telomere-analysis

Illustrative computational analyses for a review on telomere biology (Maji, Dash, Chattopadhyay; Biochimie, in revision). All outputs are predictions, not experimental measurements.

Analyses

• G4Hunter on telomeric motifs. Telomeric repeat motifs from 49 eukaryotic species (19 distinct motifs; species.tsv) are tiled to 600 nt and scored with G4Hunter (Bedrat et al., NAR 2016).
• G4Hunter on the hTERT proximal promoter. A 400-nt window (chr5:1,294,991-1,295,390, GRCh38; -2 to -401 relative to the TERT ATG) is scored on the G-rich template (plus) strand, for the wild type and the C228T (-124C>T) and C250T (-146C>T) hotspot mutations, modelled as G>A on the plus strand.
• DNA language-model embeddings. Each distinct motif is tiled to 1,200 nt and embedded with Nucleotide Transformer v2 (500M, multi-species; attention-masked mean pooling of the final layer). Species that share a motif share an embedding, so only the 19 distinct motifs are embedded. Similarity is shown as a cosine-similarity heatmap with average-linkage clustering.
• Physicochemical descriptors. Structures of 17 small molecules are retrieved from PubChem by name (CIDs recorded in results/compounds_resolved.tsv), reduced to the parent structure (largest fragment), and described with RDKit.

Files

File	Description
species.tsv	Species panel: motif (5' to 3', G-rich strand), clade, primary source, verification status
compounds.tsv	Compound list, PubChem query names, optional SMILES overrides with their source
panel.py	Shared helpers (panel loading, tiling, labels, colours)
g4hunter.py	G4Hunter on telomeric motifs and the hTERT promoter
nt_emb_extr.py	Nucleotide Transformer embeddings and similarity figure (optional UMAP)
drug.py	PubChem retrieval and RDKit descriptors
download_ucsc.py	Fetches the hTERT promoter window from the UCSC REST API into htert_promoter.json
check_results.py	Consistency checks on the panel and all outputs

Reproducing

conda env create -f env.yml
conda activate nt-env
pip install rdkit

python g4hunter.py
python nt_emb_extr.py
python drug.py
python check_results.py

nt_emb_extr.py needs a GPU for reasonable speed (about a minute on 24 GB) and caches embeddings in results/motif_embeddings.npz; pass --recompute to ignore the cache and --umap for an additional UMAP of the distinct motifs. drug.py caches PubChem look-ups in results/compounds_resolved.tsv; pass --refresh to re-query. download_ucsc.py only needs rerunning to regenerate htert_promoter.json.

Outputs

Output	Use
figures/Fig_insilico_g4	Main in silico figure: (A) motif scores, (B) hTERT promoter, (C) hotspot zoom
figures/Fig_g4hunter_motifs, figures/Fig_htert_promoter	The same panels as separate files
figures/FigS1_nt_motif_similarity	Supplementary Figure S1
figures/FigS2_drug_descriptors	Supplementary Figure S2
results/g4hunter_telomeric_motifs.csv, results/g4hunter_telomeric_species.csv	G4Hunter scores (Tables S2 and S4)
results/htert_g4hunter_tracks.csv, results/htert_hotspot_scores.csv	hTERT window scores
results/motif_embeddings.npz, results/motif_cosine_similarity.csv	Embeddings (19 x 1,024) and similarity matrix
results/compounds_resolved.tsv, results/drug_descriptors.csv	Structures with PubChem CIDs, and descriptors

Environment

Ubuntu, single NVIDIA GPU (24 GB). Requires transformers==4.44.2 (the NT v2 custom code is incompatible with transformers 5.x); weights are loaded in fp32 with torch.autocast for the forward pass.

Limitations

• The embeddings reflect the sequence of the repeat unit only; they carry no information on phylogeny, repeat heterogeneity, telomere length or chromatin.
• G4Hunter is a sequence-based propensity score; the |score| > 1.0 cut-off is heuristic, and the mean score of a strictly periodic tract is fixed by the composition of its repeat unit.
• The hTERT analysis predicts local changes in G-richness only. The established mechanism by which C228T and C250T activate TERT is the creation of ETS/GABP binding sites (Bell et al., Science 2015).
• Descriptors summarise general drug-likeness and do not predict G4 binding or activity.

References

• Dalla-Torre H. et al. Nucleotide Transformer: building and evaluating robust foundation models for human genomics. Nat. Methods 22, 287-297 (2025).
• Bedrat A., Lacroix L., Mergny J.-L. Re-evaluation of G-quadruplex propensity with G4Hunter. Nucleic Acids Res. 44, 1746-1759 (2016).
• Lyčka M. et al. TeloBase: a community-curated database of telomere sequences across the tree of life. Nucleic Acids Res. 52, D311-D321 (2024).
• Horn S. et al. TERT promoter mutations in familial and sporadic melanoma. Science 339, 959-961 (2013).
• Huang F. W. et al. Highly recurrent TERT promoter mutations in human melanoma. Science 339, 957-959 (2013).
• Bell R. J. A. et al. The transcription factor GABP selectively binds and activates the mutant TERT promoter in cancer. Science 348, 1036-1039 (2015).
• McInnes L., Healy J., Melville J. UMAP: Uniform Manifold Approximation and Projection for dimension reduction. arXiv:1802.03426 (2018).
• Landrum G. RDKit: open-source cheminformatics. https://www.rdkit.org
• Kim S. et al. PubChem 2023 update. Nucleic Acids Res. 51, D1373-D1380 (2023).