From 8439b41274cea70dc63dd131e67da52f3bb2d480 Mon Sep 17 00:00:00 2001
From: Vladimir Shitov <vladimirshitov98@gmail.com>
Date: Sun, 20 Sep 2026 15:07:17 +0200
Subject: [PATCH 1/7] senkin_tmp: expose the LightGBM compute knobs

New arguments --lgbm_learning_rate, --lgbm_n_folds, --lgbm_max_bin and
--lgbm_n_jobs (worker processes over protein targets, from the library's
train_lightgbm_kfold). The network fold count (--n_folds) is now separate
from the LightGBM one. Defaults unchanged for now.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---
 .../senkin_tmp_train/config.vsh.yaml          | 30 ++++++++++++++++++-
 .../senkin_tmp/senkin_tmp_train/script.py     | 15 ++++++++--
 2 files changed, 41 insertions(+), 4 deletions(-)

diff --git a/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml b/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
index 46642bb3..49be92e3 100644
--- a/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
+++ b/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
@@ -17,7 +17,14 @@ arguments:
   - name: "--n_folds"
     type: integer
     default: 5
-    description: Number of cross-validation folds for LightGBM and neural network training.
+    description: Number of cross-validation folds for the neural networks (one network per fold, predictions averaged).
+  - name: "--lgbm_n_folds"
+    type: integer
+    default: 5
+    description: |
+      Number of cross-validation folds for the LightGBM models, whose out-of-fold predictions are
+      features of the neural networks. Fewer folds cut the LightGBM time proportionally at a small
+      cost in the quality of these features.
   - name: "--lgbm_boost_rounds"
     type: integer
     default: 10000
@@ -40,6 +47,27 @@ arguments:
     type: integer
     default: 100
     description: TSVD components for reducing LightGBM predictions before NN input.
+  - name: "--lgbm_learning_rate"
+    type: double
+    default: 0.01
+    description: |
+      LightGBM learning rate. The original solution used 0.01 with up to 10000 rounds, which takes
+      days on the benchmark datasets; a higher rate with fewer rounds reaches the same LightGBM
+      quality in a fraction of the time.
+  - name: "--lgbm_max_bin"
+    type: integer
+    default: 63
+    description: |
+      Number of histogram bins per feature in LightGBM (library default 255). 63 halves the memory
+      of the binned whole-transcriptome inputs held by every worker process (3 instead of 6.6 GB per
+      fold on the NeurIPS 2022 CITE data) at identical validation loss and slightly faster rounds.
+  - name: "--lgbm_n_jobs"
+    type: integer
+    default: -1
+    description: |
+      Number of worker processes training different protein targets in parallel; the allocated CPUs
+      are divided among them. -1 uses one process per allocated CPU (each with one thread), which is
+      several times faster than one target at a time with all threads.
 engines:
   - type: docker
     image: nvidia/cuda:12.8.1-cudnn-runtime-ubuntu24.04
diff --git a/src/methods/senkin_tmp/senkin_tmp_train/script.py b/src/methods/senkin_tmp/senkin_tmp_train/script.py
index 0c6c1eaf..7bd8ac11 100644
--- a/src/methods/senkin_tmp/senkin_tmp_train/script.py
+++ b/src/methods/senkin_tmp/senkin_tmp_train/script.py
@@ -29,10 +29,14 @@
     "input_test_mod1":  "resources_test/task_predict_modality/openproblems_neurips2021/bmmc_cite/normal/test_mod1.h5ad",
     "output": "output_model.pkl",
     "n_folds": 5,
+    "lgbm_n_folds": 5,
     "lgbm_boost_rounds": 10000,
     "lgbm_early_stopping": 100,
     "nn_epochs": 100,
     "n_tsvd_components": 100,
+    "lgbm_learning_rate": 0.01,
+    "lgbm_max_bin": 63,
+    "lgbm_n_jobs": -1,
 }
 meta = {"name": "senkin_tmp", "resources_dir": "src/methods/senkin_tmp/senkin_tmp_train", "cpus": None}
 ## VIASH END
@@ -148,6 +152,7 @@ def _feature_names(adata):
 Y_prot_raw = to_dense(adata_prot_train.layers.get("counts", adata_prot_train.X), dtype=np.float64)
 
