pygliph 0.1.0

Pure-Python port of the R package turboGliph — GLIPH and GLIPH2 specificity-group clustering of T-cell receptor CDR3 repertoires.

py-gliph

Pure-Python port of the R package turboGliph — implementations of GLIPH (Glanville et al., Nature 2017) and GLIPH2 (Huang et al., Nat. Biotechnol. 2020): Grouping of Lymphocyte Interactions by Paratope Hotspots.

pygliph clusters T-cell receptors (TCRs) into specificity groups predicted to bind the same MHC-restricted peptide antigen, based on shared CDR3 motifs. It is a standalone, dependency-light re-implementation that does not require R or rpy2.

PyPI / import name	pygliph
Repository	omicverse/py-gliph
License	Apache-2.0
Upstream	turboGliph 0.99.2 (GPL-3, R)
Numerical parity	deterministic parts bit-exact vs turboGliph

Install

pip install pygliph              # once published
# or, from a checkout:
pip install -e .

Dependencies: numpy, scipy, pandas, networkx (and matplotlib for plot_network, an optional extra).

What it does

GLIPH/GLIPH2 build specificity groups from two kinds of CDR3 similarity:

• Local (motif) similarity — enriched short continuous CDR3 motifs (2–4-mers). GLIPH2 scores enrichment with Fisher's exact test (hypergeometric) against a naive reference TCR pool; motif clusters are position-restricted (a motif may shift by < 3 residues). GLIPH v1 instead uses repeated random sampling from the reference.
• Global similarity — CDR3s of equal length differing by a single amino acid (Hamming distance 1). GLIPH2 uses position-specific structures with an optional BLOSUM62 interchangeability constraint.
• Specificity-group construction — connected components of the combined local + global similarity graph.
• Per-group scoring — network size, CDR3-length restriction, V-gene bias, clonal-expansion enrichment and shared-HLA enrichment combined into a total.score (GLIPH2's convergence_groups.txt).
• N/P-residue up-weighting — local cluster significance is boosted for motifs overlapping non-germline (N/P-nucleotide) encoded residues.

Quick start

import pygliph as pg

# bundled example: ~2000 TCRs of known specificity
data = pg.datasets.load_gliph_input_data()

# GLIPH2
res = pg.gliph2(data, sim_depth=1000)
res["cluster_properties"]            # scored specificity groups
res["motif_enrichment"]["selected_motifs"]   # enriched local motifs
res["global_enrichment"]             # global-similarity structures
res["cluster_list"]                  # {tag: member DataFrame}

# original GLIPH (v1)
res1 = pg.turbo_gliph(data, sim_depth=1000)

# configurable hybrid
res2 = pg.gliph_combined(data, local_method="fisher", global_method="fisher")

API

Function	Purpose
gliph2	GLIPH2 algorithm — Fisher-test local motifs + global structures
turbo_gliph	original GLIPH — resampling local motifs + Hamming globals
gliph_combined	configurable hybrid of GLIPH / GLIPH2
cluster_scoring	per-cluster enrichment scoring
find_motifs	continuous / discontinuous motif enumeration
de_novo_TCRs	de-novo CDR3 generation from a specificity group
plot_network	specificity-group network visualisation
save_gliph_output / load_gliph_output	result-folder I/O
datasets	bundled example data and reference resources

Each result is a dict mirroring turboGliph: motif_enrichment, global_enrichment, connections, cluster_properties, cluster_list, parameters (and sample_log for GLIPH v1).

Input format

A pandas.DataFrame (or a plain list of CDR3b strings) with column CDR3b and any of the optional columns TRBV, patient, HLA, counts — the same schema as turboGliph.

R parity

tests/ runs the same input through turboGliph (R) and pygliph and asserts agreement. The deterministic parts are bit-exact:

• find_motifs continuous and discontinuous motif counts;
• GLIPH2 local-motif hypergeometric ("Fisher") p-values, fold change, selected motifs and their position ranges;
• GLIPH2 global-similarity structures and their Fisher scores;
• the clone-network edge list (directed 4-tuples);
• cluster membership, sizes and cluster-level Fisher scores;
• de_novo_TCRs position weight matrices and sample-sequence scores.

Unavoidable approximate parts: any score derived from random resampling — GLIPH v1's repeated-random-sampling local-motif selection, and the total.score/length/V-gene/clonal-expansion sub-scores in both algorithms — depends on the RNG. R's sample.int (Mersenne-Twister) and NumPy's PCG64 produce different draws, so these are checked for high set / rank agreement (correlation > 0.95) rather than bit equality. Fix random_state for reproducible Python runs.

References

• Glanville, J. et al. Identifying specificity groups in the T cell receptor repertoire. Nature 547, 94–98 (2017).
• Huang, H. et al. Analyzing the Mycobacterium tuberculosis immune response by T-cell receptor clustering with GLIPH2 and genome-wide antigen screening. Nat. Biotechnol. 38, 1194–1202 (2020).
• turboGliph: https://github.com/HetzDra/turboGliph

License

Apache-2.0. The upstream R package turboGliph is GPL-3; pygliph is an independent re-implementation from the published algorithms and the turboGliph source, and ships the small reference data tables exported from turboGliph for parity testing.