feat: make the FlexCNN multitask model the default (4.1.1)

The fused-trunk multitask model trained across 6,543 LC setups, bundled since
4.1.0 as an opt-in, is now what every core function and the command line load
when no model is given. The 4.0 default stays bundled as
`deeplc.core.LEGACY_MULTITASK_MODEL` so existing workflows can pin it.

An uncalibrated `predict()` on a multitask model that carries setup names now
reports the setup named by `DEFAULT_TASK_NAME` (`PXD005573_mcp`, the 200-minute
gradient the DeepLC 1.x to 3.x models were trained on) rather than head 0, which
for the new default was whichever setup sorted first. Calibration, fine-tuning
and `return_matrix=True` are unchanged.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: RobbinBouwmeester

CHANGELOG.md

@@ -6,6 +6,21 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to
 [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [4.1.1] - 2026-08-26
+
+### Changed
+
+- The default model is now the fused-trunk multitask model trained across 6,543 LC
+  setups (`multitask_flexcnn_model.pt`, bundled since 4.1.0 as an opt-in). Every core
+  function and the command line use it when no `model` is given. The 4.0 default,
+  `multitask_model.pt`, stays bundled as `deeplc.core.LEGACY_MULTITASK_MODEL`; pass it
+  as `model=` to reproduce 4.0 and 4.1.0 predictions exactly.
+- Uncalibrated `predict()` on a multitask model that carries setup names reports the
+  setup named by `deeplc.core.DEFAULT_TASK_NAME` (`PXD005573_mcp`, the 200-minute
+  gradient the DeepLC 1.x to 3.x models were trained on) instead of head 0, which for
+  the new default was an arbitrary setup. `return_matrix=True` is unchanged, and so are
+  calibration and fine-tuning, which select or fit the setup from the reference.
+
 ## [4.1.0] - 2026-08-20
 
 ### Added

README.md

@@ -16,8 +16,8 @@
 
 DeepLC predicts retention times for peptides carrying any modification. It does this by leveraging 
 a deep learning model based on atomic composition features. Starting with v4, DeepLC comes with a
-multitask pretrained model covering multiple LC setups, enabling accurate predictions out of the
-box. For best results on a specific dataset, predictions can be calibrated or fine-tuned
+multitask pretrained model covering multiple LC setups (6,543 setups since v4.1.1), enabling
+accurate predictions out of the box. For best results on a specific dataset, predictions can be calibrated or fine-tuned
 using a small reference set of identified PSMs.
 
 ## Citation

deeplc/core.py

@@ -22,7 +22,20 @@
 LOGGER = logging.getLogger(__name__)
 
 DEEPLC_DIR = Path(__file__).resolve().parent
-DEFAULT_MODEL = DEEPLC_DIR / "package_data" / "models" / "multitask_model.pt"
+#: The model every core function uses when none is given: the fused-trunk multitask
+#: model trained across 6,543 LC setups (see :data:`FLEXCNN_MULTITASK_MODEL`).
+DEFAULT_MODEL = DEEPLC_DIR / "package_data" / "models" / "multitask_flexcnn_model.pt"
+
+#: The 4.0 default, a shared-trunk model with one head per LC setup, kept so that
+#: existing workflows can pin it: ``predict(psms, model=LEGACY_MULTITASK_MODEL)``.
+LEGACY_MULTITASK_MODEL = DEEPLC_DIR / "package_data" / "models" / "multitask_model.pt"
+
+#: The LC setup an uncalibrated ``predict()`` reports for a multitask model. The
+#: default model has 6,543 setups and no reference to choose between them, so the
+#: setup of the DeepLC 1.x to 3.x training data (PXD005573, 200-minute gradient) is
+#: used, which keeps uncalibrated output on the gradient earlier versions reported.
+#: Calibration and fine-tuning pick or fit the setup from the reference instead.
+DEFAULT_TASK_NAME = "PXD005573_mcp"
 
 #: Below this many reference PSMs, fine-tuning measured worse than calibration on
 #: every held-out setup tried, so it is warned about rather than silently attempted.
@@ -40,10 +53,9 @@
 #: near 1 %.
 MAX_FINETUNE_ERROR_FRACTION = 0.15
 
