Overview
The AIDO.Tissue model is developed on spatial single-cell transcriptomic data. To incorporate spatial cell information, K nearest neighbor cells (K=8 in our case) are retrieved for each center cell. The center cell and neighbor cell expression vectors are concatenated as model input. Two-dimension positional rotary embedding are introduced to encode both the gene and cell information. The first dim is gene index and second dim is cell index. The overall training scheme is similar to scFoundation with an asymmetric encoder-decoder architecture.
Results
We evelauted the model on two spatial data task, including predicting niche label and cell density. Please refer to AIDO.tissue-downstream-tasks for more details. Below shows the brief description and evaluation metrics.
niche label type prediction: the task is to predict niche type of each cell given spatial expression data (in total 6 types).
| Model | F1-score |
|---|---|
| AIDO.Tissue | 0.67 |
| scFoundation | 0.53 |
| Nicheformer | 0.50 |
cell density prediction: the task is to predict neighbor cell number of a target cell given the expression profiles.
| Task | Mean absolute error | R square |
|---|---|---|
| AIDO.Tissue | 4.44 | 0.55 |
| scFoundation | 6.17 | 0.11 |
| Nicheformer | 7.08 | -0.07 |
Finetuning AIDO.Tissue for spatial single cell downstream tasks
We introduce how to finetune and evaluate our pre-trained AIDO.Tissue foundation models for downstream tasks. These tasks can be classified into the following categories:
- Cell-level classification tasks: niche label type prediction
- Cell-level regression tasks: cell density prediction
Note: All the following scripts should be run under ModelGenerator/.
Download data
The related data is deposited at https://huggingface.co/datasets/genbio-ai/tissue-downstream-tasks. Please download the data and put under ModelGenerator/downloads as cell_density or niche_type_classification. Under each sub-directory, there are three files denote different split (xx.train.h5ad, xx.val.h5ad, xx.test.h5ad).
For each .h5ad, several obs attributes should be included to reprezent the spatial (coordinate) information (like x, y), the label information (like niche_label). All the column fields will be specified in the following config.yaml file.
Note: the file scRNA_genename_and_index.tsv includes all the corresponding gene name and index in h5ad file.
Cell-level classification tasks
niche label type prediction
We fully finetune AIDO.Tissue for niche label type prediction.
Finetuning script
CUDA_VISIBLE_DEVICES=7 nohup mgen fit --config experiments/AIDO.Tissue/niche_type_classfification.yaml > logs/nohup/AIDO.Tissue.niche_type_classfification.yaml.log 2>&1 &
Note:
The filter_columns includes label column and spatial coordinate column. rename_columns keep unchanged and will be used for running.
Evaluation script
Once finished run, there will be several ckpt file under the specified output directory default_root_dir. Then we can use the ckpt to evaluate on test dataset.
CUDA_VISIBLE_DEVICES=6 nohup mgen test --config experiments/AIDO.Tissue/niche_type_classfification.yaml \
--ckpt_path ckpt_path \
> ckpt_path.pred.log 2>&1 &
Note: ckpt_path is the finetuned checkpoint path.
Cell-level regression tasks
cell density prediction
The config file is like experiments/AIDO.Tissue/cell_density_regression.yaml, all the fintuning running and evaluation are similar as classification task.
Dump embedding
We can dump embedding for a .h5ad file. The script is as:
CUDA_VISIBLE_DEVICES=3 nohup mgen predict --config experiments/AIDO.Tissue/emb.xenium.yaml > logs/nohup/AIDO.Tissue.emb.xenium.log 2>&1 &
The output file will be under specified output_dir like ./logs/emb.xenium/lightning_logs/pred_output. Each batch will be saved and a merged one will also be generated as predict_predictions.pt. The predict_predictions.pt file satcks all batches:
>>> import torch
>>> file_all = 'predict_predictions.pt'
>>> d_all = torch.load(file_all, map_location='cpu')
>>> d_all.keys()
dict_keys(['predictions', 'ids'])
>>> len(d_all['predictions']) # this equal to #sample
586
>>> len(d_all['ids']) # ids are numeric index corresponding to .h5ad file
586
>>> d_all['predictions'].shape # (B, L, D), L is max sequence length of all samples
torch.Size([586, 90, 128])
We can retrieve all the gene embedding and aggregate into cell embedding (like max pooling):
>>> d_all_maxpooling = [d_all['predictions'][i,:,:] for i in range(d_all['predictions'].shape[0])]
>>> d_all_maxpooling = [i[~torch.any(i.isnan(), dim=1)] for i in d_all_maxpooling]
>>> d_all_maxpooling = torch.cat([i.max(dim=0)[0].view(1,-1) for i in d_all_maxpooling])
>>> d_all_maxpooling.shape
torch.Size([586, 128])