Inputs

The Artificial Intelligence Predictive Oncology Research Toolkit (AIxPORT) utilities exchange datasets via RO-Crate bundles. Each bundle is a directory that carries a ro-crate-metadata.json file. This page documents the expected layout for the *_train_rocrate and *_test_rocrate folders consumed by aixportcmd.py train and aixportcmd.py predict.

Note

To streamline downstream matching, DRE expects directory names to end with _train_rocrate or _test_rocrate. The prediction tooling will raise an error if those suffixes are missing.

Feature Tables

Every training and testing crate must include the genomic feature panel used by the algorithms. The files below are placed at the top-level of the crate directory.

  • gene2ind.txt:

    Tab-delimited mapping from zero-based gene indices to gene symbols.

    0       ABCB1
1       ABCC3
2       ABL1
  • cell2ind.txt:

    Tab-delimited mapping from cell (genotype) indices to cell line names.

    0       201T_LUNG
1       22RV1_PROSTATE
2       2313287_STOMACH
  • cell2mutation.txt:

    Comma-delimited matrix whose rows correspond to cell2ind.txt entries and whose columns follow gene2ind.txt ordering. Values are 1 when a gene carries a non-synonymous mutation and 0 otherwise.

    0,0,1,0,0,0..
0,0,0,0,1,0..
0,0,0,0,0,0..
  • cell2cndeletion.txt:

    Comma-delimited matrix with the same shape as cell2mutation.txt. A value of 1 denotes a copy-number deletion event; 0 means no deletion.

    0,0,0,0,0,0..
0,1,0,0,0,0..
0,0,0,0,1,0..
  • cell2cnamplification.txt:

    Comma-delimited matrix indicating copy-number amplification events.

    0,0,0,0,0,0..
0,0,0,1,0,0..
0,1,0,0,0,0..

References:

  1. Park, S., Silva, E., Singhal, A. et al. A deep learning model of tumor cell architecture elucidates response and resistance to CDK4/6 inhibitors. Nat Cancer (2024). https://doi.org/10.1038/s43018-024-00740-1

Training RO-Crates

Additionally to the feature tables, training crates include the training data:

  • training_data.txt:

    Tab-delimited table with one row per training observation. Columns are:

    1. Cell identifier matching cell2ind.txt (column 1)

    2. Drug SMILES string (column 2)

    3. Observed response value (floating point, column 3)

    4. Optional data source label (column 4)

    HS633T_SOFT_TISSUE      CC1=C(C(=O)N(C2=NC(=NC=C12)NC3=NC=C(C=C3)N4CCNCC4)C5CCCC5)C(=O)C        0.6695136077442607      GDSC2
KINGS1_CENTRAL_NERVOUS_SYSTEM   CC1=C(C(=O)N(C2=NC(=NC=C12)NC3=NC=C(C=C3)N4CCNCC4)C5CCCC5)C(=O)C        0.6444092636032414      GDSC1

Running aixportcmd.py train writes algorithm-specific output RO-Crates under <output>/trainedmodels. Each subdirectory includes the fitted model (for example, model.pt or model.pkl), and a train_predictions.txt file that captures in-sample predictions.

Testing RO-Crates

Testing crates use the same feature tables as the training crates and includes a file with test data on which predictions are to be made:

  • test_data.txt:

    Tab-delimited table with one row per inference request. Columns mirror training_data.txt (cell identifier, SMILES string, numeric response if available, and optional source label).

    EW24_BONE       CC1=C(C(=O)N(C2=NC(=NC=C12)NC3=NC=C(C=C3)N4CCNCC4)C5CCCC5)C(=O)C        0.98852067122827        GDSC1
OCILY7_HAEMATOPOIETIC_AND_LYMPHOID_TISSUE       CC1=C(C(=O)N(C2=NC(=NC=C12)NC3=NC=C(C=C3)N4CCNCC4)C5CCCC5)C(=O)C        0.2728634745574858      GDSC1

aixportcmd.py predict pairs each *_test_rocrate with the corresponding trained model directory (<dataset>_train_rocrate_<algorithm>). The command generates per-algorithm output directories beneath <output>/predictions that contain RO-Crate metadata, prediction scores, and any algorithm-specific logs.