Since the full CINEMA-OT dataset (containing all three stimulations) is now available via a dataloader (or will be as soon as the PR is merged), I would suggest adding a section on synergy in the tutorial. The example provided in the documentation demonstrates the general usage but doesn't illustrate how to downstream process the result of the synergy method.

The following code can be used for the tutorial:

adata = pt.dt.cinemaot_full()
sc.pp.pca(adata)
model = pt.tl.Cinemaot()
combo = model.synergy(adata,pert_key='perturbation',base='No stimulation',A='IFNb',B='IFNg',AB='IFNb+ IFNg',thres=0.5,smoothness=1e-5,eps=1e-3,solver='Sinkhorn')
sc.pp.pca(combo)
sc.pp.neighbors(combo)
sc.tl.umap(combo)
synergyscore = np.linalg.norm(combo.X[:,np.abs(np.mean(combo.X,axis=0))>0.15],axis=1)
sc.pl.umap(combo,color=['cell_type0528'],wspace=0.5,size=0.3*synergyscore**2)

Hello pertpy developer,

@Zethson

Thank you for developing this great package and its associated tutorials scgen_perturbation_prediction.ipynb. I run the model.train but meet the following errors. Do you know how to address them? Look forward to learning from your insights. Many thanks in advance!

model.train(max_epochs=100, batch_size=32, early_stopping=True, early_stopping_patience=25)

Traceback (most recent call last):
File "", line 1, in
File "/Users/soopo/.pyenv/versions/3.9.10/lib/python3.9/site-packages/scvi/model/base/_training_mixin.py", line 78, in train
runner = self._train_runner_cls(
File "/Users/soopo/.pyenv/versions/3.9.10/lib/python3.9/site-packages/scvi/train/_trainrunner.py", line 85, in init
self.trainer = self._trainer_cls(
File "/Users/soopo/.pyenv/versions/3.9.10/lib/python3.9/site-packages/scvi/train/_trainer.py", line 139, in init
super().init(
File "/Users/soopo/.pyenv/versions/3.9.10/lib/python3.9/site-packages/lightning/pytorch/utilities/argparse.py", line 70, in insert_env_defaults
return fn(self, **kwargs)
File "/Users/soopo/.pyenv/versions/3.9.10/lib/python3.9/site-packages/lightning/pytorch/trainer/trainer.py", line 399, in init
self._accelerator_connector = _AcceleratorConnector(
File "/Users/soopo/.pyenv/versions/3.9.10/lib/python3.9/site-packages/lightning/pytorch/trainer/connectors/accelerator_connector.py", line 157, in init
self._set_parallel_devices_and_init_accelerator()
File "/Users/soopo/.pyenv/versions/3.9.10/lib/python3.9/site-packages/lightning/pytorch/trainer/connectors/accelerator_connector.py", line 390, in _set_parallel_devices_and_init_accelerator
self._devices_flag = accelerator_cls.parse_devices(self._devices_flag)
File "/Users/soopo/.pyenv/versions/3.9.10/lib/python3.9/site-packages/lightning/pytorch/accelerators/cpu.py", line 48, in parse_devices
return _parse_cpu_cores(devices)
File "/Users/soopo/.pyenv/versions/3.9.10/lib/python3.9/site-packages/lightning/fabric/accelerators/cpu.py", line 85, in _parse_cpu_cores
raise TypeError("devices selected with CPUAccelerator should be an int > 0.")
TypeError: devices selected with CPUAccelerator should be an int > 0.

Tutorial on expression prediction evaluation
https://github.com/theislab/pertpy/blob/hackathon-metrics/pertpy/expression_prediction_evaluation.ipynb

Comments from Lukas:

• Show how these plots were generated with some of Pertpy's functions
• Explain well how one could evaluate their own method using the functions and your approach
• Make draft PR here

• Title is off
• Broken links

It defaults to the value that we're changing in the notebook

In it's current version, the Distance metrics tutorial throws several errors. Most of them can be easily fixed by implementing the following things:

• Remove verbose as a parameter from pairwise method whenever it's used in the tutorial
• Specify that 'X_pca' is the obsm_key parameter when initializing a Distance object (obsm_key='X_pca')
• Use correct distance name: There is one section titled PseudobulkDistance and a Distance object is initialized with the distance 'pseudobulk'. This distance, however, doesn't exist, hence an error is thrown. Is it possible that pseudobulk was renamed and is now called euclidean?

All those things are easily fixed, I just need conformation on the distance that is supposed to be used in the Pseudobulk section.

Since my data had four timepoints-treated conditions, I tried to use 'random_forest_regressor ' estimator which was described for continuous data in the tutorial:
augur = pt.tl.Augur("random_forest_regressor")
treat_augur=augur.load(treat,label_col='group',cell_type_col='celltype')
treat_augur, treat_augur_results = augur.predict(treat_augur, random_state=None, n_threads=4)
However, I got an error in the predicting step:
AttributeError: RandomForestRegressor has none of the following attributes: predict_proba.
What should I do?

First, the summary shows an overview over the model properties: * Number of samples/cell types * The reference cell type. * The formula used

Currently on pertpy version 0.6.0.

• cell 4: train.obs["condition"] raises KeyError: 'condition'. You actually want train.obs["label"] == "stim")
• similarly, "condition" should be replaced with "label" for the rest of the notebook
• "Let’s remove stimulated CD4T cells from both the training set. This is just for the sake of this notebook, in practice, you do not need to do this step" this is incorrect. You're removing CD4T cells because you want to generalize your stimulated condition to them.
• cell 7 does not work on cpu or non-jax gpu, raising the following:
  RuntimeError: Backend 'cuda' failed to initialize: module 'jaxlib.xla_extension' has no attribute 'GpuAllocatorConfig'. Available backends are ['cpu']
  You need to add the parameter accelerator='cpu'
• cell type names have been changed, specifically from "CD4T" to "CD4 T cells"