 folds = KFold(n_splits=par["n_folds"], shuffle=True, random_state=666)
+lgbm_folds = KFold(n_splits=par["lgbm_n_folds"], shuffle=True, random_state=666)
 n_tsvd = par["n_tsvd_components"]
 boost_rounds = par["lgbm_boost_rounds"]
 early_stop = par["lgbm_early_stopping"]
@@ -157,9 +162,12 @@ def _feature_names(adata):
 # meta["cpus"], so the threads oversubscribe and thrash -- the same class of slowdown
 # fixed for guanlab in #59. Leave the library default when cpus is unknown (local runs).
 _n_threads = meta.get("cpus")
-if _n_threads:
-    for _p in (lgbm_params_1, lgbm_params_2, lgbm_params_3, lgbm_params_4):
+for _p in (lgbm_params_1, lgbm_params_2, lgbm_params_3, lgbm_params_4):
+    if _n_threads:
         _p["num_threads"] = _n_threads
+    _p["learning_rate"] = par["lgbm_learning_rate"]
+    _p["max_bin"] = par["lgbm_max_bin"]
+lgbm_n_jobs = par["lgbm_n_jobs"]
 
 # ---------------------------------------------------------------------------
 # LightGBM — 4 models, train+test passed together (original design)
@@ -168,8 +176,9 @@ def _feature_names(adata):
 def _lgbm(X_all, Y, params, description):
     logger.info(f"Training LightGBM {description}...")
     return get_lgbm_predictions(
-        X_all[train_idx], Y, X_all[test_idx], folds, params,
+        X_all[train_idx], Y, X_all[test_idx], lgbm_folds, params,
         n_tsvd_components=n_tsvd, num_boost_round=boost_rounds, early_stopping_rounds=early_stop,
+        n_jobs=lgbm_n_jobs,
     )
 
 lgbm1_svd_all = _lgbm(X_lognorm_all, Y_prot_train, lgbm_params_1, "model 1 (log-normalized RNA -> proteins)")

From 2f6107e9fbda00cda9c45433c39cbf56722928ea Mon Sep 17 00:00:00 2001
From: Vladimir Shitov <vladimirshitov98@gmail.com>
Date: Sun, 20 Sep 2026 18:15:25 +0200
Subject: [PATCH 2/7] senkin_tmp: pin the library to the parallel LightGBM
 branch (to be replaced by main after the library release)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---
 src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml | 2 +-
 src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml   | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml b/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml
index 27151ff6..2baaa1c0 100644
--- a/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml
+++ b/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml
@@ -18,7 +18,7 @@ engines:
     image: openproblems/base_pytorch_nvidia:1
     setup:
       - type: docker
-        run: pip install --no-cache-dir --no-deps git+https://github.com/lueckenlab/senkin-tmp-cite-pred.git
+        run: pip install --no-cache-dir --no-deps git+https://github.com/lueckenlab/senkin-tmp-cite-pred.git@9826aae71d9d1584247ee21b8f39d244b0329af4
       - type: python
         packages:
           - lightgbm>=4.0
diff --git a/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml b/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
index 49be92e3..0932f48c 100644
--- a/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
+++ b/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
@@ -94,7 +94,7 @@ engines:
         github:
           - openproblems-bio/core#subdirectory=packages/python/openproblems
       - type: docker
-        run: pip install --no-cache-dir --no-deps git+https://github.com/lueckenlab/senkin-tmp-cite-pred.git
+        run: pip install --no-cache-dir --no-deps git+https://github.com/lueckenlab/senkin-tmp-cite-pred.git@9826aae71d9d1584247ee21b8f39d244b0329af4
 runners:
   - type: executable
   - type: nextflow

