diff --git a/src/clm/commands/write_structural_prior_CV.py b/src/clm/commands/write_structural_prior_CV.py
index 7a48a018..e5c1b855 100644
--- a/src/clm/commands/write_structural_prior_CV.py
+++ b/src/clm/commands/write_structural_prior_CV.py
@@ -242,6 +242,41 @@ def write_structural_prior_CV(
     logger.info("Reading sample file from generative model")
     gen = read_csv_file(sample_file)
 
+    # some SMILES may be invalid when tabulate_molecules used a different
+    # rdkit version -- validate only generated SMILES that are candidates to
+    # match a test molecule
+    gen_sorted = gen.sort_values("mass", kind="stable")
+    gen_masses = gen_sorted["mass"].values
+    lo_vals = test["mass_range"].apply(lambda r: r[0]).values
+    hi_vals = test["mass_range"].apply(lambda r: r[1]).values
+    lefts = np.searchsorted(gen_masses, lo_vals, side="left")
+    rights = np.searchsorted(gen_masses, hi_vals, side="right")
+    candidate_positions = set()
+    for left, right in zip(lefts, rights):
+        candidate_positions.update(range(left, right))
+    candidates = gen_sorted.iloc[sorted(candidate_positions)]
+    invalid_idx = candidates.index[
+        candidates["smiles"].progress_apply(
+            lambda s: clean_mol(s, raise_error=False) is None
+        )
+    ]
+
+    gen = gen.drop(invalid_idx)
+
+    n_candidates = len(candidates)
+    n_invalid = len(invalid_idx)
+
+    # log if invalid SMILES were detected and removed
+    if n_invalid > 0:
+        examples = gen.loc[invalid_idx, "smiles"].head(5).tolist()
+
+        logger.warning(
+            f"Removed {n_invalid} invalid SMILES among "
+            f"{n_candidates} candidates to match a test molecule "
+            f"(possibly due to a different rdkit version). "
+            f"Examples: {examples}"
+        )
+
     inputs = {"model": gen.assign(source="model")}
 
     if pubchem_file: