ldclust 0.2.0__tar.gz

ldclust-0.2.0/PKG-INFO

@@ -0,0 +1,112 @@
+Metadata-Version: 2.4
+Name: ldclust
+Version: 0.2.0
+Summary: LD-based clustering of SNPs with a haplotype-testing protocol
+Author-email: Gennady Khvorykh <info@inzilico.com>
+Keywords: linkage disequilibrium,SNP,clustering,haplotype,bioinformatics
+Classifier: Intended Audience :: Science/Research
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Scientific/Engineering :: Bio-Informatics
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+Requires-Dist: numpy>=1.24
+Requires-Dist: scipy>=1.10
+Requires-Dist: h5py
+Requires-Dist: pandas
+Requires-Dist: scikit-learn>=1.3
+Provides-Extra: hdbscan
+Requires-Dist: hdbscan>=0.8; extra == "hdbscan"
+Provides-Extra: graph
+Requires-Dist: networkx>=3; extra == "graph"
+Provides-Extra: igraph
+Requires-Dist: igraph; extra == "igraph"
+Requires-Dist: leidenalg; extra == "igraph"
+Requires-Dist: infomap; extra == "igraph"
+Provides-Extra: all
+Requires-Dist: hdbscan>=0.8; extra == "all"
+Requires-Dist: networkx>=3; extra == "all"
+Requires-Dist: igraph; extra == "all"
+Requires-Dist: leidenalg; extra == "all"
+Requires-Dist: infomap; extra == "all"
+
+# ldclust
+
+LD-based clustering of SNPs with a haplotype-testing protocol: load an
+LD (r²) matrix, define distances or a spectral embedding, cluster SNPs
+into LD blocks, write PLINK `--hap-assoc` hlist files, run the
+haplotype-association endpoint, and select associated blocks.
+
+The package grew out of a benchmark of 25 clustering method families on
+LD matrices (simulated and real chromosome 22); every wrapper returns
+`(labels, probs_or_None, seconds)` and consumes one of four substrates,
+so methods are drop-in comparable.
+
+## Install
+
+```bash
+pip install .                    # core: numpy, scipy, h5py, pandas, scikit-learn
+pip install ".[all]"             # + hdbscan, networkx, igraph, leidenalg, infomap
+```
+
+The SBM family needs [graph-tool](https://graph-tool.skewed.de), which
+is not pip-installable:
+
+```bash
+conda create -n sbm -c conda-forge python=3.10 numpy scipy h5py graph-tool
+conda run -n sbm python -m ldclust.cli --help
+```
+
+## Usage
+
+Command line (one clustering run + benchmark artifacts: blocks csv,
+labels npy, PLINK hlist, `summary.tsv` row):
+
+```bash
+ldclust -p <prefix> --method fof --eps 0.5 --out-dir data/bench
+ldclust -p <prefix> --method hc --deep-split 2 --min-cluster-size 3
+ldclust -p <prefix> --method dpblocks --tag-thr 0.3 --max-len 30
+ldclust --help
+```
+
+`<prefix>` names `prefix.ld.h5` + `prefix.snplist` (an r² matrix with
+SNP ids; pandas-fixed and plain h5py layouts both load); the `gmm` /
+`dpgmm` methods additionally read `<prefix>.gmm.raw` (PLINK `--recodeA`
+dosages) to build the relationship-matrix embedding.
