Warn when intervals fall in the cropped-out region

AUTHORS.md

@@ -9,3 +9,4 @@
 * M. Hasan Celik <celik.hasan@external.gene.com>
 * Carlo de Donno <carlo.de_donno@roche.com>
 * Johannes Hingerl <hingerl.johannes@gene.com>
+* Liudeng Zhang <zhangliudeng@gmail.com>

src/grelu/lightning/__init__.py

@@ -1120,7 +1120,7 @@ def input_intervals_to_output_bins(
         Returns:start and end indices of the output bins corresponding
             to each input interval.
         """
-        return pd.DataFrame(
+        result = pd.DataFrame(
             {
                 "start": intervals.start.apply(
                     self.input_coord_to_output_bin, args=(start_pos,)
@@ -1132,6 +1132,35 @@ def input_intervals_to_output_bins(
             }
         )
 
+        # Check for intervals in the cropped-out region
+        issues = []
+        n_negative = (result["start"] < 0).sum()
+        if n_negative > 0:
+            issues.append(
+                f"{n_negative} interval(s) have start positions in the cropped-out "
+                "region, resulting in negative output bin indices."
+            )
+        seq_len = self.data_params.get("train", {}).get("seq_len")
+        if seq_len is not None:
+            bin_size = self.data_params["train"]["bin_size"]
+            crop_len = self.model_params["crop_len"]
+            max_bins = seq_len // bin_size - 2 * crop_len
+            n_over = (result["end"] > max_bins).sum()
+            if n_over > 0:
+                issues.append(
+                    f"{n_over} interval(s) have end positions beyond the cropped-out "
+                    f"region, resulting in output bin indices exceeding the maximum "
+                    f"({max_bins})."
+                )
+        if issues:
+            warnings.warn(
+                " ".join(issues)
+                + " Consider removing these intervals or manually setting their"
+                " boundaries to 0 or the maximum number of output bins."
+            )
+
+        return result
+
 
 class LightningModelEnsemble(pl.LightningModule):
     """

tests/test_lightning.py

@@ -1,3 +1,5 @@
+import warnings
+
 import numpy as np
 import pandas as pd
 import torch
@@ -458,3 +460,16 @@ def test_input_intervals_to_output_bins():
     assert output.equals(pd.DataFrame({"start": [3], "end": [12]}))
     output = crop_bin_model.input_intervals_to_output_bins(intervals=intervals)
     assert output.equals(pd.DataFrame({"start": [0], "end": [5]}))
+
+
+def test_input_intervals_to_output_bins_cropped_warning():
+    """Warn when intervals fall in the cropped-out region."""
+    # crop_model has crop_len=3, bin_size=1, so coords 0-2 produce negative bins
+    intervals = pd.DataFrame({"chrom": ["chr1"], "start": [0], "end": [2]})
+    with warnings.catch_warnings(record=True) as w:
+        warnings.simplefilter("always")
+        output = crop_model.input_intervals_to_output_bins(intervals=intervals)
+        assert len(w) == 1
+        assert "cropped-out region" in str(w[0].message)
+    # Values are still returned (warning, not error)
+    assert (output["start"] < 0).any()