Fix CI: don't require uninstalled Ensembl release 115

tests/test_frameshift_cterminus_regression.py

@@ -25,13 +25,30 @@
 yielded 73 aa ending ``...F-R-K-K-Q-N``.
 """
 
+import pytest
 from pyensembl import cached_release
 
 from varcode import Variant
 from varcode.effects import FrameShift
 
 ensembl_grch38 = cached_release(115)
 
+# This regression is pinned against the exact reported variant (ATM p.F61fs
+# on MANE Select ENST00000675843), which only exists in newer Ensembl
+# releases. Skip cleanly when release 115 isn't installed -- CI's data mirror
+# (openvax/ensembl-data) tops out at GRCh38.95, so 115 is unavailable there.
+# The same bug CLASS is exercised on an installed release (81) by
+# tests/test_annotator_divergence_scenarios.py (CFTR p.L127fs, BRCA1 p.R71fs)
+# and tests/test_protein_diff_parity.py, so CI coverage of #396/#397 does not
+# depend on release 115.
+try:
+    ensembl_grch38.transcript_by_id("ENST00000675843")
+except Exception as _exc:  # pyensembl raises if the GTF DB isn't downloaded
+    pytest.skip(
+        "Ensembl release 115 not installed (%s); ATM regression covered on "
+        "release 81 elsewhere." % type(_exc).__name__,
+        allow_module_level=True)
+
 
 def _atm_f61fs_effect(annotator):
     # VCF-style anchored representation of chr11:g.108227882delT

tests/test_timings.py

@@ -13,8 +13,16 @@
 from __future__ import print_function, division, absolute_import
 import time
 
+from pyensembl import cached_release
+
 from varcode.util import random_variants
 
+# Pin to an installed release. genome_for_reference_name("GRCh38") (the
+# random_variants default) resolves to pyensembl's latest release, which is
+# not necessarily downloaded (CI installs 75/81/95 only); release 81 matches
+# the rest of the suite.
+ensembl_grch38 = cached_release(81)
+
 def _time_variant_annotation(variant_collection):
     start_t = time.time()
     effects = variant_collection.effects()
@@ -30,12 +38,14 @@ def test_effect_timing(
         n_warmup_variants=5):
     warmup_collection = random_variants(
         n_warmup_variants,
-        random_seed=None)
+        random_seed=None,
+        ensembl=ensembl_grch38)
     warmup_collection.effects()
 
     variant_collection = random_variants(
         n_variants,
-        random_seed=random_seed)
+        random_seed=random_seed,
+        ensembl=ensembl_grch38)
     elapsed_t = _time_variant_annotation(variant_collection)
     print("Elapsed: %0.4f for %d variants" % (elapsed_t, n_variants))
     assert elapsed_t / n_variants < 0.1, \

varcode/util.py

@@ -26,13 +26,26 @@ def random_variants(
         genome_name="GRCh38",
         deletions=True,
         insertions=True,
-        random_seed=None):
+        random_seed=None,
+        ensembl=None):
     """
     Generate a VariantCollection with random variants that overlap
     at least one complete coding transcript.
+
+    Parameters
+    ----------
+    ensembl : pyensembl.EnsemblRelease, optional
+        Explicit genome to draw transcripts from. When ``None`` (default)
+        the genome is resolved from ``genome_name`` via
+        :func:`genome_for_reference_name`, which picks pyensembl's *latest*
+        release for that assembly. Pass a pinned release (e.g.
+        ``cached_release(81)``) when the caller needs deterministic data
+        that is actually installed — ``genome_for_reference_name`` can
+        resolve to a release that hasn't been downloaded.
     """
     rng = random.Random(random_seed)
-    ensembl = genome_for_reference_name(genome_name)
+    if ensembl is None:
+        ensembl = genome_for_reference_name(genome_name)
 
     if ensembl in _transcript_ids_cache:
         transcript_ids = _transcript_ids_cache[ensembl]
@@ -52,33 +65,55 @@ def random_variants(
             if not transcript.complete:
                 continue
 
-            exon = rng.choice(transcript.exons)
-            base1_genomic_position = rng.randint(exon.start, exon.end)
-            transcript_offset = transcript.spliced_offset(base1_genomic_position)
-            seq = transcript.sequence
+            try:
+                exon = rng.choice(transcript.exons)
+                base1_genomic_position = rng.randint(exon.start, exon.end)
+                transcript_offset = transcript.spliced_offset(
+                    base1_genomic_position)
+                seq = transcript.sequence
 
-            ref = str(seq[transcript_offset])
-            if transcript.on_backward_strand:
-                ref = reverse_complement(ref)
+                ref = str(seq[transcript_offset])
+                if transcript.on_backward_strand:
+                    ref = reverse_complement(ref)
 
-            alt_nucleotides = [x for x in STANDARD_NUCLEOTIDES if x != ref]
+                alt_nucleotides = [x for x in STANDARD_NUCLEOTIDES if x != ref]
 
-            if insertions:
-                nucleotide_pairs = [
-                    x + y
-                    for x in STANDARD_NUCLEOTIDES
-                    for y in STANDARD_NUCLEOTIDES
-                ]
-                alt_nucleotides.extend(nucleotide_pairs)
-            if deletions:
-                alt_nucleotides.append("")
-            alt = rng.choice(alt_nucleotides)
-            variant = Variant(
-                transcript.contig,
-                base1_genomic_position,
-                ref=ref,
-                alt=alt,
-                ensembl=ensembl)
+                if insertions:
+                    nucleotide_pairs = [
+                        x + y
+                        for x in STANDARD_NUCLEOTIDES
+                        for y in STANDARD_NUCLEOTIDES
+                    ]
+                    alt_nucleotides.extend(nucleotide_pairs)
+                if deletions:
+                    alt_nucleotides.append("")
+                alt = rng.choice(alt_nucleotides)
+                variant = Variant(
+                    transcript.contig,
+                    base1_genomic_position,
+                    ref=ref,
+                    alt=alt,
+                    ensembl=ensembl)
+                # Force the lazy contig validation NOW, through the exact path
+                # effect prediction uses: Variant.transcripts calls
+                # _check_that_genome_has_contig, which reads a process-wide
+                # valid-contig cache keyed by reference name. A variant built on
+                # an alternate/patch scaffold (e.g. 'CHR_HSCHR6_MHC_MCF_CTG1')
+                # constructs fine but raises here; skipping now guarantees we
+                # never return a variant that blows up a later .effects().
+                # Going through .transcripts -- not a separately computed contig
+                # set -- is what makes this consistent with the caller (earlier
+                # attempts compared against the wrong set and let scaffolds
+                # through).
+                overlapping = variant.transcripts
+            except ValueError:
+                # Alternate/patch-scaffold contig, or a transcript with a
+                # sequence/offset edge case: skip and draw another rather than
+                # failing the whole generator on an unlucky (often unseeded)
+                # pick.
+                continue
+            if not overlapping:
+                continue
             variants.append(variant)
         else:
             return VariantCollection(variants)