From dedbd432e406520389d9f620f5297c66f46bb046 Mon Sep 17 00:00:00 2001
From: Vladimir Shitov <vladimirshitov98@gmail.com>
Date: Sun, 20 Sep 2026 18:15:49 +0200
Subject: [PATCH 3/7] senkin_tmp: list joblib in the image packages (the
 library is installed with --no-deps)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---
 src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml | 1 +
 src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml   | 1 +
 2 files changed, 2 insertions(+)

diff --git a/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml b/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml
index 2baaa1c0..e4d2afca 100644
--- a/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml
+++ b/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml
@@ -24,6 +24,7 @@ engines:
           - lightgbm>=4.0
           - tensorflow>=2.12
           - scikit-learn>=1.1
+          - joblib>=1.4
           - mudata>=0.2
           - muon>=0.1
           - fast-array-utils
diff --git a/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml b/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
index 0932f48c..8ce0d648 100644
--- a/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
+++ b/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
@@ -88,6 +88,7 @@ engines:
           - jsonschema
           - lightgbm>=4.0
           - scikit-learn>=1.1
+          - joblib>=1.4
           - mudata>=0.2
           - muon>=0.1
           - fast-array-utils

From 259d6a9db3659d26095f4232eaa8f1eb31a5391c Mon Sep 17 00:00:00 2001
From: Vladimir Shitov <vladimirshitov98@gmail.com>
Date: Sun, 20 Sep 2026 19:00:39 +0200
Subject: [PATCH 4/7] senkin_tmp: pass the allocated memory to the LightGBM
 stage as its worker budget

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---
 src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml | 2 +-
 src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml   | 2 +-
 src/methods/senkin_tmp/senkin_tmp_train/script.py         | 6 ++++--
 3 files changed, 6 insertions(+), 4 deletions(-)

diff --git a/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml b/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml
index e4d2afca..d2f2268e 100644
--- a/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml
+++ b/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml
@@ -18,7 +18,7 @@ engines:
     image: openproblems/base_pytorch_nvidia:1
     setup:
       - type: docker
-        run: pip install --no-cache-dir --no-deps git+https://github.com/lueckenlab/senkin-tmp-cite-pred.git@9826aae71d9d1584247ee21b8f39d244b0329af4
+        run: pip install --no-cache-dir --no-deps git+https://github.com/lueckenlab/senkin-tmp-cite-pred.git@7f5b7977fc92b01423b7adaddfbeb1687766c69c
       - type: python
         packages:
           - lightgbm>=4.0
diff --git a/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml b/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
index 8ce0d648..c9d53455 100644
--- a/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
+++ b/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
@@ -95,7 +95,7 @@ engines:
         github:
           - openproblems-bio/core#subdirectory=packages/python/openproblems
       - type: docker
-        run: pip install --no-cache-dir --no-deps git+https://github.com/lueckenlab/senkin-tmp-cite-pred.git@9826aae71d9d1584247ee21b8f39d244b0329af4
+        run: pip install --no-cache-dir --no-deps git+https://github.com/lueckenlab/senkin-tmp-cite-pred.git@7f5b7977fc92b01423b7adaddfbeb1687766c69c
 runners:
   - type: executable
   - type: nextflow
diff --git a/src/methods/senkin_tmp/senkin_tmp_train/script.py b/src/methods/senkin_tmp/senkin_tmp_train/script.py
index 7bd8ac11..843ade42 100644
--- a/src/methods/senkin_tmp/senkin_tmp_train/script.py
+++ b/src/methods/senkin_tmp/senkin_tmp_train/script.py
@@ -38,7 +38,7 @@
     "lgbm_max_bin": 63,
     "lgbm_n_jobs": -1,
 }
-meta = {"name": "senkin_tmp", "resources_dir": "src/methods/senkin_tmp/senkin_tmp_train", "cpus": None}
+meta = {"name": "senkin_tmp", "resources_dir": "src/methods/senkin_tmp/senkin_tmp_train", "cpus": None, "memory_gb": None}
 ## VIASH END
 
 sys.path.append(meta["resources_dir"])
@@ -168,6 +168,8 @@ def _feature_names(adata):
     _p["learning_rate"] = par["lgbm_learning_rate"]
     _p["max_bin"] = par["lgbm_max_bin"]
 lgbm_n_jobs = par["lgbm_n_jobs"]
+# Let the library keep the LightGBM worker processes within the allocated memory (a tenth is left for the rest).
+lgbm_memory_budget_gb = meta["memory_gb"] * 0.9 if meta.get("memory_gb") else None
 
