Usage on proteins¶

Although Novae was mainly designed for spatial transcriptomics, it can also be used for spatial proteomics.

The main difference with the normal tutorial is that we will not be able to use the pretrained model, so we'll re-train a Novae model from scratch.

In [1]:

import novae

import novae

Loading the data objects¶

Similarly as in the other tutorials, we need to load one or multiple AnnData object(s). Here, we load the first 2 head&neck slides which have protein information.

See here for more details on all four possible input modes for Novae.

In [4]:

adatas = novae.utils.load_dataset(
    top_k=2,
    tissue="head_and_neck",
    custom_filter=lambda df: df["n_proteins"] > 0,
)

adatas = novae.utils.load_dataset( top_k=2, tissue="head_and_neck", custom_filter=lambda df: df["n_proteins"] > 0, )

[INFO] (novae.utils._data) Found 2 h5ad file(s) matching the filters.

10H064210813H0208914_down.h5ad:   0%|          | 0.00/230M [00:00<?, ?B/s]

10H064210813H0208914_up.h5ad:   0%|          | 0.00/184M [00:00<?, ?B/s]

In [5]:

adatas

adatas

Out[5]:

[AnnData object with n_obs × n_vars = 140139 × 63
     obs: 'annot_level1', 'ID', 'annot_level0', 'novae_sid'
     uns: 'annot_level0_colors', 'annot_level1_colors', 'spatial_neighbors'
     obsm: 'X_pca', 'X_umap', 'spatial'
     layers: 'counts'
     obsp: 'spatial_connectivities', 'spatial_distances',
 AnnData object with n_obs × n_vars = 112155 × 63
     obs: 'annot_level1', 'ID', 'annot_level0', 'novae_sid'
     uns: 'annot_level0_colors', 'annot_level1_colors', 'spatial_neighbors'
     obsm: 'X_pca', 'X_umap', 'spatial'
     layers: 'counts'
     obsp: 'spatial_connectivities', 'spatial_distances']

Compute the neighbors graph¶

Again, this step is similar to the spatial transcriptomics tutorial. The only difference is that, in our specific case, the coordinate system is in pixels (not microns), so radius is much higher that what we usually set.

Please adjust radius based on your coordinate system, and make sure the plot below looks correct.

In [23]:

novae.utils.spatial_neighbors(adatas, radius=300)

novae.utils.spatial_neighbors(adatas, radius=300)

[INFO] (novae.utils._build) Computing graph on 140,139 cells (coord_type=generic, delaunay=True, radius=[0.0, 300.0], n_neighs=None)
[INFO] (novae.utils._build) Computing graph on 112,155 cells (coord_type=generic, delaunay=True, radius=[0.0, 300.0], n_neighs=None)

We can show the graph of connectivities, which is a good quality control. Nodes in red are cells that are connected to very few other cells (ngh_threshold=2, by default). In particular:

1. You should have a relatively low amount of "red" cells. If so, decrease the radius parameter.
2. Regions that are far apart should not be connected. If so, increase the radius parameter.

If the graph is not looking right, check the docs of novae.utils.spatial_neighbors to adapt it.

In [24]:

novae.plot.connectivities(adatas)

novae.plot.connectivities(adatas)

Preprocessing¶

This is the first main difference. For spatial proteomics, we recommend using the preprocessing below.

Check the docs of novae.utils.quantile_scaling for more details.

In [8]:

novae.settings.auto_preprocessing = False # prevent Novae from running normalize_total / log1p

novae.utils.quantile_scaling(adatas)

novae.settings.auto_preprocessing = False # prevent Novae from running normalize_total / log1p novae.utils.quantile_scaling(adatas)

Training a new model¶

Since we have here 63 proteins, we will use 62 (n_proteins - 1) as an embedding size.

In [20]:

embedding_size = adatas[0].n_vars - 1
embedding_size

embedding_size = adatas[0].n_vars - 1 embedding_size

Out[20]:

62

We create a new Novae model. Since it doesn't have any embedding for the proteins, it will run a PCA to initialize the embeddings of the proteins.

In [12]:

model = novae.Novae(adatas, embedding_size=embedding_size)

model = novae.Novae(adatas, embedding_size=embedding_size)

[INFO] (novae.module.embed) Running PCA embedding initialization

Then, we train it (here, 4 epochs only).

In [13]:

model.fit(max_epochs=4)

model.fit(max_epochs=4)

GPU available: True (mps), used: False
TPU available: False, using: 0 TPU cores
HPU available: False, using: 0 HPUs
/Users/quentinblampey/Library/Caches/pypoetry/virtualenvs/novae-ezkWKrh6-py3.10/lib/python3.10/site-packages/lightning/pytorch/trainer/setup.py:177: GPU available but not used. You can set it by doing `Trainer(accelerator='gpu')`.

  | Name          | Type              | Params | Mode 
------------------------------------------------------------
0 | cell_embedder | CellEmbedder      | 7.8 K  | train
1 | encoder       | GraphEncoder      | 114 K  | train
2 | augmentation  | GraphAugmentation | 0      | train
3 | swav_head     | SwavHead          | 16.4 K | train
------------------------------------------------------------
138 K     Trainable params
0         Non-trainable params
138 K     Total params
0.555     Total estimated model params size (MB)
76        Modules in train mode
0         Modules in eval mode
/Users/quentinblampey/Library/Caches/pypoetry/virtualenvs/novae-ezkWKrh6-py3.10/lib/python3.10/site-packages/lightning/pytorch/trainer/connectors/data_connector.py:424: The 'train_dataloader' does not have many workers which may be a bottleneck. Consider increasing the value of the `num_workers` argument` to `num_workers=9` in the `DataLoader` to improve performance.

Training: |          | 0/? [00:00<?, ?it/s]

`Trainer.fit` stopped: `max_epochs=4` reached.

We compute the cells latent space.

In [14]:

model.compute_representations()

model.compute_representations()

Computing representations:   0%|          | 0/548 [00:00<?, ?it/s]

Computing representations:   0%|          | 0/439 [00:00<?, ?it/s]

And we assign the domains.

In [15]:

model.assign_domains()

model.assign_domains()

Out[15]:

'novae_domains_7'

Now, we can show the domains:

In [17]:

novae.plot.domains(adatas, cell_size=100)

novae.plot.domains(adatas, cell_size=100)

[INFO] (novae.utils._validate) Using obs_key='novae_domains_7' by default.