+
+Library:
+
+```python
+import ldclust as ld
+
+ids, r2, layout = ld.load_ld(prefix)          # SNP ids + r2 matrix
+labels, _, sec = ld.cluster_hc(ld.distances(r2, "d1"),
+                               deep_split=2)  # WGCNA-style, dynamic tree cut
+ld.write_hlist("blocks.hlist", ids, labels)   # PLINK --hap input
+assoc = ld.run_hap_assoc(prefix, "blocks.hlist", "run1",
+                         timeout_sec=600)     # PLINK 1.07 --hap-assoc
+selected = ld.select_blocks(ld.parse_assoc_hap(assoc))  # OMNIBUS p <= 5e-6
+```
+
+## Methods (25 families)
+
+| substrate | methods |
+|---|---|
+| precomputed distances (1 − r², √(1 − r²)) | hdbscan, optics, dbscan |
+| spectral embeddings of r² / relationship matrix | soptics, spectral (NJW), gmm, dpgmm, kmeans, minibatch |
+| weighted r² graph | fof (percolation), louvain, leiden, cnm, mcl, lpa, cw (Chinese Whispers), walktrap, infomap, sbm (DC/nested, MDL) |
+| raw r² similarity | ap (affinity propagation), pam (similarity k-medoids), cdhit (CD-HIT-style seed-and-recruit) |
+| hierarchical | hc (average/complete linkage + dynamic tree cut, WGCNA recipe) |
+| ordered SNPs | dpblocks (Zhang et al. 2002 DP, tag-SNP cost) |
+| ensembles | consensus (co-association: DTC on 1 − C or percolation on C) |
+
+Noise label convention: −1 where a method defines it (hdbscan family,
+dynamic tree cut); all other methods assign every SNP. Every wrapper is
+deterministic unless documented (gmm/dpgmm/sbm/minibatch are seeded).
+
+## Development
+
+```bash
+python3 tests/test_smoke.py    # synthetic-matrix smoke tests, no PLINK
+```
+
+Author: Gennady Khvorykh (`info@inzilico.com`).

ldclust-0.2.0/README.md

@@ -0,0 +1,81 @@
+# ldclust
+
+LD-based clustering of SNPs with a haplotype-testing protocol: load an
+LD (r²) matrix, define distances or a spectral embedding, cluster SNPs
+into LD blocks, write PLINK `--hap-assoc` hlist files, run the
+haplotype-association endpoint, and select associated blocks.
+
+The package grew out of a benchmark of 25 clustering method families on
+LD matrices (simulated and real chromosome 22); every wrapper returns
+`(labels, probs_or_None, seconds)` and consumes one of four substrates,
+so methods are drop-in comparable.
+
+## Install
+
+```bash
+pip install .                    # core: numpy, scipy, h5py, pandas, scikit-learn
+pip install ".[all]"             # + hdbscan, networkx, igraph, leidenalg, infomap
+```
+
+The SBM family needs [graph-tool](https://graph-tool.skewed.de), which
+is not pip-installable:
+
+```bash
+conda create -n sbm -c conda-forge python=3.10 numpy scipy h5py graph-tool
+conda run -n sbm python -m ldclust.cli --help
+```
+
+## Usage
+
+Command line (one clustering run + benchmark artifacts: blocks csv,
+labels npy, PLINK hlist, `summary.tsv` row):
+
+```bash
+ldclust -p <prefix> --method fof --eps 0.5 --out-dir data/bench
+ldclust -p <prefix> --method hc --deep-split 2 --min-cluster-size 3
+ldclust -p <prefix> --method dpblocks --tag-thr 0.3 --max-len 30
+ldclust --help
+```
+
+`<prefix>` names `prefix.ld.h5` + `prefix.snplist` (an r² matrix with
+SNP ids; pandas-fixed and plain h5py layouts both load); the `gmm` /
+`dpgmm` methods additionally read `<prefix>.gmm.raw` (PLINK `--recodeA`
+dosages) to build the relationship-matrix embedding.
+
+Library:
+
+```python
+import ldclust as ld
+
+ids, r2, layout = ld.load_ld(prefix)          # SNP ids + r2 matrix
+labels, _, sec = ld.cluster_hc(ld.distances(r2, "d1"),
+                               deep_split=2)  # WGCNA-style, dynamic tree cut
+ld.write_hlist("blocks.hlist", ids, labels)   # PLINK --hap input
+assoc = ld.run_hap_assoc(prefix, "blocks.hlist", "run1",
+                         timeout_sec=600)     # PLINK 1.07 --hap-assoc
+selected = ld.select_blocks(ld.parse_assoc_hap(assoc))  # OMNIBUS p <= 5e-6
+```
+
+## Methods (25 families)
+
+| substrate | methods |
+|---|---|
+| precomputed distances (1 − r², √(1 − r²)) | hdbscan, optics, dbscan |
+| spectral embeddings of r² / relationship matrix | soptics, spectral (NJW), gmm, dpgmm, kmeans, minibatch |
+| weighted r² graph | fof (percolation), louvain, leiden, cnm, mcl, lpa, cw (Chinese Whispers), walktrap, infomap, sbm (DC/nested, MDL) |
+| raw r² similarity | ap (affinity propagation), pam (similarity k-medoids), cdhit (CD-HIT-style seed-and-recruit) |
+| hierarchical | hc (average/complete linkage + dynamic tree cut, WGCNA recipe) |
+| ordered SNPs | dpblocks (Zhang et al. 2002 DP, tag-SNP cost) |
+| ensembles | consensus (co-association: DTC on 1 − C or percolation on C) |
+
+Noise label convention: −1 where a method defines it (hdbscan family,
+dynamic tree cut); all other methods assign every SNP. Every wrapper is
+deterministic unless documented (gmm/dpgmm/sbm/minibatch are seeded).
