ProTransAtlas
A comprehensive evidence-based resource for prokaryotic transposable elements, integrating HMM-based annotation, protein-level weak labels, sequence clustering, deep learning embeddings, and curated negative references.
Database Overview
ProTransAtlas is a prokaryotic transposon evidence database constructed from large-scale HMM (Hidden Markov Model) scanning of bacterial and archaeal proteomes. The database currently indexes 7,069,013 HMM hits from 142 unique HMM accessions representing 123 distinct protein domain descriptions, covering 7,060,499 unique proteins with associated genomic coordinates (contig, start, end, strand) for 10,651,187 protein records.
The HMM profiles originate from two major sources: NCBI-curated family models (NCBIFAM, NCBI Protein Clusters) and Pfam domain models. Each HMM description was manually classified into an evidence hierarchy (A1 through U1; see Section 2) based on its specificity for transposition-related functions, following the framework established by the ISfinder database (Siguier et al., 2006, PMID: 16381877) and the comprehensive IS family review by Siguier, Gourbeyre & Chandler (2014, PMID: 24499397).
Evidence-Level Classification
Each HMM model is assigned an evidence level reflecting the strength and specificity of its association with transposition:
| Level | Meaning | HMM Hits | Models |
|---|---|---|---|
| A1 | Family-specific core transposases (IS families, Tn3, Mu, RAYT, TnsA) | 2,787,823 | 37 |
| A2 | Family-specific domains or partial core proteins (N-terminal, C-terminal, domain components) | 470,208 | 27 |
| B1 | Transposition system components (TnsC ATPase, Tra conjugative proteins, excisionases) | 210,136 | 11 |
| B2 | Accessory, regulatory, and transfer-support proteins (TnpC, TnpV, TnpW, DUF2080) | 110,363 | 8 |
| C1 | Host-domesticated transposable element-derived proteins (e.g., PEG10) | 4,688 | 1 |
| D1 | Other mobile genetic element proteins (phage integrases, conjugal transfer, viral RdRP) | 192,779 | 5 |
| D2 | Non-transposon or low-specificity proteins (TIM barrel, NHL repeat, chromodomain, etc.) | 1,448,046 | 12 |
| U1 | Generic/unresolved transposition-related domains (generic transposase, integrase, RT, RNase H, DDE) | 1,844,970 | 11 |
This hierarchical classification allows researchers to filter annotations by confidence level. A1 models provide high-confidence transposase identification suitable as positive training seeds, while D1 and D2 models serve as hard negative and background negative controls respectively. U1 models represent unresolved domains requiring additional contextual evidence for classification.
Protein-Level Training Labels
Building on the evidence hierarchy, each of the 7,060,499 unique proteins was assigned a training label by aggregating all its HMM hits and applying a rule-based classification system (version protein-label-v1.0).
train_positive
2,784,330
Prokaryotic A1 core transposase proteins used as positive training seeds. These proteins carry IS-family, Tn3, Mu, or RAYT transposase domain hits with no conflicting negative evidence.
hard_negative
192,495
D1 proteins from other mobile genetic elements (phage, plasmids, conjugative systems). Critical for distinguishing transposons from other MGE classes.
background_negative
1,446,447
D2 low-specificity or cellular proteins. Each carries exactly one D2 HMM hit and no positive transposon evidence. Provides broad background coverage but should not be treated as absolute negatives.
discovery_pool
1,842,822
U1 unresolved proteins carrying generic transposition-related domains (integrase, RT, RNase H, generic transposase) that cannot be reliably classified with current HMM evidence alone.
conditional_positive
297,620
A2 partial/domain evidence. May represent fragmented transposases or multi-domain proteins requiring structural context for confident classification.
auxiliary_component
320,113
B1/B2 system and accessory components. Not core catalytic proteins but important for full transposition system annotation.
The training label system is designed for weak supervision: labels are HMM-derived and should be treated as noisy annotations. For supervised learning, we recommend using the 4,423,272 proteins in the supervised training set (train_positive + hard_negative + background_negative). The 2,140,442 discovery-eligible proteins serve as a pool for positive-unlabeled (PU) learning and novel family discovery. Users should apply homology-based clustering before train/validation/test splitting to prevent data leakage.
MMseqs2 Sequence Clustering
To reduce redundancy and identify sequence subfamilies, we performed MMseqs2 (v16.747c6) clustering on all 2,784,330 positive training proteins grouped by their seed IS family annotation. Clustering was performed with 50% minimum sequence identity and 80% coverage of the shorter sequence (-c 0.8 --cov-mode 1), producing:
37
IS Families
2,784,330
Input Proteins
670,144
Clusters (4.2× compression)
The largest families are IS630 (427,127 proteins → 86,838 clusters), IS110 (334,754 → 72,461), IS21 (251,094 → 62,188), IS607 (221,627 → 33,209), and IS481 (158,387 → 37,444). Cluster representatives and complete cluster membership are available via the Clusters page, with per-family Newick trees downloadable for import into iTOL or FigTree.
Deep Learning Embeddings
Each cluster representative sequence was encoded using ESM2 (Evolutionary Scale Modeling), a protein language model pre-trained on 65 million UniRef sequences. We used the t6_8M variant (8 million parameters) which produces 320-dimensional per-protein embeddings via mean pooling of residue-level representations across the full sequence (excluding BOS/EOS tokens).
676,724
Encoded Proteins
320
Dimensions
~90 seq/s
GPU Throughput
Encoding was performed on an NVIDIA RTX 4080 SUPER GPU over ~2 hours. The embeddings capture evolutionary and structural information encoded in the protein sequences and are suitable for downstream machine learning tasks including family classification, novel element discovery via embedding-space clustering, and transfer learning for transposon annotation.
An interactive UMAP projection of 11,100 representative sequences (300 per family) is available as a clickable scatter plot, where each point links to the corresponding protein detail page. This visualization enables visual exploration of embedding-space relationships among transposon families.
Independent Curated Negative Reference
To complement HMM-derived negative labels, we constructed an independent set of 14,267 reviewed core-cellular proteins from UniProtKB/Swiss-Prot (curated-negative-v1.0). These proteins were selected from five functional groups with well-established non-mobile cellular roles:
5,000
Ribosomal
5,000
tRNA Ligases
2,567
RNA Polymerase
1,479
Chaperones
221
Topoisomerase
All records were filtered to remove exact sequence duplicates across groups, and sequences containing mobile-element annotation keywords (transposase, integrase, recombinase, reverse transcriptase, phage, plasmid, insertion sequence, retron) were excluded. This set is restricted to bacterial (taxonomy_id:2) and archaeal (taxonomy_id:2157) sequences and is available for download as FASTA and TSV.
These curated negatives provide a high-confidence, source-independent complement to the D1/D2 HMM-derived negatives and are particularly valuable for training models that must distinguish transposon proteins from essential cellular machinery.
BLAST Search & Data Access
The database provides a built-in BLASTP interface against all 670,144 cluster representative sequences (max 100 query sequences per submission), with downloadable results as TSV. A full-text search engine supports case-insensitive queries across protein IDs, HMM accessions, family descriptions, and annotations, with results grouped into Models, Families, and Proteins sections and full download support (TSV and FASTA).
All data products are available for direct download: protein-level training labels (Parquet 68 MB, TSV 1.1 GB), curated negative references (FASTA 8.3 MB, TSV 12 MB), cluster representative sequences (per-family FASTA, 186 MB total), and phylogenetic Newick trees (37 files).
Scientific Impact & Use Cases
Novel transposon family discovery. The evidence hierarchy and training labels enable positive-unlabeled (PU) learning approaches. The 1.84M proteins in the discovery pool, together with the 63M+ unannotated proteins from the source proteomes, constitute a rich search space for identifying previously uncharacterized transposon families through embedding-space clustering and outlier detection.
Automated genome annotation. The ESM2 embeddings provide a unified representation space where transposon proteins can be identified by proximity to known family representatives. A classifier trained on these embeddings can annotate transposons in newly sequenced genomes without requiring HMM scanning, potentially improving both speed and sensitivity for large-scale metagenomic applications.
Evolutionary studies. The per-family sequence clusters and phylogenetic trees enable comparative analyses of IS family expansion, host range, and sequence divergence. The inclusion of genomic coordinates allows investigation of insertion site preferences, genomic context, and the role of transposons in genome plasticity.
Benchmarking and method development. The combination of HMM-based weak labels, curated independent negatives, and deep learning embeddings makes this database suitable as a benchmark for evaluating transposon annotation tools, protein language model fine-tuning strategies, and semi-supervised learning methods in genomics.
Key References
IS classification & annotation:
- Siguier P, Perochon J, Lestrade L, Mahillon J, Chandler M. ISfinder: the reference centre for bacterial insertion sequences. Nucleic Acids Res. 2006. PMID: 16381877
- Siguier P, Gourbeyre E, Chandler M. Bacterial insertion sequences: their genomic impact and diversity. FEMS Microbiol Rev. 2014. PMID: 24499397
Sequence clustering:
- Steinegger M, Söding J. MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets. Nat Biotechnol. 2017. PMID: 29035372
Protein embeddings:
- Lin Z, Akin H, Rao R, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science. 2023. PMID: 36927031
IS110 mechanism:
- Durrant MG, Perry NT, Pai JJ, et al. Bridge RNAs direct programmable recombination of target and donor DNA. Nature. 2024. PMID: 38926615
- Siddiquee R, Pong CH, Hall RM, Ataide SF. A programmable seekRNA guides target selection by IS1111 and IS110 type insertion sequences. Nat Commun. 2024. PMID: 38898016
UMAP:
- McInnes L, Healy J, Melville J. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction. arXiv. 2018. arXiv:1802.03426
Phylogenetic trees:
- Price MN, Dehal PS, Arkin AP. FastTree 2 – Approximately Maximum-Likelihood Trees for Large Alignments. PLoS ONE. 2010. PMID: 20224823
- Katoh K, Standley DM. MAFFT Multiple Sequence Alignment Software Version 7. Mol Biol Evol. 2013. PMID: 23329690