Add EnsembleMF: ensembled two-tower matrix factorization

[PascalIversen]

Adds EnsembleMF, a two-tower matrix factorization model for drug response prediction.

Basically, I found that the graph convolutional part of my earlier idea does not really drive the performance. So this is the easiest model with which I get SOTA performance >RF.

AI summary:

Predicts the cell line x drug response matrix as a dot product of latent factors. The factors are
not free parameters but the outputs of small residual MLPs, one over gene expression and one over
Morgan fingerprints, so a held-out cell line can be scored from its features alone. On top of the
dot product sit per-cell, per-drug and global bias terms, a free per-drug id embedding (every drug
is seen during training, so an id-indexed latent is learnable and captures behaviour that
fingerprints only approximate), and a small interaction head. And then we ensemble the whole thing

[Copilot]

Pull request overview

Adds a new EnsembleMF multi-drug model to drevalpy for drug response prediction, implementing an ensembled two-tower matrix factorization with feature-derived latent factors (gene expression + Morgan fingerprints) and associated bias/interaction terms. The PR also wires the model into the factory/registry, documents it in Sphinx docs, and adds regression/smoke tests.

Changes:

• Introduce drevalpy.models.EnsembleMF implementation, hyperparameters, and save/load support.
• Register EnsembleMF in MODEL_FACTORY and include it in the global model test matrix.
• Add docs + dedicated tests covering training, prediction, save/load, and unknown-id fallback behavior.

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
tests/test_ensemble_mf.py	New tests for factory registration, training/prediction, save/load round-trip, and unknown-id fallback.
tests/test_drp_model.py	Adds EnsembleMF to the factory registration assertion.
tests/models/test_global_models.py	Includes EnsembleMF in parametrized global-model tests with reduced hyperparams for speed.
drevalpy/models/EnsembleMF/hyperparameters.yaml	Adds default hyperparameter grid for EnsembleMF.
drevalpy/models/EnsembleMF/ensemble_mf.py	Implements the ensembled two-tower MF model, training loop, prediction logic, and persistence.
drevalpy/models/EnsembleMF/init.py	Exposes EnsembleMF from the subpackage.
drevalpy/models/init.py	Registers EnsembleMF in exports and MODEL_FACTORY.
docs/usage.rst	Documents EnsembleMF in the model table and updates the table header wording.
docs/drevalpy.models.rst	Adds drevalpy.models.EnsembleMF to the documented modules list.
docs/drevalpy.models.EnsembleMF.rst	New Sphinx automodule page for EnsembleMF.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Pull request overview

Copilot reviewed 10 out of 10 changed files in this pull request and generated no new comments.

Suppressed comments (3)

drevalpy/models/EnsembleMF/ensemble_mf.py:258

• Similarly, fitting the StandardScaler on mat[np.isin(...)] will error cryptically if none of the training cell_line_ids have features. Since this can happen when IDs don’t overlap, it’s better to validate the overlap and raise a targeted ValueError.

        if len(train_ids) > 0:
            self._scaler = StandardScaler().fit(mat[np.isin(cell_ids, np.unique(train_ids))])

drevalpy/models/EnsembleMF/ensemble_mf.py:383

• The automatic train/val split uses torch.randperm on the global RNG state and can also produce an empty training set when there’s only 1 training pair. This contradicts the docstring claim that seeding is scoped to a forked RNG state, and can make runs/tests nondeterministic or crash on tiny datasets.

Consider forking/seeding RNG locally for the split and handling len(y) < 2 by skipping the split (or using the same pair for both train and val).

        else:
            perm = torch.randperm(len(y), device=self.device)
            n_val = max(1, int(0.1 * len(y)))
            val = (ci[perm[:n_val]], di[perm[:n_val]], y[perm[:n_val]])
            ci, di, y = ci[perm[n_val:]], di[perm[n_val:]], y[perm[n_val:]]

drevalpy/models/EnsembleMF/ensemble_mf.py:240

• When fitting the rank reference, mat[np.isin(...)] can be empty if none of the training cell_line_ids are present in the provided feature dataset. In that case, np.sort yields an empty reference and the subsequent scaler fit will fail with a generic sklearn error before the intended "No training pairs matched..." error is raised.

Add an explicit overlap check and raise a clear ValueError here.

This issue also appears on line 257 of the same file.

            if len(train_ids) > 0:
                self._rank_reference = np.sort(mat[np.isin(cell_ids, np.unique(train_ids))], axis=0)