+
+## Development
+
+```bash
+python3 tests/test_smoke.py    # synthetic-matrix smoke tests, no PLINK
+```
+
+Author: Gennady Khvorykh (`info@inzilico.com`).

ldclust-0.2.0/ldclust/__init__.py

@@ -0,0 +1,20 @@
+"""ldclust: LD-based SNP clustering + haplotype-testing protocol."""
+
+from .library import (PROTOCOL_P, DISTANCES, load_ld, load_ld_h5, load_ids,
+                      drop_nan_snps, distances, spectral_embedding,
+                      cluster_hdbscan, cluster_optics, cluster_soptics,
+                      cluster_dbscan, cluster_fof, cluster_louvain,
+                      spectral_eig, cluster_spectral, cluster_ap,
+                      cluster_mcl, cluster_lpa, cluster_leiden,
+                      cluster_walktrap, cluster_infomap, cluster_pam,
+                      cluster_cnm, cluster_cw, cluster_kmeans,
+                      cluster_minibatch_kmeans, load_dosage,
+                      cross_product_matrix, cluster_gmm, cluster_hc,
+                      dynamic_tree_cut, cluster_dpblocks, cluster_sbm,
+                      cluster_cdhit, coassociation, cluster_consensus,
+                      cluster_dpgmm, write_blocks,
+                      write_hdbscan_file, write_hlist, run_hap_assoc,
+                      parse_assoc_hap, select_blocks, causal_snps,
+                      snp_positions, clustering_stats, pairwise_ari)
+
+__version__ = "0.2.0"

ldclust-0.2.0/ldclust/cli.py

@@ -0,0 +1,431 @@
+"""
+ldclust command line: run one clustering method on an LD matrix and
+write the benchmark artifacts (blocks csv, labels npy, hlist for PLINK
+--hap-assoc, and a row in <out-dir>/summary.tsv).
+
+Methods:
+  hdbscan  HDBSCAN on precomputed distances (--dist d1|d2, default d2)
+  optics   OPTICS on precomputed distances d2 (sklearn)
+  soptics  OPTICS on the spectral embedding of the r2 matrix
+  fof      friends-of-friends percolation on r2 >= --eps (no noise label)
+  dbscan   DBSCAN on precomputed distances d2 with --eps
+  louvain  modularity communities on the weighted r2 graph
+  spectral normalized-affinity spectral clustering (Ng et al. 2002)
+  ap       affinity propagation on the raw r2 similarity
+  mcl      Markov clustering on the weighted r2 graph
+  lpa      asynchronous label propagation on the r2 graph
+  leiden   Leiden RBConfiguration partition on the r2 graph
+  walktrap Walktrap communities on the r2 graph
+  infomap  Infomap two-level partition on the r2 graph
+  pam      similarity k-medoids on the raw r2
+  cnm      greedy modularity (CNM) on the r2 graph
+  cw       Chinese Whispers on the r2 graph
+  gmm      Gaussian mixture on the cross-product-matrix embedding
+  kmeans   K-Means on the Euclidean embedding of 1 - r2
+  minibatch Mini-Batch K-Means on the same embedding
+  hc       agglomerative linkage on 1-r2 + dynamic tree cut (noise = -1)
+  dpblocks dynamic-programming contiguous blocks minimizing tag SNPs
+  sbm      degree-corrected stochastic block model (needs graph-tool;
+           run inside its conda environment - see README)
+  cdhit    CD-HIT-style greedy seed-and-recruit on r2 >= --eps
+  consensus evidence-accumulation consensus of existing runs' labels
+           (--bases tag1,tag2,... --consensus dtc|fof)
+  dpgmm    Dirichlet-process GMM on the gmm embedding (--k = cap only)
+
+Usage:
+  ldclust -p <prefix> --method <name> [--eps 0.5] [--dist d2]
+      [--out-dir data/bench] [--check <labels.npy>]
+  (prefix names prefix.ld.h5 + prefix.snplist; --method hdbscan|gmm
+  also read prefix.gmm.raw via PLINK --recodeA output)
+
+Author: Gennady Khvorykh, info@inzilico.com
+Started: 2026-09-21
+"""
+
+import argparse
+import os
+import time
+
+import numpy as np
+
+from . import __version__
+from . import library as ld
+
+
+def build_parser():
+    parser = argparse.ArgumentParser(
+        prog="ldclust",
+        description="LD-based clustering runner (ldclust package)")
