fix(eval-labels): CodeRabbit review — splice_donor consensus + multi-parent GFF rows

- splice_donor regex now requires the terminal T (GT[AG]AGT consensus), not GT[AG]AG
- euk_windows: split comma-separated GFF Parent IDs so a shared exon/CDS attaches to every transcript

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Signed-off-by: Polina Binder <pbinder@nvidia.com>

Author: polinabinder1

bionemo-recipes/interpretability/sparse_autoencoders/recipes/evo2/scripts/euk_windows.py

@@ -67,11 +67,13 @@ def parse_gff(gff_path):
                 tx_gene[tid] = a.get("Parent", "").replace("gene:", "")
                 tx_biotype[tid] = a.get("biotype", "")
             elif typ == "exon":
-                tid = a.get("Parent", "").replace("transcript:", "")
-                tx_exon[tid].append((s, e))
+                for tid in a.get("Parent", "").replace("transcript:", "").split(","):
+                    if tid:
+                        tx_exon[tid].append((s, e))
             elif typ == "CDS":
-                tid = a.get("Parent", "").replace("transcript:", "")
-                tx_cds[tid].append((s, e))
+                for tid in a.get("Parent", "").replace("transcript:", "").split(","):
+                    if tid:
+                        tx_cds[tid].append((s, e))
     genes = {}
     for tid, gid in tx_gene.items():
         if gene_biotype.get(gid) != "protein_coding" or tx_biotype.get(tid) != "protein_coding":

bionemo-recipes/interpretability/sparse_autoencoders/recipes/evo2/scripts/labelers.py

@@ -225,7 +225,7 @@ def _kozak(ctx):
 @labeler("splice_donor", complex=True)
 def _sd(ctx):
     # 5' donor consensus GT(A/G)AGT — mark the GT
-    return _dna_mask(ctx, _starts(ctx.dna, r"GT[AG]AG"))
+    return _dna_mask(ctx, _starts(ctx.dna, r"GT[AG]AGT"))
 
 
 @labeler("splice_acceptor", complex=True)