 # ---------------------------------------------------------------------------
 # LightGBM — 4 models, train+test passed together (original design)
@@ -178,7 +180,7 @@ def _lgbm(X_all, Y, params, description):
     return get_lgbm_predictions(
         X_all[train_idx], Y, X_all[test_idx], lgbm_folds, params,
         n_tsvd_components=n_tsvd, num_boost_round=boost_rounds, early_stopping_rounds=early_stop,
-        n_jobs=lgbm_n_jobs,
+        n_jobs=lgbm_n_jobs, memory_budget_gb=lgbm_memory_budget_gb,
     )
 
 lgbm1_svd_all = _lgbm(X_lognorm_all, Y_prot_train, lgbm_params_1, "model 1 (log-normalized RNA -> proteins)")

From 0689d27617a71b9540e5962fe8991486af439147 Mon Sep 17 00:00:00 2001
From: Vladimir Shitov <vladimirshitov98@gmail.com>
Date: Sun, 20 Sep 2026 21:13:57 +0200
Subject: [PATCH 5/7] senkin_tmp: LightGBM defaults that fit the benchmark's 8
 h on 16 CPUs

learning rate 0.1, at most 200 rounds with early stopping 20, 3 LightGBM
folds (networks keep 5), 63 histogram bins, one worker per CPU within the
allocated memory. Measured on the NeurIPS 2021/2022 CITE datasets: the
LightGBM stage keeps its per-target quality within 0.003 of a 1000-round
run, and the final predictions are unchanged compared with 300 rounds and
5 folds (2021: Pearson per cell 0.822 either way).

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---
 .../senkin_tmp_predict/config.vsh.yaml        |  2 +-
 .../senkin_tmp_train/config.vsh.yaml          | 21 +++++++++++--------
 .../senkin_tmp/senkin_tmp_train/script.py     |  8 +++----
 3 files changed, 17 insertions(+), 14 deletions(-)

diff --git a/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml b/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml
index d2f2268e..54a48c28 100644
--- a/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml
+++ b/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml
@@ -18,7 +18,7 @@ engines:
     image: openproblems/base_pytorch_nvidia:1
     setup:
       - type: docker
-        run: pip install --no-cache-dir --no-deps git+https://github.com/lueckenlab/senkin-tmp-cite-pred.git@7f5b7977fc92b01423b7adaddfbeb1687766c69c
+        run: pip install --no-cache-dir --no-deps git+https://github.com/lueckenlab/senkin-tmp-cite-pred.git@896f2657d60d67e8abf040c3e193e9d9e0f74f99
       - type: python
         packages:
           - lightgbm>=4.0
diff --git a/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml b/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
index c9d53455..1a53c9c7 100644
--- a/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
+++ b/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
@@ -20,20 +20,24 @@ arguments:
     description: Number of cross-validation folds for the neural networks (one network per fold, predictions averaged).
   - name: "--lgbm_n_folds"
     type: integer
-    default: 5
+    default: 3
     description: |
       Number of cross-validation folds for the LightGBM models, whose out-of-fold predictions are
       features of the neural networks. Fewer folds cut the LightGBM time proportionally at a small
       cost in the quality of these features.
   - name: "--lgbm_boost_rounds"
     type: integer
-    default: 10000
-    description: Maximum LightGBM boosting rounds (early stopping applies).
+    default: 200
+    description: |
+      Maximum LightGBM boosting rounds (early stopping applies). The original solution used up to
+      10000 rounds at learning rate 0.01, which takes days on the benchmark datasets; 200 rounds at
+      learning rate 0.1 give the same LightGBM quality as 1000 rounds within 0.003 (per-target
+      Pearson) and the same final predictions.
     info:
       test_default: 50
   - name: "--lgbm_early_stopping"
     type: integer
-    default: 100
+    default: 20
     description: LightGBM early stopping patience (rounds without improvement).
     info:
       test_default: 10
@@ -49,11 +53,10 @@ arguments:
     description: TSVD components for reducing LightGBM predictions before NN input.
   - name: "--lgbm_learning_rate"
     type: double
-    default: 0.01
+    default: 0.1
     description: |
-      LightGBM learning rate. The original solution used 0.01 with up to 10000 rounds, which takes
-      days on the benchmark datasets; a higher rate with fewer rounds reaches the same LightGBM
-      quality in a fraction of the time.
+      LightGBM learning rate. The original solution used 0.01 with up to 10000 rounds; 0.1 with a few
+      hundred rounds reaches the same LightGBM quality in a fraction of the time.
   - name: "--lgbm_max_bin"
     type: integer
     default: 63
@@ -95,7 +98,7 @@ engines:
         github:
           - openproblems-bio/core#subdirectory=packages/python/openproblems
       - type: docker
-        run: pip install --no-cache-dir --no-deps git+https://github.com/lueckenlab/senkin-tmp-cite-pred.git@7f5b7977fc92b01423b7adaddfbeb1687766c69c
+        run: pip install --no-cache-dir --no-deps git+https://github.com/lueckenlab/senkin-tmp-cite-pred.git@896f2657d60d67e8abf040c3e193e9d9e0f74f99
 runners:
   - type: executable
   - type: nextflow
diff --git a/src/methods/senkin_tmp/senkin_tmp_train/script.py b/src/methods/senkin_tmp/senkin_tmp_train/script.py
index 843ade42..c484e90f 100644
--- a/src/methods/senkin_tmp/senkin_tmp_train/script.py
+++ b/src/methods/senkin_tmp/senkin_tmp_train/script.py
@@ -29,12 +29,12 @@
     "input_test_mod1":  "resources_test/task_predict_modality/openproblems_neurips2021/bmmc_cite/normal/test_mod1.h5ad",
     "output": "output_model.pkl",
     "n_folds": 5,
-    "lgbm_n_folds": 5,
-    "lgbm_boost_rounds": 10000,
-    "lgbm_early_stopping": 100,
+    "lgbm_n_folds": 3,
+    "lgbm_boost_rounds": 200,
+    "lgbm_early_stopping": 20,
     "nn_epochs": 100,
     "n_tsvd_components": 100,
-    "lgbm_learning_rate": 0.01,
+    "lgbm_learning_rate": 0.1,
     "lgbm_max_bin": 63,
     "lgbm_n_jobs": -1,
 }