+    parser.add_argument("-p", "--prefix", required=True,
+                        help="/path/to/prefix of prefix.ld.h5 file")
+    parser.add_argument("--method", required=True,
+                        choices=["hdbscan", "optics", "soptics", "fof",
+                                 "dbscan", "louvain", "spectral", "ap",
+                                 "mcl", "lpa", "leiden", "walktrap",
+                                 "infomap", "pam", "cnm", "cw", "gmm",
+                                 "kmeans", "minibatch", "hc", "dpblocks",
+                                 "sbm", "cdhit", "consensus", "dpgmm"])
+    parser.add_argument("--dist", choices=["d1", "d2"], default="d2")
+    parser.add_argument("--eps", type=float, default=None,
+                        help="threshold for fof and cdhit (r2), dbscan "
+                             "(distance), consensus-fof (co-association)")
+    parser.add_argument("--min-samples", type=int, default=5)
+    parser.add_argument("--xi", type=float, default=0.05)
+    parser.add_argument("--variance", type=float, default=0.95)
+    parser.add_argument("--k", type=int, default=None,
+                        help="components: soptics embedding dim / pam "
+                             "blocks / gmm components / dpgmm cap")
+    parser.add_argument("--dim", type=int, default=50,
+                        help="embedding dimension for gmm/dpgmm")
+    parser.add_argument("--gamma", type=float, default=None,
+                        help="resolution for louvain")
+    parser.add_argument("--theta", type=float, default=None,
+                        help="eigenvalue cutoff for spectral "
+                             "(k = #(w > theta))")
+    parser.add_argument("--preference", type=float, default=None,
+                        help="self-similarity (diagonal) for ap")
+    parser.add_argument("--inflation", type=float, default=None,
+                        help="inflation for mcl")
+    parser.add_argument("--cutoff", type=float, default=0.05,
+                        help="edge cutoff on r2 for graph methods; "
+                             "consensus bases count; sbm")
+    parser.add_argument("--steps", type=int, default=4,
+                        help="random-walk length for walktrap")
+    parser.add_argument("--linkage", choices=["average", "complete"],
+                        default="average", help="agglomeration for hc")
+    parser.add_argument("--min-cluster-size", type=int, default=3,
+                        help="dynamic tree cut minimum block size for hc")
+    parser.add_argument("--deep-split", type=int, default=None,
+                        choices=[0, 1, 2, 3, 4],
+                        help="dynamic tree cut deepSplit preset for hc")
+    parser.add_argument("--tag-thr", type=float, default=None,
+                        help="r2 tag-coverage threshold for dpblocks")
+    parser.add_argument("--max-len", type=int, default=None,
+                        help="maximum block length in SNPs for dpblocks")
+    parser.add_argument("--no-deg-corr", action="store_true",
+                        help="plain SBM (no degree correction) for sbm")
+    parser.add_argument("--binary", action="store_true",
+                        help="ignore r2 weights (binary graph) for sbm")
+    parser.add_argument("--nested", action="store_true",
+                        help="nested SBM (level-0 partition) for sbm")
+    parser.add_argument("--priority", choices=["degree", "order"],
+                        default="degree", help="seed priority for cdhit")
+    parser.add_argument("--recruit", choices=["seed", "member"],
+                        default="seed", help="recruit semantics for cdhit")
+    parser.add_argument("--bases", default=None,
+                        help="comma-separated base run tags for consensus "
+                             "(labels.npy in --out-dir)")
+    parser.add_argument("--consensus", choices=["dtc", "fof"], default="dtc",
+                        help="consensus function on the co-association "
+                             "matrix")
+    parser.add_argument("--conc", type=float, default=None,
+                        help="DP weight concentration prior for dpgmm "
+                             "(default: sklearn 1/cap)")
+    parser.add_argument("--out-dir", default="data/bench")
+    parser.add_argument("--check", default=None,
+                        help="optional labels.npy to verify determinism")
+    parser.add_argument("--version", action="version",
+                        version="%(prog)s " + __version__)
+    return parser
+
+
+def main(argv=None):
+    t1 = time.time()
+    parser = build_parser()
+    args = parser.parse_args(argv)
+
+    if args.method == "fof" and args.eps is None:
+        parser.error("--eps is required for fof")
+    if args.method == "dbscan" and args.eps is None:
+        parser.error("--eps is required for dbscan")
+    if args.method == "louvain" and args.gamma is None:
+        parser.error("--gamma is required for louvain")
+    if args.method == "spectral" and args.theta is None:
+        parser.error("--theta is required for spectral")
+    if args.method == "ap" and args.preference is None:
+        parser.error("--preference is required for ap")
+    if args.method == "mcl" and args.inflation is None:
+        parser.error("--inflation is required for mcl")
+    if args.method == "leiden" and args.gamma is None:
+        parser.error("--gamma is required for leiden")
+    if args.method == "pam" and args.k is None:
+        parser.error("--k is required for pam")
+    if args.method == "gmm" and args.k is None:
+        parser.error("--k is required for gmm")
+    if args.method == "kmeans" and args.k is None:
+        parser.error("--k is required for kmeans")
+    if args.method == "minibatch" and args.k is None:
+        parser.error("--k is required for minibatch")
+    if args.method == "dpgmm" and args.k is None:
+        parser.error("--k (component cap) is required for dpgmm")
+    if args.method == "hc" and args.deep_split is None:
+        parser.error("--deep-split is required for hc")
+    if args.method == "cdhit" and args.eps is None:
+        parser.error("--eps is required for cdhit")
+    if args.method == "consensus" and not args.bases:
+        parser.error("--bases is required for consensus")
+    if (args.method == "consensus" and args.consensus == "fof"
+            and args.eps is None):
+        parser.error("--eps (co-association threshold) is required for "
+                     "fof consensus")
+    if args.method == "dpblocks" and (args.tag_thr is None
+                                      or args.max_len is None):
+        parser.error("--tag-thr and --max-len are required for dpblocks")
+
+    os.makedirs(args.out_dir, exist_ok=True)
+    name = os.path.basename(args.prefix)
+
+    ids, r2, layout = ld.load_ld(args.prefix)
+    r2, ids, n_dropped = ld.drop_nan_snps(r2, ids)
+    if n_dropped:
+        print(f"Dropped {n_dropped} SNPs with NaN")
+    print(f"Loaded {ids.size} SNPs ({layout})")
+
+    substrate = "precomputed-" + args.dist
+    if args.method == "hdbscan":
+        labels, probs, sec = ld.cluster_hdbscan(ld.distances(r2, args.dist))
+        tag = f"hdbscan.{args.dist}"
+    elif args.method == "optics":
+        labels, probs, sec = ld.cluster_optics(ld.distances(r2, "d2"),
+                                               args.min_samples, args.xi)
+        tag = "optics"
+    elif args.method == "soptics":
+        ksfx = f".emb{args.k}" if args.k else ""
+        emb_file = os.path.join(args.out_dir, f"{name}{ksfx}.embedding.npy")
+        if os.path.isfile(emb_file):
+            coords = np.load(emb_file)
+            print(f"Loaded embedding from {emb_file} (k = {coords.shape[1]})")
