Cell line transfer

[jason-lequyer]

Hi, congrats on the fascinating preprint.

I'm attempting to run cell line transfer using the script you provide but am running into dimensionality issues with he Norman data I'm trying to apply it to (see below). I was wondering if you could supply the exact Norman h5ad you used that will be compatible with this, and also if possible the exact Replogle h5ad files needed to run the k562->RPE1 transfer that is done in your preprint.

Thanks a lot!
Jason

Traceback (most recent call last):
File "/home/user/Downloads/MORPH-main/analysis/cell_line_transfer/cell_line_transfer_ctrl_finetune.py", line 197, in
y_hat, x_recon, z_mu, z_logvar = model_ft_ctrl(x, c_1, c_2)
File "/home/user/anaconda3/envs/morph/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1194, in _call_impl
return forward_call(*input, **kwargs)
File "/home/user/Downloads/MORPH-main/morph/model.py", line 154, in forward
mu, var = self.encode(x)
File "/home/user/Downloads/MORPH-main/morph/model.py", line 89, in encode
h = self.leakyrelu(self.fc1(x))
File "/home/user/anaconda3/envs/morph/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1194, in _call_impl
return forward_call(*input, **kwargs)
File "/home/user/anaconda3/envs/morph/lib/python3.10/site-packages/torch/nn/modules/linear.py", line 114, in forward
return F.linear(input, self.weight, self.bias)
RuntimeError: mat1 and mat2 shapes cannot be multiplied (32x5044 and 5000x512)
Traceback (most recent call last):
File "/home/user/Downloads/MORPH-main/analysis/cell_line_transfer/cell_line_transfer_ctrl_finetune.py", line 197, in
y_hat, x_recon, z_mu, z_logvar = model_ft_ctrl(x, c_1, c_2)
File "/home/user/anaconda3/envs/morph/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1194, in _call_impl
return forward_call(*input, **kwargs)
File "/home/user/Downloads/MORPH-main/morph/model.py", line 154, in forward
mu, var = self.encode(x)
File "/home/user/Downloads/MORPH-main/morph/model.py", line 89, in encode
h = self.leakyrelu(self.fc1(x))
File "/home/user/anaconda3/envs/morph/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1194, in _call_impl
return forward_call(*input, **kwargs)
File "/home/user/anaconda3/envs/morph/lib/python3.10/site-packages/torch/nn/modules/linear.py", line 114, in forward
return F.linear(input, self.weight, self.bias)
RuntimeError: mat1 and mat2 shapes cannot be multiplied (32x5044 and 5000x512)

[JqVicky]

Hi Jason,

Thanks so much for your message, and apologies for the delayed response!

For the transferring to the Norman dataset, you'll need to start by training a new pretrained model on a source dataset that is compatible with the Norman data. This is necessary because:

1. The two available pretrained models were trained for K562 → RPE1 transfer in the preprint using the Replogle data.
2. These models predict transcriptomics for genes that do not match those in the Norman data (see step 3 of this guide).
3. Because of this gene set mismatch, the existing models cannot be directly fine-tuned on the Norman data.

We apologize for this limitation. Here's how you can use the current setup to perform transfer learning to the Norman dataset:

1. Choose a source dataset — for example, the Replogle K562 dataset.
2. Subset both the source and Norman datasets to their overlapping genes, following demo/data_tutorial_transfer_learning.ipynb.
3. Normalize and preprocess the data as shown in demo/data_tutorial.ipynb.
4. Train a MORPH model on the source dataset to obtain a pretrained model, following the steps in the user demo.
5. Fine-tune or transfer the pretrained model to the Norman dataset using the tutorial.

Note: If you use the Replogle K562 dataset as the source, we recommend fine-tuning the model using a subset of perturbations (not just control cells), since Replogle K562 is CRISPRi (knockdown) while the Norman dataset is CRISPRa (activation).

We’ll be uploading the .h5ad files needed to reproduce the K562 → RPE1 transfer in the preprint as soon as possible and keep you updated.

Thanks,
Jiaqi