-#: Fused-trunk multitask model, trained across 6,543 LC setups. Not the default:
-#: switching would change every prediction, so the choice is left to the caller
-#: until the calibration path is adapted to its low-rank head.
-FLEXCNN_MULTITASK_MODEL = DEEPLC_DIR / "package_data" / "models" / "multitask_flexcnn_model.pt"
+#: Fused-trunk multitask model, trained across 6,543 LC setups. The default since
+#: 4.1.1; the name is kept for callers that pass it explicitly.
+FLEXCNN_MULTITASK_MODEL = DEFAULT_MODEL
 
 
 def predict(
@@ -65,8 +77,10 @@ def predict(
         Additional keyword arguments to pass to the prediction function.
     return_matrix
         If True, return the full prediction matrix of shape ``(n, n_heads)`` when using a
-        multitask model. If False (default), return a 1D array of shape ``(n,)`` using
-        head 0 when model output is 2D.
+        multitask model. If False (default), return a 1D array of shape ``(n,)`` for the
+        setup named by :data:`DEFAULT_TASK_NAME` when the model knows its setups, and head
+        0 otherwise. Uncalibrated output is on that setup's gradient; use
+        :func:`predict_and_calibrate` to map it onto your own.
 
     Returns
     -------
@@ -96,7 +110,7 @@ def predict(
         and "task_idx" not in kwargs
         and _model_ops.supports_task_subset(loaded_model)
     ):
-        kwargs["task_idx"] = [0]
+        kwargs["task_idx"] = [_default_task_idx(loaded_model)]
 
     result = _model_ops.predict(
         model=loaded_model,
@@ -106,10 +120,24 @@ def predict(
         **kwargs,
     ).numpy()
     if not return_matrix:
-        return result[:, 0]
+        return result[:, 0 if "task_idx" in kwargs else _default_task_idx(loaded_model)]
     return result
 
 
+def _default_task_idx(model: torch.nn.Module) -> int:
+    """
+    Index of the setup an uncalibrated prediction reports.
+
+    :data:`DEFAULT_TASK_NAME` when the model carries setup names and lists it, else 0.
+    A model without names, or a model of a single setup, has nothing to choose from.
+    """
+    names = getattr(model, "task_names", None) or []
+    try:
+        return list(names).index(DEFAULT_TASK_NAME)
+    except ValueError:
+        return 0
+
+
 def calibrate(
     psm_list_reference: PSMList,
     model: torch.nn.Module | PathLike | str | None = None,

docs/source/migration.rst

@@ -60,9 +60,10 @@ Model checkpoints
 =================
 
 Legacy ``.hdf5`` checkpoints from v3 are not compatible with v4. The bundled model
-has been retrained as a PyTorch multitask model (``multitask_model.pt``). Custom
-``.hdf5`` checkpoints cannot be loaded; retrain using the v4 API. The new model
-and should serve as an ideal starting point for fine-tuning to any custom setup.
+has been retrained as a PyTorch multitask model (``multitask_flexcnn_model.pt``
+since 4.1.1, covering 6,543 LC setups; ``multitask_model.pt`` in 4.0 and 4.1.0).
+Custom ``.hdf5`` checkpoints cannot be loaded; retrain using the v4 API. The new
+model should serve as an ideal starting point for fine-tuning to any custom setup.
 
 
 Backend: TensorFlow → PyTorch

docs/source/models.rst

@@ -5,11 +5,22 @@ Prediction models
 Default model
 =============
 
-DeepLC 4.0 ships a pretrained multitask model (``multitask_model.pt``) as the
-default. This model was trained jointly across multiple LC setups and outputs
-one retention time prediction per setup. The best-fitting output head is
-selected automatically during calibration based on Pearson correlation to the
-observed retention times in the reference set.
+DeepLC ships a pretrained multitask model as the default. Since 4.1.1 this is
+``multitask_flexcnn_model.pt``: a fused-trunk convolutional model with a low-rank
+multitask head, trained jointly across 6,543 LC setups from public repositories.
+It outputs one retention time prediction (in minutes) per setup. The best-fitting
+setup is selected automatically during calibration based on Pearson correlation to
+the observed retention times in the reference set, and fine-tuning fits a new setup
+head (66 parameters) on the reference with the trunk frozen.
+
+Without calibration, :func:`deeplc.predict` reports the setup named by
+:data:`deeplc.core.DEFAULT_TASK_NAME` (``PXD005573_mcp``, the 200-minute gradient
+that DeepLC 1.x to 3.x models were trained on), or the full matrix with
+``return_matrix=True``. The setup names are available as ``model.task_names``.
+
+The 4.0 default, ``multitask_model.pt`` (shared trunk, one head per setup), stays
+bundled as :data:`deeplc.core.LEGACY_MULTITASK_MODEL` and can be passed as
+``model=`` to any core function to reproduce 4.0 and 4.1.0 predictions.
 