+        else:
+            coords, k, _ = ld.spectral_embedding(r2, args.variance, args.k)
+            np.save(emb_file, coords)
+            print(f"Built spectral embedding (k = {k}), saved to {emb_file}")
+        labels, probs, sec = ld.cluster_soptics(coords, args.min_samples,
+                                                args.xi)
+        tag = "soptics" + (f"k{args.k}" if args.k else "")
+        substrate = "spectral-embedding"
+    elif args.method == "fof":
+        labels, probs, sec = ld.cluster_fof(r2, args.eps)
+        tag = f"fof{args.eps}"
+        substrate = "r2-graph"
+    elif args.method == "louvain":
+        labels, probs, sec = ld.cluster_louvain(r2, args.gamma)
+        tag = f"louvain{args.gamma}"
+        substrate = "r2-graph"
+    elif args.method == "spectral":
+        labels, probs, sec = ld.cluster_spectral(r2, args.theta)
+        tag = f"spectral{args.theta}"
+        substrate = "normalized-affinity"
+    elif args.method == "ap":
+        labels, probs, sec = ld.cluster_ap(r2, args.preference)
+        tag = f"ap{args.preference}"
+        substrate = "r2-similarity"
+    elif args.method == "mcl":
+        labels, probs, sec = ld.cluster_mcl(r2, args.inflation, args.cutoff)
+        tag = f"mcl{args.inflation}"
+        substrate = "r2-graph"
+    elif args.method == "lpa":
+        labels, probs, sec = ld.cluster_lpa(r2, args.cutoff)
+        tag = "lpa"
+        substrate = "r2-graph"
+    elif args.method == "leiden":
+        labels, probs, sec = ld.cluster_leiden(r2, args.gamma, args.cutoff)
+        tag = f"leiden{args.gamma}"
+        substrate = "r2-graph"
+    elif args.method == "walktrap":
+        labels, probs, sec = ld.cluster_walktrap(r2, args.steps, args.cutoff)
+        tag = f"walktrap{args.steps}"
+        substrate = "r2-graph"
+    elif args.method == "infomap":
+        labels, probs, sec = ld.cluster_infomap(r2, args.cutoff)
+        tag = "infomap"
+        substrate = "r2-graph"
+    elif args.method == "pam":
+        labels, probs, sec = ld.cluster_pam(r2, args.k)
+        tag = f"pam{args.k}"
+        substrate = "r2-similarity"
+    elif args.method == "cnm":
+        labels, probs, sec = ld.cluster_cnm(r2, args.cutoff)
+        tag = "cnm"
+        substrate = "r2-graph"
+    elif args.method == "cw":
+        labels, probs, sec = ld.cluster_cw(r2, args.cutoff)
+        tag = "cw"
+        substrate = "r2-graph"
+    elif args.method == "gmm":
+        # relationship matrix from the genotype dosages, then GMM on
+        # its spectral embedding with a full per-component covariance
+        gmm_prefix = args.prefix + ".gmm"  # genotype prefix must expose .raw
+        emb_file = os.path.join(args.out_dir,
+                                f"{name}.gmm{args.dim}.embedding.npy")
+        if os.path.isfile(emb_file):
+            coords = np.load(emb_file)
+            print(f"Loaded embedding from {emb_file} (d = {coords.shape[1]})")
+        else:
+            k_grm, raw_snps = ld.cross_product_matrix(gmm_prefix)
+            if not np.array_equal(raw_snps, ids):
+                print("Cross-product SNP order differs from snplist - "
+                      "reordering")
+                pos_of = {s: i for i, s in enumerate(raw_snps)}
+                order = np.array([pos_of[s] for s in ids])
+                k_grm = k_grm[np.ix_(order, order)]
+            coords, dim_used, _ = ld.spectral_embedding(k_grm, k=args.dim)
+            del k_grm
+            np.save(emb_file, coords)
+            print(f"Built cross-product embedding (d = {dim_used}), "
+                  f"saved to {emb_file}")
+        labels, probs, sec = ld.cluster_gmm(k=args.k, dim=args.dim,
+                                            coords=coords)
+        tag = f"gmm{args.k}d{args.dim}"
+        substrate = "cross-product-embedding"
+    elif args.method == "kmeans":
