diff --git a/.github/workflows/deepretro_tests.yml b/.github/workflows/deepretro_tests.yml
index adab4976..53192cf7 100644
--- a/.github/workflows/deepretro_tests.yml
+++ b/.github/workflows/deepretro_tests.yml
@@ -60,4 +60,3 @@ jobs:
       - name: Run tests
         run: |
           pytest deepretro/tests/ -v
-
diff --git a/deepretro/algorithms/stability_checker.py b/deepretro/algorithms/stability_checker.py
index 430afbdb..a66ef428 100644
--- a/deepretro/algorithms/stability_checker.py
+++ b/deepretro/algorithms/stability_checker.py
@@ -70,6 +70,8 @@
 * `is_valid_smiles` — quick check that a SMILES string parses.
 """
 
+from __future__ import annotations
+
 from rdkit import Chem
 from rdkit.Chem import Descriptors
 from rdkit.Chem.rdMolDescriptors import (
diff --git a/deepretro/utils/metrics.py b/deepretro/utils/metrics.py
index 625e8cb4..2dd4f832 100644
--- a/deepretro/utils/metrics.py
+++ b/deepretro/utils/metrics.py
@@ -3,7 +3,6 @@
 from typing import Sequence
 
 import numpy as np
-from sklearn.metrics import precision_recall_curve
 
 
 def find_optimal_threshold(
@@ -42,6 +41,10 @@ def find_optimal_threshold(
     >>> f1 > 0.0
     True
     """
+    # Import lazily so docs and lightweight tooling can import this module
+    # without pulling in sklearn's full scipy stack at module import time.
+    from sklearn.metrics import precision_recall_curve
+
     precision, recall, thresholds = precision_recall_curve(y_true, probabilities)
     f1_scores = 2 * (precision * recall) / (precision + recall + 1e-10)
     best_idx = np.argmax(f1_scores)
diff --git a/docs/source/conf.py b/docs/source/conf.py
index 4b657ac2..4869defa 100644
--- a/docs/source/conf.py
+++ b/docs/source/conf.py
@@ -175,11 +175,17 @@
     'diskcache',
     'dotenv',
     'jax',
+    'lightgbm',
+    'numpy',
+    'pandas',
     'PIL',
+    'rdkit',
     'rootutils',
+    'sklearn',
     'structlog',
     'tensorflow',
     'torch',
+    'xgboost',
 ]
 
 # Todo settings