 Training a model from scratch
 ==============================

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "deeplc"
-version = "4.1.0"
+version = "4.1.1"
 description = "DeepLC: Retention time prediction for (modified) peptides using Deep Learning."
 readme = "README.md"
 license = { file = "LICENSE" }

tests/test_core.py

@@ -71,6 +71,45 @@ def test_predict_returns_matrix_when_flag_set():
     assert result.shape[1] > 1
 
 
+def test_default_model_is_the_flexcnn_multitask_model():
+    """Since 4.1.1 the fused-trunk model of 6,543 setups is what a bare call loads."""
+    assert deeplc.core.DEFAULT_MODEL == deeplc.core.FLEXCNN_MULTITASK_MODEL
+    assert deeplc.core.DEFAULT_MODEL.name == "multitask_flexcnn_model.pt"
+    assert deeplc.core.LEGACY_MULTITASK_MODEL.name == "multitask_model.pt"
+    assert deeplc.core.LEGACY_MULTITASK_MODEL.exists()
+
+
+def test_uncalibrated_predict_reports_the_default_setup():
+    """
+    A bare ``predict`` returns the column of :data:`DEFAULT_TASK_NAME`, not head 0.
+
+    The default model lists thousands of setups, and head 0 is whichever sorted first.
+    The PXD005573 setup keeps uncalibrated output on the gradient DeepLC 1.x to 3.x
+    reported, so downstream code that never calibrated sees comparable numbers.
+    """
+    model = deeplc.core._model_ops.load_model(deeplc.core.DEFAULT_MODEL, device="cpu")
+    idx = list(model.task_names).index(deeplc.core.DEFAULT_TASK_NAME)
+    assert idx != 0
+
+    psm_list = _make_psm_list(_PEPTIDES)
+    single = deeplc.core.predict(psm_list, predict_kwargs={"device": "cpu"})
+    matrix = deeplc.core.predict(psm_list, return_matrix=True, predict_kwargs={"device": "cpu"})
+    np.testing.assert_allclose(single, matrix[:, idx], rtol=1e-5, atol=1e-4)
+    assert not np.allclose(single, matrix[:, 0])
+    assert np.isfinite(single).all()
+
+
+def test_legacy_multitask_model_still_loads_and_predicts():
+    """The 4.0 default remains bundled and usable when pinned explicitly."""
+    result = deeplc.core.predict(
+        _make_psm_list(_PEPTIDES),
+        model=deeplc.core.LEGACY_MULTITASK_MODEL,
+        predict_kwargs={"device": "cpu"},
+    )
+    assert result.shape == (len(_PEPTIDES),)
+    assert np.isfinite(result).all()
+
+
 def test_predict_and_calibrate_auto_selects_reference():
     # 200 PSMs cycling through _PEPTIDES; 100 with qvalue<=0.01, 100 with qvalue=1.0.
     # auto-selection picks the 100 low-qvalue PSMs as reference.

tests/test_flexcnn.py

@@ -549,8 +549,10 @@ def test_bundled_model_predicts_in_minutes():
     assert -100.0 < out.min() < 60.0
     assert 20.0 < out.max() < 1000.0
 
+    # Uncalibrated output reports the DeepLC 1.x to 3.x setup, not whichever sorted first.
+    default_idx = list(model.task_names).index(core.DEFAULT_TASK_NAME)
     single = core.predict(peptides, model=path)
-    np.testing.assert_allclose(single, out[:, 0], rtol=1e-5)
+    np.testing.assert_allclose(single, out[:, default_idx], rtol=1e-5)
 
 
 def test_small_reference_set_warns_and_widens_validation(tmp_path, caplog):

tests/test_model_ops.py

@@ -9,7 +9,7 @@
 from torch.utils.data import Dataset
 
 from deeplc._architecture import DeepLCModel
-from deeplc.core import DEFAULT_MODEL
+from deeplc.core import LEGACY_MULTITASK_MODEL
 from deeplc.data import split_datasets
 
 
@@ -59,15 +59,15 @@ def test_train_rejects_empty_validation_loader():
 
 
 @pytest.mark.skipif(
-    not DEFAULT_MODEL.exists(),
+    not LEGACY_MULTITASK_MODEL.exists(),
     reason="multitask model not bundled",
 )
 def test_load_multitask_model_without_prior_shim():
     """multitask_model.pt must load even when the legacy module is not pre-registered."""
     # Remove any previously registered shim so the test is self-contained.
     sys.modules.pop("multitask_model", None)
 
-    model = _model_ops.load_model(DEFAULT_MODEL, device="cpu")
+    model = _model_ops.load_model(LEGACY_MULTITASK_MODEL, device="cpu")
 
     assert isinstance(model, DeepLCModel)