Fix novel on cells without counts in the hvg peaks

CHANGELOG.md

@@ -112,6 +112,8 @@
 
 * `senkin_tmp_train`, `senkin_tmp_predict`: Store the model bundle (the test predictions, as the solution is transductive) as an h5ad file in a directory instead of a pickle. The train image is built from `nvidia/cuda` with the current pandas 3, the predict image `openproblems/base_pytorch_nvidia:1` ships pandas 2, and unpickling the bundled `var` DataFrame raised `NotImplementedError` in `NDArrayBacked.__setstate__`, so senkin_tmp finished both CITE training runs of run_2026-09-23 but never produced a prediction. The predict step still reads the old pickle until the test resources are regenerated (PR #73).
 
+* `novel`: Guard the TF-IDF and LSI of the ATAC input against cells without counts in the selected peaks. These turned into NaN, so the validation loss never improved, no model was saved and `novel_predict` failed with a missing `tensor.pt` on the multiome swap datasets (PR #76).
+
 * `senkin_tmp_predict`: Compare the cell and protein names of the stored predictions as lists. The train and predict images run different pandas versions, so the same names were read back with a different index dtype and `Index.equals()` failed on every dataset (PR #75).
 
 * `scipenn`: Exit 99 (non-applicable) instead of raising a `ValueError` on anything other than GEX -> ADT (PR #77).

src/methods/novel/helper_functions.py

@@ -29,11 +29,13 @@ def fit(self, X):
     def transform(self, X):
         if not self.fitted:
             raise RuntimeError('Transformer was not fitted on any data')
+        row_sums = np.asarray(X.sum(axis=1)).reshape(-1, 1)
+        row_sums[row_sums == 0] = 1
         if scipy.sparse.issparse(X):
-            tf = X.multiply(1 / X.sum(axis=1))
+            tf = X.multiply(1 / row_sums)
             return tf.multiply(self.idf)
         else:
-            tf = X / X.sum(axis=1, keepdims=True)
+            tf = X / row_sums
             return tf * self.idf
 
     def fit_transform(self, X):
@@ -75,7 +77,9 @@ def transform(self, adata):
         X_norm = np.log1p(X_norm * 1e4)
         X_lsi = self.pcaTransformer.transform(X_norm)
         X_lsi -= X_lsi.mean(axis=1, keepdims=True)
-        X_lsi /= X_lsi.std(axis=1, ddof=1, keepdims=True)
+        X_lsi_std = X_lsi.std(axis=1, ddof=1, keepdims=True)
+        X_lsi_std[X_lsi_std == 0] = 1
+        X_lsi /= X_lsi_std
         lsi_df = pd.DataFrame(X_lsi, index = adata_use.obs_names)
         return lsi_df