Merge pull request #89 from Genentech/fix-ism

fix bug in ism_predict on multiple sequences
Genentech · Jan 31, 2025 · ab65884 · ab65884
2 parents 292aa93 + 23aa16f
commit ab65884
Show file tree

Hide file tree

Showing 2 changed files with 37 additions and 4 deletions.
diff --git a/src/grelu/interpret/score.py b/src/grelu/interpret/score.py
@@ -17,7 +17,7 @@
 
 
 def ISM_predict(
-    seqs: Union[pd.DataFrame, np.ndarray, str],
+    seqs: Union[pd.DataFrame, np.ndarray, str, List[str]],
     model: Callable,
     genome: Optional[str] = None,
     prediction_transform: Optional[Callable] = None,
@@ -87,11 +87,16 @@ def ISM_predict(
     )
     # B, L, 4, T, L
 
-    # Calculate log ratio w.r.t reference sequence
     if compare_func is not None:
+
+        # Slice the prediction corresponding to each reference sequence
         ref_bases = [BASE_TO_INDEX_HASH[seq[start_pos]] for seq in seqs]
-        ref_pred = preds[:, [0], [ref_bases], :]  # B, 1, 1, T, L
-        preds = get_compare_func(compare_func, tensor=False)(preds, ref_pred)
+        ref_preds = np.concatenate(
+            [preds[None, None, None, i, 0, x] for i, x in enumerate(ref_bases)]
+        )  # B, L, 1, T, L
+
+        # Compare all predictions to the prediction for the corresponding reference sequence
+        preds = get_compare_func(compare_func, tensor=False)(preds, ref_preds)
 
     # Convert into a dataframe
     if return_df:

diff --git a/tests/test_interpret.py b/tests/test_interpret.py
@@ -93,13 +93,41 @@ def test_marginalize_patterns():
 
 
 def test_ISM_predict():
+
+    # Single sequence
     seq = "AA"
     expected_preds = np.array([[4.0, 1.0, 2.0, 2.0], [4.0, 1.0, 2.0, 2.0]]).T
     preds = ISM_predict(seq, model, compare_func=None)
     assert np.allclose(preds.values, expected_preds)
     preds = ISM_predict(seq, model, compare_func="log2FC")
     assert np.allclose(preds.values, np.log2(expected_preds / 4))
 
+    # Multiple sequences
+    seqs = ["AAA", "CCC"]
+    expected_preds = np.expand_dims(
+        np.array(
+            [
+                [
+                    [4.0, 2.0, 2.6666667, 2.6666667],
+                    [4.0, 2.0, 2.6666667, 2.6666667],
+                    [4.0, 2.0, 2.6666667, 2.6666667],
+                ],
+                [
+                    [0.0, -2.0, -1.3333334, -1.3333334],
+                    [0.0, -2.0, -1.3333334, -1.3333334],
+                    [0.0, -2.0, -1.3333334, -1.3333334],
+                ],
+            ]
+        ),
+        (3, 4),
+    )
+    preds = ISM_predict(seqs, model, compare_func=None, return_df=False)
+    assert np.allclose(preds, expected_preds)
+    preds = ISM_predict(seqs, model, compare_func="log2FC", return_df=False)
+    assert np.allclose(
+        preds, np.log2(np.stack([expected_preds[0] / 4, -expected_preds[1] / 2]))
+    )
+
 
 def test_get_attributions():
     seq = generate_random_sequences(n=1, seq_len=50, seed=0, output_format="strings")[0]