From da87618c3b32c8e7ca92cf97989bbcd755a87ac6 Mon Sep 17 00:00:00 2001
From: Vladimir Shitov <vladimirshitov98@gmail.com>
Date: Sun, 20 Sep 2026 22:42:19 +0200
Subject: [PATCH 6/7] senkin_tmp: free the log-normalized matrix after LightGBM
 model 1; pin the library

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---
 src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml | 2 +-
 src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml   | 2 +-
 src/methods/senkin_tmp/senkin_tmp_train/script.py         | 4 +++-
 3 files changed, 5 insertions(+), 3 deletions(-)

diff --git a/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml b/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml
index 54a48c28..b3469f38 100644
--- a/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml
+++ b/src/methods/senkin_tmp/senkin_tmp_predict/config.vsh.yaml
@@ -18,7 +18,7 @@ engines:
     image: openproblems/base_pytorch_nvidia:1
     setup:
       - type: docker
-        run: pip install --no-cache-dir --no-deps git+https://github.com/lueckenlab/senkin-tmp-cite-pred.git@896f2657d60d67e8abf040c3e193e9d9e0f74f99
+        run: pip install --no-cache-dir --no-deps git+https://github.com/lueckenlab/senkin-tmp-cite-pred.git@be8bee65c146316e579ae092c43247284ba801fe
       - type: python
         packages:
           - lightgbm>=4.0
diff --git a/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml b/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
index 1a53c9c7..c316b2d2 100644
--- a/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
+++ b/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
@@ -98,7 +98,7 @@ engines:
         github:
           - openproblems-bio/core#subdirectory=packages/python/openproblems
       - type: docker
-        run: pip install --no-cache-dir --no-deps git+https://github.com/lueckenlab/senkin-tmp-cite-pred.git@896f2657d60d67e8abf040c3e193e9d9e0f74f99
+        run: pip install --no-cache-dir --no-deps git+https://github.com/lueckenlab/senkin-tmp-cite-pred.git@be8bee65c146316e579ae092c43247284ba801fe
 runners:
   - type: executable
   - type: nextflow
diff --git a/src/methods/senkin_tmp/senkin_tmp_train/script.py b/src/methods/senkin_tmp/senkin_tmp_train/script.py
index c484e90f..a0cbd564 100644
--- a/src/methods/senkin_tmp/senkin_tmp_train/script.py
+++ b/src/methods/senkin_tmp/senkin_tmp_train/script.py
@@ -184,6 +184,8 @@ def _lgbm(X_all, Y, params, description):
     )
 
