Examples
Interactive TMAP visualizations across chemistry, biology, and machine learning.
EGFR Kinase Inhibitors - SAR Navigation
Structure-activity relationships for 10,466 EGFR kinase inhibitors from ChEMBL (CHEMBL203). 677 activity cliffs detected, SAR continuity score 0.88.
Protein Fold Space - SCOPe ASTRAL 40%
15,177 protein domains across 1,257 folds from SCOPe ASTRAL 40%. Cross-class boundary edges: 19.1%. Mean subtree purity: 0.88.
Protein Function 3D - EC Class Map
35,000 proteins organized by embedding similarity. Color by EC class or phylum, search UniProt accessions, and open AlphaFold-backed 3D structure cards.
Oxford Pets - Breed Classifier Audit
3,669 images, 37 breeds, 90.4% linear probe accuracy. 302 boundary edges (8.2%) reveal confused breed pairs. Traces failure paths from misclassified to correct predictions.
MNIST Digits - Cosine Metric
70,000 handwritten digits (784D pixel vectors) with cosine metric. Trace paths between morphologically similar digits: 3 to 8, 4 to 9, 7 to 1.
Enamine Chemical Cluster - Molecular Properties
~6,000 molecules from Enamine cluster 65053. Multiple color layers: MW, LogP, ring count, QED, Murcko scaffolds. Demonstrates filtering and search panels.
Fashion-MNIST - Category Confusion Analysis
70,000 clothing items (784D) with cosine metric and image tooltips. 14.3% boundary edges. Top confused: Shirt vs T-shirt (1,844 edges), Coat vs Pullover (1,584 edges).
PBMC 3k - Single-Cell RNA-seq
2,638 peripheral blood cells, 8 cell types, euclidean metric on PCA. 5.8% boundary edges, subtree purity 0.88. Pseudotime and gene marker color layers.
Tox21 - Multi-Endpoint Toxicity
7,831 compounds screened across 12 toxicity endpoints. Each endpoint is a filterable color layer: toggle between NR-AR, NR-AhR, SR-MMP, SR-p53, and 8 more.
Approved Drugs - ATC Classification
3,039 approved drugs from ChEMBL (phase 4), filtered to MW >= 100. Colored by ATC therapeutic class. 62.8% boundary edges, subtree purity 0.47. SMILES tooltips, drug names.
COCONUT - Natural Products
25,000 natural products from the COCONUT database (738k total). NP classifier pathways (Terpenoids, Alkaloids, Shikimates, Polyketides, Fatty acids). 15.5% pathway boundary edges, subtree purity 0.76.
ESM-2 Protein Embeddings - Sequence Families
3,164 protein sequences embedded with ESM-2 (650M parameter model), cosine metric. Colored by protein family and sequence length.
ESM Atlas - Metagenomic Protein Universe
50,000 metagenomic proteins from the ESM Atlas, embedded with ESMC-600M and folded with ESMFold2. Two maps - raw 1,152-d embeddings and 16,384-d sparse-autoencoder features - with click-to-view predicted 3D structures.
Oral Peptide Design Landscape
1,879 cyclic peptide analogues across 5 drug scaffolds (Cyclosporine, Linaclotide, Octreotide, Desmopressin, Pasireotide), encoded with MAP4 atom-pair fingerprints and mapped by Jaccard similarity.
CH103 Antibody Lineage - Affinity Maturation
459 GenBank CH103 heavy-chain sequences aligned to the published germline UCA. Tree hops from the UCA track nucleotide divergence (Spearman 0.253), with TMAP edges 3.6x tighter in mutation burden than random pairs.
SARS-CoV-2 Spike Evolution - Nextstrain Lineages
24,820 SARS-CoV-2 spike sequences from the Nextstrain open build, spanning 59 clades and 2,486 Pango lineages. Tree hops from Wuhan-Hu-1 strongly track collection date (Spearman 0.855).
Arabidopsis Root - Ground Tissue Atlas
22,600 root cells (Endodermis, Cortex, Quiescent Center) from Shahan et al. 2022. Only 2.4% boundary edges; tree depth from the quiescent center correlates with developmental pseudotime (Spearman 0.877).
Lung Regeneration - Club-to-Ciliated Trajectory
20,000 lung epithelial cells across a regeneration time course. 86.8% same-state edges, subtree purity 0.74. A trajectory path traces Club cells through 8 label transitions to ciliated fate.
Direct Reprogramming - Fibroblast to iEP
12,000 cells sampled across a fibroblast-to-iEP direct reprogramming time course. TMAP pseudotime correlates with reference pseudotime (Spearman 0.771); an example path traces day 0 to day 28.
Pancreas Development - scVelo Pseudotime Audit
3,696 pancreatic endocrine cells (Bergen et al. 2020). TMAP tree-hop distance from the scVelo root correlates with RNA-velocity latent time at Spearman 0.897, and with velocity pseudotime at 0.931.
Caltech-256 - Object Confusion Map
30,607 images across 257 categories (DINOv2 embeddings). 96.5% of edges stay within the same broad object group; the most surprising cross-category bridge is sneaker <-> tennis-shoes (132 edges).
CUB-200 Birds - Species Confusion Map
11,788 bird images across 200 species (ResNet-50 embeddings). 81.5% of edges stay within the same family, but species-level subtree purity drops to 0.430.
EMNIST - Character Confusion Map
20,000 handwritten characters across 47 classes (raw pixels, cosine metric). 70.6% of edges connect the same character; '0'/'O' (654 edges) and '1'/'L' (534) dominate the rest.
Oxford Flowers 102 - Morphological Similarity Map
8,189 flower images across 102 species (ResNet-50 embeddings). 83.4% of edges stay within the same morphological group; species-level purity averages 0.649.
50,000 Words - Semantic Map of English Vocabulary
50,000 English words (WordNet nouns) embedded with MiniLM, cosine metric. Semantic categories overlap heavily (91.5% cross-category edges) - the real signal is local paths, like dog to wolf in 46 hops.
20 Newsgroups - Topic Map
18,846 Usenet posts across 20 newsgroups, MiniLM embeddings, cosine metric. 73.5% same-topic edges, mean subtree purity 0.62 (0.78 at the supercategory level).
Indo-European Languages - Cognate Map
160 Indo-European languages mapped by shared cognate vocabulary (IE-CoR dataset, Jaccard metric). 91.2% of tree edges connect same-family languages - a 4.7x lift over the 19.4% random-pair baseline.
Earth Microbiome Project - 16S Habitat Atlas
26,181 samples from the Earth Microbiome Project, OTU presence/absence with Jaccard metric. 95.2% same-habitat edges at the broad EMPO level, 92.0% at the detailed level.
Public Space Catalogue Atlas
15,379 tracked space objects from CelesTrak, orbital elements with Euclidean metric. Regime-boundary edges: just 0.1% - mean subtree purity by orbital regime is 0.998.
NASA C-MAPSS - Turbofan Degradation
20,631 cycles from 100 simulated turbofan engines (NASA C-MAPSS), 15 sensors, cosine metric. Mean remaining-useful-life gradient across tree edges: 42.9 cycles.
OpenSky Flight-Route Atlas
200,000 flights from OpenSky, seven-feature route fingerprints with Euclidean metric. Route-region boundary edges: just 0.8% - each origin-destination region keeps its own neighborhood.