+        # K-Means on the Euclidean embedding of the distance-converted LD
+        # matrix (d = 1 - r2 is squared Euclidean; classical MDS coords)
+        emb_file = os.path.join(
+            args.out_dir, f"{name}.kmeans{args.dim}.embedding.npy")
+        if os.path.isfile(emb_file):
+            coords = np.load(emb_file)
+            print(f"Loaded embedding from {emb_file} (d = {coords.shape[1]})")
+        else:
+            coords, dim_used, _ = ld.spectral_embedding(r2, k=args.dim)
+            np.save(emb_file, coords)
+            print(f"Built r2 embedding (d = {dim_used}), saved to {emb_file}")
+        labels, probs, sec = ld.cluster_kmeans(r2, args.k, args.dim)
+        tag = f"kmeans{args.k}d{args.dim}"
+        substrate = "r2-embedding"
+    elif args.method == "minibatch":
+        # Mini-Batch K-Means on the same embedding as kmeans (shared cache)
+        emb_file = os.path.join(
+            args.out_dir, f"{name}.kmeans{args.dim}.embedding.npy")
+        if os.path.isfile(emb_file):
+            coords = np.load(emb_file)
+            print(f"Loaded embedding from {emb_file} (d = {coords.shape[1]})")
+        else:
+            coords, dim_used, _ = ld.spectral_embedding(r2, k=args.dim)
+            np.save(emb_file, coords)
+            print(f"Built r2 embedding (d = {dim_used}), saved to {emb_file}")
+        labels, probs, sec = ld.cluster_minibatch_kmeans(r2, args.k, args.dim)
+        tag = f"mbk{args.k}d{args.dim}"
+        substrate = "r2-embedding"
+    elif args.method == "hc":
+        # WGCNA-style: scipy linkage on the condensed d1 = 1 - r2 (the
+        # analogue of WGCNA's 1 - |cor|), then the dynamic tree cut
+        # (Langfelder et al. 2008, tree stage) on the dendrogram
+        labels, probs, sec = ld.cluster_hc(ld.distances(r2, "d1"),
+                                           args.linkage,
+                                           args.min_cluster_size,
+                                           args.deep_split)
+        tag = f"dtc{args.deep_split}" + ("" if args.linkage == "average"
+                                         else args.linkage[0])
+        substrate = "precomputed-d1"
+    elif args.method == "dpblocks":
+        # Zhang et al. 2002 DP framework, r2-tagging cost: optimal
+        # contiguous partition of the ordered SNPs minimizing total tag
+        # SNPs (every SNP covered at r2 >= tag_thr; greedy set cover per
+        # candidate block)
+        labels, probs, sec = ld.cluster_dpblocks(r2, args.tag_thr,
+                                                 args.max_len)
+        tag = f"dp{args.tag_thr}L{args.max_len}"
+        substrate = "r2-ordered"
+    elif args.method == "sbm":
+        # degree-corrected SBM on the weighted r2 graph (Peixoto 2017 MDL,
+        # graph-tool); run this CLI inside the graph-tool environment
+        labels, probs, sec = ld.cluster_sbm(r2, args.cutoff,
+                                            deg_corr=not args.no_deg_corr,
+                                            weighted=not args.binary,
+                                            nested=args.nested)
+        tag = ("sbm" + ("nd" if args.no_deg_corr else "")
+               + ("b" if args.binary else "w") + str(args.cutoff)
+               + ("n" if args.nested else ""))
+        substrate = "r2-graph-mdl"
+    elif args.method == "cdhit":
+        # CD-HIT-style greedy seed-and-recruit: hub SNPs (weighted degree)
+        # seed blocks first, recruits need r2 >= eps to the seed (no
+        # transitive chaining - the anti-FoF)
+        labels, probs, sec = ld.cluster_cdhit(r2, args.eps, args.priority,
+                                              args.recruit)
+        tag = (f"cd{args.eps}" + ("o" if args.priority == "order" else "")
+               + ("m" if args.recruit == "member" else ""))
+        substrate = "r2-similarity"
+    elif args.method == "consensus":
