base on State is a machine learning model that predicts cellular perturbation response across diverse contexts # Predicting cellular responses to perturbation across diverse contexts with State > Train State transition models or pretrain State embedding models. See the State [paper](https://arcinstitute.org/manuscripts/State). ## Associated repositories - Model evaluation framework: [cell-eval](https://github.com/ArcInstitute/cell-eval) - Dataloaders and preprocessing: [cell-load](https://github.com/ArcInstitute/cell-load) ## Installation ### Installation from PyPI This package is distributed via [`uv`](https://docs.astral.sh/uv). ```bash uv tool install arc-state ``` ### Installation from Source ```bash git clone [email protected]:ArcInstitute/state.git cd state uv run state ``` When making fundamental changes to State, install an editable version with the `-e` flag. ```bash git clone [email protected]:ArcInstitute/state.git cd state uv tool install -e . ``` ## CLI Usage You can access the CLI help menu with: ```state --help``` Output: ``` usage: state [-h] {emb,tx} ... positional arguments: {emb,tx} options: -h, --help show this help message and exit ``` ## State Transition Model (ST) To start an experiment, write a TOML file (see `examples/zeroshot.toml` or `examples/fewshot.toml` to start). The TOML file specifies the dataset paths (containing h5ad files) as well as the machine learning task. Training an ST example below. ```bash state tx train \ data.kwargs.toml_config_path="examples/fewshot.toml" \ data.kwargs.embed_key=X_hvg \ data.kwargs.num_workers=12 \ data.kwargs.batch_col=batch_var \ data.kwargs.pert_col=target_gene \ data.kwargs.cell_type_key=cell_type \ data.kwargs.control_pert=TARGET1 \ training.max_steps=40000 \ training.val_freq=100 \ training.ckpt_every_n_steps=100 \ training.batch_size=8 \ training.lr=1e-4 \ model.kwargs.cell_set_len=64 \ model.kwargs.hidden_dim=328 \ model=pertsets \ wandb.tags="[test]" \ output_dir="$HOME/state" \ name="test" ``` The cell lines and perturbations specified in the TOML should match the values appearing in the `data.kwargs.cell_type_key` and `data.kwargs.pert_col` used above. To evaluate STATE on the specified task, you can use the `tx predict` command: ```bash state tx predict --output_dir $HOME/state/test/ --checkpoint final.ckpt ``` It will look in the `output_dir` above, for a `checkpoints` folder. If you instead want to use a trained checkpoint for inference (e.g. on data not specified) in the TOML file: ```bash state tx infer --output $HOME/state/test/ --output_dir /path/to/model/ --checkpoint /path/to/model/final.ckpt --adata /path/to/anndata/processed.h5 --pert_col gene --embed_key X_hvg ``` Here, `/path/to/model/` is the folder downloaded from [HuggingFace](https://huggingface.co/arcinstitute). ## TOML Configuration Files State experiments are configured using TOML files that define datasets, training splits, and evaluation scenarios. The configuration system supports both **zeroshot** (unseen cell types) and **fewshot** (limited perturbation examples) evaluation paradigms. ### Configuration Structure #### Required Sections **`[datasets]`** - Maps dataset names to their file system paths ```toml [datasets] replogle = "/path/to/replogle/dataset/" # YOU CAN ADD MORE ``` **`[training]`** - Specifies which datasets participate in training ```toml [training] replogle = "train" # Include all replogle data in training (unless overridden below) ``` #### Optional Evaluation Sections **`[zeroshot]`** - Reserves entire cell types for validation/testing ```toml [zeroshot] "replogle.jurkat" = "test" # All jurkat cells go to test set "replogle.k562" = "val" # All k562 cells go to validation set ``` **`[fewshot]`** - Specifies perturbation-level splits within cell types ```toml [fewshot] [fewshot."replogle.rpe1"] # Configure splits for rpe1 cell type val = ["AARS", "TUFM"] # These perturbations go to validation test = ["NUP107", "RPUSD4"] # These perturbations go to test # Note: All other perturbations in rpe1 automatically go to training ``` ### Configuration Examples #### Example 1: Pure Zeroshot Evaluation ```toml # Evaluate generalization to completely unseen cell types [datasets] replogle = "/data/replogle/" [training] replogle = "train" [zeroshot] "replogle.jurkat" = "test" # Hold out entire jurkat cell line "replogle.rpe1" = "val" # Hold out entire rpe1 cell line [fewshot] # Empty - no perturbation-level splits ``` #### Example 2: Pure Fewshot Evaluation ```toml # Evaluate with limited examples of specific perturbations [datasets] replogle = "/data/replogle/" [training] replogle = "train" [zeroshot] # Empty - all cell types participate in training [fewshot] [fewshot."replogle.k562"] val = ["AARS"] # Limited AARS examples for validation test = ["NUP107", "RPUSD4"] # Limited examples of these genes for testing [fewshot."replogle.jurkat"] val = ["TUFM"] test = ["MYC", "TP53"] ``` #### Example 3: Mixed Evaluation Strategy ```toml # Combine both zeroshot and fewshot evaluation [datasets] replogle = "/data/replogle/" [training] replogle = "train" [zeroshot] "replogle.jurkat" = "test" # Zeroshot: unseen cell type [fewshot] [fewshot."replogle.k562"] # Fewshot: limited perturbation examples val = ["STAT1"] test = ["MYC", "TP53"] ``` ### Important Notes - **Automatic training assignment**: Any cell type not mentioned in `[zeroshot]` automatically participates in training, with perturbations not listed in `[fewshot]` going to the training set - **Overlapping splits**: Perturbations can appear in both validation and test sets within fewshot configurations - **Dataset naming**: Use the format `"dataset_name.cell_type"` when specifying cell types in zeroshot and fewshot sections - **Path requirements**: Dataset paths should point to directories containing h5ad files - **Control perturbations**: Ensure your control condition (specified via `control_pert` parameter) is available across all splits ### Validation The configuration system will validate that: - All referenced datasets exist at the specified paths - Cell types mentioned in zeroshot/fewshot sections exist in the datasets - Perturbations listed in fewshot sections are present in the corresponding cell types - No conflicts exist between zeroshot and fewshot assignments for the same cell type ## State Embedding Model (SE) After following the same installation commands above: ```bash state emb fit --conf ${CONFIG} ``` To run inference with a trained State checkpoint, e.g., the State trained to 4 epochs: ```bash state emb transform \ --checkpoint "/large_storage/ctc/userspace/aadduri/SE-600M" \ --input "/large_storage/ctc/datasets/replogle/rpe1_raw_singlecell_01.h5ad" \ --output "/home/aadduri/vci_pretrain/test_output.h5ad" \ ``` ## Licenses State code is [licensed](LICENSE) under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 (CC BY-NC-SA 4.0). The model weights and output are licensed under the [Arc Research Institute State Model Non-Commercial License](MODEL_LICENSE.md) and subject to the [Arc Research Institute State Model Acceptable Use Policy](MODEL_ACCEPTABLE_USE_POLICY.md). Any publication that uses this source code or model parameters should cite the State [paper](https://arcinstitute.org/manuscripts/State). ", Assign "at most 3 tags" to the expected json: {"id":"14117","tags":[]} "only from the tags list I provide: [{"id":77,"name":"3d"},{"id":89,"name":"agent"},{"id":17,"name":"ai"},{"id":54,"name":"algorithm"},{"id":24,"name":"api"},{"id":44,"name":"authentication"},{"id":3,"name":"aws"},{"id":27,"name":"backend"},{"id":60,"name":"benchmark"},{"id":72,"name":"best-practices"},{"id":39,"name":"bitcoin"},{"id":37,"name":"blockchain"},{"id":1,"name":"blog"},{"id":45,"name":"bundler"},{"id":58,"name":"cache"},{"id":21,"name":"chat"},{"id":49,"name":"cicd"},{"id":4,"name":"cli"},{"id":64,"name":"cloud-native"},{"id":48,"name":"cms"},{"id":61,"name":"compiler"},{"id":68,"name":"containerization"},{"id":92,"name":"crm"},{"id":34,"name":"data"},{"id":47,"name":"database"},{"id":8,"name":"declarative-gui "},{"id":9,"name":"deploy-tool"},{"id":53,"name":"desktop-app"},{"id":6,"name":"dev-exp-lib"},{"id":59,"name":"dev-tool"},{"id":13,"name":"ecommerce"},{"id":26,"name":"editor"},{"id":66,"name":"emulator"},{"id":62,"name":"filesystem"},{"id":80,"name":"finance"},{"id":15,"name":"firmware"},{"id":73,"name":"for-fun"},{"id":2,"name":"framework"},{"id":11,"name":"frontend"},{"id":22,"name":"game"},{"id":81,"name":"game-engine "},{"id":23,"name":"graphql"},{"id":84,"name":"gui"},{"id":91,"name":"http"},{"id":5,"name":"http-client"},{"id":51,"name":"iac"},{"id":30,"name":"ide"},{"id":78,"name":"iot"},{"id":40,"name":"json"},{"id":83,"name":"julian"},{"id":38,"name":"k8s"},{"id":31,"name":"language"},{"id":10,"name":"learning-resource"},{"id":33,"name":"lib"},{"id":41,"name":"linter"},{"id":28,"name":"lms"},{"id":16,"name":"logging"},{"id":76,"name":"low-code"},{"id":90,"name":"message-queue"},{"id":42,"name":"mobile-app"},{"id":18,"name":"monitoring"},{"id":36,"name":"networking"},{"id":7,"name":"node-version"},{"id":55,"name":"nosql"},{"id":57,"name":"observability"},{"id":46,"name":"orm"},{"id":52,"name":"os"},{"id":14,"name":"parser"},{"id":74,"name":"react"},{"id":82,"name":"real-time"},{"id":56,"name":"robot"},{"id":65,"name":"runtime"},{"id":32,"name":"sdk"},{"id":71,"name":"search"},{"id":63,"name":"secrets"},{"id":25,"name":"security"},{"id":85,"name":"server"},{"id":86,"name":"serverless"},{"id":70,"name":"storage"},{"id":75,"name":"system-design"},{"id":79,"name":"terminal"},{"id":29,"name":"testing"},{"id":12,"name":"ui"},{"id":50,"name":"ux"},{"id":88,"name":"video"},{"id":20,"name":"web-app"},{"id":35,"name":"web-server"},{"id":43,"name":"webassembly"},{"id":69,"name":"workflow"},{"id":87,"name":"yaml"}]" returns me the "expected json"