Graph Convolutional Matrix Factorization

docs/drevalpy.models.GCMF.rst

@@ -0,0 +1,10 @@
+GCMF
+============================
+
+GCMF family (GCMF, RGCMF, PGCMF, PRGCMF)
+----------------------------------------
+
+.. automodule:: drevalpy.models.GCMF.gcmf
+   :members:
+   :undoc-members:
+   :show-inheritance:

docs/drevalpy.models.rst

@@ -26,6 +26,7 @@ Implemented models
 
    drevalpy.models.DIPK
    drevalpy.models.DrugGNN
+   drevalpy.models.GCMF
    drevalpy.models.MOLIR
    drevalpy.models.PharmaFormer
    drevalpy.models.Precily

docs/usage.rst

@@ -248,7 +248,7 @@ needing separate model classes. By default they use gene expression and fingerpr
 See the sklearn model :ref:`flexible-inputs` or the SimpleNeuralNetwork :ref:`flexible-inputs-simplenn` for details.
 
 +---------------------------------+----------------------------+--------------------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
-| Model Name                      | Baseline / Published Model | Multi-Drug Model / Single-Drug Model | Description                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                |
+| Model Name                      | Baseline / Published / New | Multi-Drug Model / Single-Drug Model | Description                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                |
 +=================================+============================+======================================+============================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================================+
 | NaivePredictor                  | Baseline Method            | Multi-Drug Model                     | Most simple method. Predicts the mean response of all drugs in the training set.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           |
 +---------------------------------+----------------------------+--------------------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
@@ -302,7 +302,14 @@ See the sklearn model :ref:`flexible-inputs` or the SimpleNeuralNetwork :ref:`fl
 +---------------------------------+----------------------------+--------------------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
 | Precily                         | Published Model            | Multi-Drug Model                     | `Precily <https://doi.org/10.1038/s41467-022-33291-z>`_ from Chawla et al. Uses GSVA pathway-activity scores with SMILESVec drug embeddings. Features are concatenated and passed through multiple linear layers with ReLU and Dropout.                                                                                                                                                                                                                                                                                                                                                    |
 +---------------------------------+----------------------------+--------------------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
-
+| GCMF                            | New Model                  | Multi-Drug Model                     | Graph-convolutional matrix factorization. Predicts the response matrix as a dot product of cell-line and drug latent factors that are learned end-to-end by graph convolutions over a k-NN gene-expression similarity graph (cell lines) and a Morgan-fingerprint Tanimoto graph (drugs); each convolution smooths a node's embedding over its graph neighbours. Adds per-cell/per-drug biases and an optional MLP interaction head.                                                                                                                                                       |
++---------------------------------+----------------------------+--------------------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| RGCMF                           | New Model                  | Multi-Drug Model                     | Relational, multi-graph GCMF: several similarity or prior-knowledge graphs per side (multi-omics cell graphs; pathway and bioassay drug relations) are fused by a relational graph convolution. Strongest single variant on leave-cell-out in our experiments.                                                                                                                                                                                                                                                                                                                             |
++---------------------------------+----------------------------+--------------------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| PGCMF                           | New Model                  | Multi-Drug Model                     | Probabilistic GCMF with a heteroscedastic Gaussian negative-log-likelihood head that emits a calibrated per-prediction aleatoric uncertainty; point accuracy on par with the deterministic GCMF.                                                                                                                                                                                                                                                                                                                                                                                           |
++---------------------------------+----------------------------+--------------------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
+| PRGCMF                          | New Model                  | Multi-Drug Model                     | Probabilistic RGCMF: the relational variant combined with the heteroscedastic Gaussian-NLL uncertainty head.                                                                                                                                                                                                                                                                                                                                                                                                                                                                               |
++---------------------------------+----------------------------+--------------------------------------+--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------+
 
 Available Datasets
 ------------------

drevalpy/models/GCMF/__init__.py

@@ -0,0 +1,5 @@
+"""Module for the GCMF (Graph Convolutional Matrix Factorization) model."""
+
+from .gcmf import GCMF, PGCMF, PRGCMF, RGCMF
+
+__all__ = ["GCMF", "PGCMF", "RGCMF", "PRGCMF"]