+        # evidence-accumulation consensus: co-association matrix of the
+        # base partitions, then DTC on 1 - C (C is PSD like r2) or
+        # percolation
+        bases = args.bases.split(",")
+        base_labels = [np.load(os.path.join(
+            args.out_dir, f"{name}.{b}.labels.npy")) for b in bases]
+        ds = args.deep_split if args.deep_split is not None else 2
+        labels, probs, sec = ld.cluster_consensus(
+            base_labels, method=args.consensus, thr=args.eps, deep_split=ds,
+            min_cluster_size=args.min_cluster_size)
+        if args.consensus == "dtc":
+            tag = f"consdtc{ds}"
+        else:
+            tag = f"consfof{args.eps:g}"
+        substrate = "coassociation-of:" + "+".join(bases)
+    elif args.method == "dpgmm":
+        # Dirichlet-process GMM on the same cached embedding as gmm; --k is
+        # only an upper bound - the DP empties unsupported components
+        emb_file = os.path.join(
+            args.out_dir, f"{name}.gmm{args.dim}.embedding.npy")
+        if os.path.isfile(emb_file):
+            coords = np.load(emb_file)
+            print(f"Loaded embedding from {emb_file} (d = {coords.shape[1]})")
+        else:
+            gmm_prefix = args.prefix + ".gmm"
+            k_grm, raw_snps = ld.cross_product_matrix(gmm_prefix)
+            if not np.array_equal(raw_snps, ids):
+                print("Cross-product SNP order differs from snplist - "
+                      "reordering")
+                pos_of = {s: i for i, s in enumerate(raw_snps)}
+                order = np.array([pos_of[s] for s in ids])
+                k_grm = k_grm[np.ix_(order, order)]
+            coords, dim_used, _ = ld.spectral_embedding(k_grm, k=args.dim)
+            del k_grm
+            np.save(emb_file, coords)
+            print(f"Built cross-product embedding (d = {dim_used}), "
+                  f"saved to {emb_file}")
+        labels, probs, sec = ld.cluster_dpgmm(
+            coords=coords, n_components=args.k, dim=args.dim,
+            weight_concentration_prior=args.conc)
+        tag = f"dpgmm{args.k}" + (f"a{args.conc:g}" if args.conc else "adef")
+        substrate = "cross-product-embedding"
+    elif args.method == "dbscan":
+        labels, probs, sec = ld.cluster_dbscan(ld.distances(r2, "d2"),
+                                               args.eps, args.min_samples)
+        tag = f"dbscan{args.eps}"
+
+    stats = ld.clustering_stats(labels)
+    print(f"{tag}: {stats['n_blocks']} blocks, noise {stats['noise']:.2%}, "
+          f"median size {stats['size_median']}, max {stats['size_max']}, "
+          f"{sec:.0f} s")
+
+    if args.check:
+        ref = np.load(args.check)
+        if ref.size == labels.size:
+            print(f"Determinism check vs {args.check}: "
+                  f"{int((ref != labels).sum())} label differences")
+
+    ld.write_blocks(os.path.join(args.out_dir, f"{name}.{tag}.blocks.csv"),
+                    ids, labels)
+    np.save(os.path.join(args.out_dir, f"{name}.{tag}.labels.npy"), labels)
+    ld.write_hlist(os.path.join(args.out_dir, f"{name}.{tag}.hlist"),
+                   ids, labels)
+
+    row = {"method": args.method, "tag": tag, "substrate": substrate,
+           "params": f"eps={args.eps};min_samples={args.min_samples};"
+                     f"xi={args.xi};dist={args.dist}", **stats,
+           "seconds": sec}
+    summary = os.path.join(args.out_dir, "summary.tsv")
+    header = not os.path.isfile(summary)
+    with open(summary, "a") as f:
+        if header:
+            f.write("\t".join(row.keys()) + "\n")
+        f.write("\t".join(str(v) for v in row.values()) + "\n")
+    print("Appended summary row to", summary)
+    dur = time.strftime("%H:%M:%S", time.gmtime(time.time() - t1))
+    print(f"Done {tag}, time elapsed: {dur}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())