequence pattern The cluster's consensus: the most common monomer at each position, with markers for chem chains, branches and bridges. The colour goes from shared to variable, and ? marks a position held by less than 60% of the members. When the dataset has branches, switch the pattern between the backbone and a branch in the group header.

Columns are sortable and resizable. Find peptide shows only the peptides whose name matches. Legend explains the symbols and colours. The table keeps its layout (expanded clusters, collapsed groups, the chain shown) with the slide.

Collapse Cluster statistics to bring the endpoint charts and the sequence pattern into view:

The Summary table Endpoints group with one boxplot per cluster on a shared scale and the Δ under each value

To find a promising sub-series:

  1. Read the Sequence pattern to see what structural motif each cluster stands for.
  2. Check Quality and Outliers before you trust a cluster.
  3. Look for a starred Δ in an endpoint column. For example, a cluster with Δ +0.8 ★★ on pIC50 is clearly more potent than the rest.
  4. Select that cluster in the drawer and click Apply To Viewer to work on it in the viewer.

Step 4: Edit the clusters by hand

You can move peptides between clusters when you know better than the algorithm, for example to put a known series together.

  1. In the Summary Table, expand the clusters and tick the peptide rows.
  2. Click Move to cluster and pick a cluster, or click New cluster from selection. You can also right-click the ticked rows.
  3. To rename a cluster, point to it in the drawer's Clusters list and click the pen beside its name. Type the new name and press Enter. Names must be unique.

Every view, the quality figures and the medoids update. Your edits are saved with the slide. The drawer lists them under N manual edits: click it to see them, newest first. Remove one edit with its delete icon, or click Reset all to go back to the clusters of the run.

Running the clustering again replaces your edits

A new run starts from fresh clusters and discards the manual edits. Ideation warns you before you run.

Clusters made from a column cannot be edited: their peptides belong to a cluster by their value.

Step 5: Look at the UMAP view

The UMAP View places each peptide as a point, coloured by cluster, so that peptides of similar composition sit close together. Use the 2D / 3D switch to change the plot. Hover a point to see the peptide name.

The UMAP view in 2D with the peptides coloured by cluster, and the toolbar with the 2D/3D switch, Auto parameters, the n_neighbors and min_dist sliders and Nearest neighbours

  • Auto parameters is on by default: Ideation sets n_neighbors and min_dist from your clusters, so you rarely need to tune them. Turn it off to move the sliders yourself.
  • Click a point to select its cluster.
  • Turn on Nearest neighbours and select a cluster: each of its peptides gets a line to its closest peptides. Set how many with k (3 to 10). A line in the cluster's colour stays in the cluster, a grey line goes to another cluster.
  • Scroll to zoom. In 2D, drag to pan, and Reset zoom brings back the full plot. In 3D, drag to rotate, and Reset view brings back the starting view.

The same data and settings always give the same plot.

Do not judge cluster quality from this plot

The plot is computed from a simpler description of the peptides than the clustering, which ignores where each monomer sits. A cluster may look scattered, and two clusters may seem to overlap, without anything being wrong. Use Quality in the Summary Table instead. The nearest-neighbour lines, by contrast, use the clustering's own distances. See Understanding Peptide Clustering.

Step 6: Use the other views

  • Comparison Matrix shows how far apart the clusters' medoids are.
  • Property Distributions compares the distribution of one endpoint across the clusters.
  • Parallel Coordinates draws each cluster as a line across several endpoints, to compare their profiles.
  • Silhouette Analysis shows how well each peptide fits its cluster.

Step 7: Apply or export the clusters

The footer says what the viewer will show, for example Cluster 3 · 12 peptides.

  • Apply To Viewer filters the viewer on the selected cluster and closes the Explorer. With All peptides selected, it clears the cluster filter.
  • Export Cluster (with a cluster selected) or Export All Clusters downloads an Excel file. Its Clusters sheet has one row per cluster, with the statistics and endpoint values of the Summary Table. Its Peptides sheet has one row per peptide, with its HELM and endpoint values.

Next steps