tcrdist draft version implemented

CHANGELOG.md

@@ -8,6 +8,10 @@ and this project adheres to [Semantic Versioning][].
 [keep a changelog]: https://keepachangelog.com/en/1.0.0/
 [semantic versioning]: https://semver.org/spec/v2.0.0.html
 
+## Unreleased
+
+-   Add "TCRdist" as new metric ([#502](https://github.com/scverse/scirpy/pull/502))
+
 ## v0.16.1
 
 ### Fixes

docs/api.rst

@@ -302,3 +302,4 @@ distance metrics
    ir_dist.metrics.HammingDistanceCalculator
    ir_dist.metrics.AlignmentDistanceCalculator
    ir_dist.metrics.FastAlignmentDistanceCalculator
+   ir_dist.metrics.TCRdistDistanceCalculator

docs/tutorials/tutorial_3k_tcr.ipynb

@@ -1049,13 +1049,19 @@
     "For instance, a distance of `10` is equivalent to 2 Rs mutating into N.\n",
     "This appoach was initially proposed as *TCRdist* by Dash et al. {cite}`TCRdist`.\n",
     "\n",
-    ":::{tip}\n",
-    "You can use `metric=\"fastalignment\"` for a faster calculation at the cost of a few false-negatives (i.e. sequence pairs\n",
-    "that are actually below the distance cutoff, but are removed during a pre-filtering step). With default parameters, \n",
-    "the false-negative rate (of all sequence pairs actually below the cutoff) was ~2% on the {func}`scirpy.datasets.wu2020`\n",
-    "dataset. \n",
-    "\n",
-    "See also {class}`scirpy.ir_dist.metrics.FastAlignmentDistanceCalculator`. \n",
+    ":::{admonition} Speeding up TCR distance calculation\n",
+    ":class: tip\n",
+    "\n",
+    "Scirpy provides alternative distance metrics that are similar to `\"alignment\"`, but a lot faster: \n",
+    "\n",
+    "* `metric=\"tcrdist\"` is an implementation of [tcrdist3](https://github.com/kmayerb/tcrdist3) within scirpy. The scores\n",
+    "  are calculated differently, but it gives very similar results compared to `metric=\"alignment\"`.\n",
+    "  See also {class}`scirpy.ir_dist.metrics.TCRdistDistanceCalculator`.\n",
+    "* `metric=\"fastalignment\"` uses a heuristic to speed up the `\"alignment\"` metric at the cost of a few false-negatives (i.e. sequence pairs\n",
+    "  that are actually below the distance cutoff, but are removed during a pre-filtering step). With default parameters, \n",
+    "  the false-negative rate (of all sequence pairs actually below the cutoff) was ~2% on the {func}`scirpy.datasets.wu2020`\n",
+    "  dataset.  See also {class}`scirpy.ir_dist.metrics.FastAlignmentDistanceCalculator`.\n",
+    "   \n",
     ":::\n",
     "\n",
     "All cells with a distance between their CDR3 sequences lower than `cutoff` will be connected in the network.\n"

src/scirpy/ir_dist/__init__.py

@@ -99,6 +99,8 @@ def _get_distance_calculator(metric: MetricType, cutoff: Union[int, None], *, n_
         dist_calc = metrics.LevenshteinDistanceCalculator(n_jobs=n_jobs, **kwargs)
     elif metric == "hamming":
         dist_calc = metrics.HammingDistanceCalculator(n_jobs=n_jobs, **kwargs)
+    elif metric == "tcrdist":
+        dist_calc = metrics.TCRdistDistanceCalculator(n_jobs=n_jobs, **kwargs)
     else:
         raise ValueError("Invalid distance metric.")
 
@@ -122,6 +124,7 @@ def _ir_dist(
     airr_mod_ref: str = "airr",
     airr_key_ref: str = "airr",
     chain_idx_key_ref: str = "chain_indices",
+    **kwargs,
 ) -> Union[dict, None]:
     """\
     Computes a sequence-distance metric between all unique :term:`VJ <Chain locus>`
@@ -171,6 +174,8 @@ def _ir_dist(
         Like `airr_key`, but for `reference`.
     chain_idx_key_ref
         Like `chain_idx_key`, but for `reference`.
+    **kwargs
+        Arguments are passed to the respective :class:`~scirpy.ir_dist.metrics.DistanceCalculator` class.
 
     Returns
     -------
@@ -227,7 +232,7 @@ def _get_unique_seqs(tmp_adata, chain_type):
                 result[chain_type][tmp_key] = unique_seqs
 
     # compute distance matrices
-    dist_calc = _get_distance_calculator(metric, cutoff, n_jobs=n_jobs)
+    dist_calc = _get_distance_calculator(metric, cutoff, n_jobs=n_jobs, **kwargs)
     for chain_type in ["VJ", "VDJ"]:
         logging.info(f"Computing sequence x sequence distance matrix for {chain_type} sequences.")  # type: ignore
         result[chain_type]["distances"] = dist_calc.calc_dist_mat(