 lgbm1_svd_all = _lgbm(X_lognorm_all, Y_prot_train, lgbm_params_1, "model 1 (log-normalized RNA -> proteins)")
+del X_lognorm_all  # not needed any more; keeps the parent's footprint (and hence the memory left for workers) small
+gc.collect()
 
 X_comb_all = np.concatenate([X_clr_tsvd_all, X_raw_selected_all, X_sqrt_tsvd_all, X_sqrt_pca_all], axis=1)
 lgbm2_svd_all = _lgbm(X_comb_all, Y_prot_train, lgbm_params_2, "model 2 (CLR-TSVD + selected genes + normalized TSVD/PCA -> proteins)")
@@ -191,7 +193,7 @@ def _lgbm(X_all, Y, params, description):
 
 lgbm3_svd_all = _lgbm(X_counts_all, Y_prot_train, lgbm_params_3, "model 3 (raw counts -> proteins)")
 lgbm4_svd_all = _lgbm(X_counts_all, Y_prot_raw, lgbm_params_4, "model 4 (raw counts -> raw proteins)")
-del X_counts_all, X_lognorm_all
+del X_counts_all
 gc.collect()
 
 # ---------------------------------------------------------------------------

From 043215c2c794dd42092f9dbd75f304e8aafe482c Mon Sep 17 00:00:00 2001
From: Vladimir Shitov <vladimirshitov98@gmail.com>
Date: Mon, 21 Sep 2026 01:05:14 +0200
Subject: [PATCH 7/7] senkin_tmp: default to 100 LightGBM rounds

200 rounds barely cut the time on the 2022 data (early stopping already
ends most models around 170 rounds); 100 rounds keep the final predictions
unchanged and leave a safe margin under the 8 h limit.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---
 src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml | 8 ++++----
 src/methods/senkin_tmp/senkin_tmp_train/script.py       | 2 +-
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml b/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
index c316b2d2..ef7ec762 100644
--- a/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
+++ b/src/methods/senkin_tmp/senkin_tmp_train/config.vsh.yaml
@@ -27,12 +27,12 @@ arguments:
       cost in the quality of these features.
   - name: "--lgbm_boost_rounds"
     type: integer
-    default: 200
+    default: 100
     description: |
       Maximum LightGBM boosting rounds (early stopping applies). The original solution used up to
-      10000 rounds at learning rate 0.01, which takes days on the benchmark datasets; 200 rounds at
-      learning rate 0.1 give the same LightGBM quality as 1000 rounds within 0.003 (per-target
-      Pearson) and the same final predictions.
+      10000 rounds at learning rate 0.01, which takes days on the benchmark datasets. At learning
+      rate 0.1, 100 rounds keep the LightGBM per-target quality within 0.015 (NeurIPS 2021) / 0.003
+      (NeurIPS 2022) of 300 rounds, and the final predictions of the networks are unchanged.
     info:
       test_default: 50
   - name: "--lgbm_early_stopping"
diff --git a/src/methods/senkin_tmp/senkin_tmp_train/script.py b/src/methods/senkin_tmp/senkin_tmp_train/script.py
index a0cbd564..bc2c87e4 100644
--- a/src/methods/senkin_tmp/senkin_tmp_train/script.py
+++ b/src/methods/senkin_tmp/senkin_tmp_train/script.py
@@ -30,7 +30,7 @@
     "output": "output_model.pkl",
     "n_folds": 5,
     "lgbm_n_folds": 3,
-    "lgbm_boost_rounds": 200,
+    "lgbm_boost_rounds": 100,
     "lgbm_early_stopping": 20,
     "nn_epochs": 100,
     "n_tsvd_components": 100,