What this is

An interactive 3-D map of 2,174 graspetide precursor peptide sequences — the full curated set reported by Makarova et al. (2022). Each point is one precursor peptide. Points that sit close together have similar sequences; the spatial structure is learned, not hand-drawn.

How it was built (the short version)

1. Sequences in. Start from Makarova's curated final set of 2,174 graspetide precursor peptides (their Additional file 10, Table S3) and their 174 precursor families / clusters.
2. Turn each sequence into numbers (the embedding). Feed every precursor into ProtT5 — a protein language model (ProtTrans; Elnaggar et al. 2021) trained on hundreds of millions of protein sequences. For each peptide it returns a 1024-dimensional vector that captures "what kind of protein this is," learned purely from sequence context. Think of it as the model's fingerprint of the peptide. We mean-pool over residues to get one vector per precursor.
3. Squash 1024-D down to 3-D (UMAP). 1024 numbers per peptide is impossible to look at, so we use UMAP (Uniform Manifold Approximation and Projection; McInnes et al. 2018) to compress each vector to just 3 coordinates while keeping near-neighbors near. Those 3 coordinates are the UMAP1/UMAP2/UMAP3 axes you're spinning around.
4. Color the points (the biology overlay). Coordinates come only from sequence; everything in the color dropdown is metadata laid on top — Makarova ATP-grasp branch, GTDB taxonomic order, oral vs non-oral (HOMD), graspetide family, CDD hit, and crosslink motifs. Branch labels are mapped from each precursor's cluster to the ATP-grasp phylogeny branch (Makarova Additional file 8, Table S1); taxonomy is assigned via GTDB.
5. Render. A single self-contained HTML file using Plotly (loaded from a CDN). One scatter trace per category, with client-side search and filtering — no server, no install.

Stack: Python · ProtT5 / Hugging Face Transformers · umap-learn · pandas · Plotly. Happy to share the scripts — find me at the poster.

Dataset & provenance

• Source: Makarova KS, Blackburne B, Wolf YI, Nikolskaya A, Karamycheva S, Espinoza M, Barry CE III, Bewley CA, Koonin EV. "Phylogenomic analysis of the diversity of graspetides and proteins involved in their biosynthesis." Biology Direct (2022) 17:7.
• Precursor set: 2,174 sequences, 174 families (Additional file 10, Table S3).
• Branch assignments: ATP-grasp phylogeny branches (Additional file 8, Table S1), mapped precursor-cluster → branch (1:1 in this set).
• Taxonomy: GTDB phylum/order/genus; oral status from the Human Oral Microbiome Database (HOMD).

Cross-check with our own data

We independently identified graspetide biosynthetic gene clusters carried on integrative and conjugative elements (ICEs) in Porphyromonas gingivalis. The precursor sequences we found are present in / identical to entries in Makarova's curated precursor set (Additional file 10), and they fall squarely in Branch_5 — the Bacteroidota-specific branch (in this dataset Branch_5 is 179/209 Bacteroidota, with zero other phyla). The P. gingivalis points here (search "gingivalis") sit in that same neighborhood as Chryseobacterium (chryseoviridin) precursors.

The 4 P. gingivalis points in this plot represent the two types of graspetide operons we identified — two precursors each. Use the Header search box to isolate them: type1 → the ICE-type-1/2 operon (strains Ando, SJD2; cluster2758), and type2 → the type-3-ICE operon (strains JCVI SC001, MP4-504; cluster2680). The poster figure shows which is which.

Read the paper → Makarova et al., Biology Direct 2022 (DOI 10.1186/s13062-022-00320-2)