Tools · Custom Modeling

Custom ADMET Models

Adapt ADMET-AI predictions to your own assay: recalibrate an endpoint onto your units, or stack a model on all 41 predictions

On this page

Prices, workflows, and method papers Open in the app

Run it from the API

Submit with Submit a job and the job_type below. Price it first with Estimate job reservation cost: submitting reserves that amount from your wallet, and the charge settles at the actual runtime.

Custom ADMET Model Training `train-admet-custom`

Job type: train-admet-custom
Hardware: cpu (default)
Typical runtime: 10 min on CPU

Payload

Payload fields
Field	Type	Description
`model_name`required	`string`	Limits: `min length 1`
`admet_endpoint`required	`string`	The ADMET-AI endpoint (an admet-predict column) to adapt: recalibrated onto your assay's scale, and used with the other 40 as stacking features. One of: "admet_hia_hou", "admet_bioavailability_ma", "admet_solubility_aqsoldb", "admet_lipophilicity_astrazeneca", "admet_hydrationfreeenergy_freesolv", "admet_caco2_wang", "admet_pampa_ncats", "admet_pgp_broccatelli", "admet_bbb_martins", "admet_ppbr_az", "admet_vdss_lombardo", "admet_half_life_obach", "admet_clearance_hepatocyte_az", "admet_clearance_microsome_az", "admet_cyp1a2_veith", "admet_cyp2c19_veith", "admet_cyp2c9_veith", "admet_cyp2d6_veith", "admet_cyp3a4_veith", "admet_cyp2c9_substrate_carbonmangels", "admet_cyp2d6_substrate_carbonmangels", "admet_cyp3a4_substrate_carbonmangels", "admet_herg", "admet_clintox", "admet_ames", "admet_dili", "admet_carcinogens_lagunin", "admet_ld50_zhu", "admet_skin_reaction", "admet_nr_ar", "admet_nr_ar_lbd", "admet_nr_ahr", "admet_nr_aromatase", "admet_nr_er", "admet_nr_er_lbd", "admet_nr_ppar_gamma", "admet_sr_are", "admet_sr_atad5", "admet_sr_hse", "admet_sr_mmp", "admet_sr_p53"
`task`	`string`	Regression only for now. Default: `"regression"`One of: `"regression"`
`methods[]`	`string[]`	Candidates compared by cross-validation: linear / isotonic recalibration of the chosen endpoint, and stacked models on fingerprints, descriptors and all 41 ADMET-AI predictions. Defaults to all. One of: `"recal_linear", "recal_isotonic", "stack_linear", "stack_random_forest", "stack_lightgbm", "stack_svm"`Limits: `min items 1`
`features[]`	`string[]`	Feature kinds a stacked model sees besides the 41 ADMET-AI predictions; empty to stack on the predictions alone. Default: `["morgan_counts","rdkit2d"]`One of: `"morgan_counts", "rdkit2d"`
`prediction_transform`	`string`	Applied to the chosen endpoint's prediction before a recalibration; use log10 when your assay is reported on a log scale and the endpoint is not (for example clearance). Default: `"none"`One of: `"none", "log10"`
`morgan_radius`	`integer`	Morgan fingerprint radius. Default: `2`One of: `1, 2, 3, 4`
`morgan_bits`	`integer`	Morgan fingerprint size. Default: `2048`One of: `512, 1024, 2048, 4096`
`input_data[]`	`string[]`	Training SMILES. Limits: `min items 1, max items 10000`
`input_format`	`string`	QSAR features are computed from SMILES. Default: `"smiles"`One of: `"smiles"`
`targets[]`	`number[]`	One measured value per input_data entry, in the assay's own units (for example log10 CLint). Limits: `min items 1, max items 10000`
`splits[]`	`string[]`	Fold label per input_data entry for split_method 'provided'. One of: `"train", "training", "val", "valid", "validation", "test"`Limits: `min items 1`
`activity_table`	`string \| object[] \| artifactRef`	Training table instead of input_data + targets: CSV/JSON rows or a closed $artifact ref (e.g. bioactivity-curate's curated table).
`smiles_column`	`string`	activity_table column with SMILES. Default: `"smiles"`Limits: `min length 1`
`target_column`	`string`	activity_table column with the training target (p_activity is -log10 M). Default: `"p_activity"`Limits: `min length 1`
`split_column`	`string`	activity_table column with train / val / test labels; expanded into splits for split_method 'provided'. Limits: `min length 1`
`endpoint`	`string`	Train on rows of this endpoint only (e.g. IC50); required when the table mixes endpoints. Limits: `min length 1`
`include_censored`	`boolean`	Keep censored rows (relation > or <); dropped by default. Default: `false`
`training_data_sources[]`	`object[]`	Set by CogniChem when the activity table carries ChEMBL identifiers (CC BY-SA 3.0), so the model card names ChEMBL as a training-data source. Do not send. Limits: `max items 1`
training_data_sources.`source`required	`"chembl"`
training_data_sources.`release`	`string \| null`
`selection_metric`	`string`	Cross-validated metric that picks the saved method. Default: `"rmse"`One of: `"rmse", "mae", "r2"`
`cv_folds`	`integer`	Cross-validation folds on the training data; scaffold and cluster splits keep each group in one fold. Default: `5`Limits: `≥ 3, ≤ 10`
`split_method`	`string`	How the held-out test set is drawn (default random, or provided when splits are sent): random, scaffold (Bemis-Murcko groups), cluster (Butina on Morgan fingerprints at Tanimoto 0.4) or provided. One of: `"random", "scaffold", "cluster", "provided"`
`test_size`	`number`	Fraction held out as the test set and scored against mean and kNN-Tanimoto baselines (0 skips it). Default: `0.1`Limits: `≥ 0, ≤ 0.5`
`random_seed`	`integer`	Seed for the split, folds and models; drawn and recorded when omitted. Limits: `≥ 0`
`unit`	`string`	Unit of the targets (regression); returned with every prediction from the trained model.
`model_description`	`string`	Free-text description stored with the trained model.
`tags[]`	`string[]`	Labels stored with the trained model.

Example

from cognichem_client import CogniChem

client = CogniChem.from_env()  # reads COGNICHEM_API_KEY
payload = {
    "model_name": "admet-custom-tiny",
    "admet_endpoint": "admet_lipophilicity_astrazeneca",
    "input_data": ["CCO", "CCCO", "CCCCO", "CCCCCO", "CC(C)O", "… 35 more"],
    "targets": [-0.001, 0.389, 0.779, 1.169, 0.387, "… 35 more"],
    "methods": ["recal_linear", "recal_isotonic", "stack_linear"],
    "features": ["morgan_counts"],
    "morgan_bits": 512,
    "cv_folds": 3,
    "test_size": 0.2,
    "random_seed": 0,
    "unit": "logD",
}

estimate = client.jobs.estimate(job_type="train-admet-custom", payload=payload, resource="cpu")
print(f"Reserves ${estimate.cost:.2f}")

job = client.jobs.submit(
    job_name="my-train-admet-custom-run",
    job_type="train-admet-custom",
    payload=payload,
    resource="cpu",
)
status = client.jobs.wait(job.process_id)
if status.status == "completed":
    client.jobs.result(job.process_id, save_path=".")

curl -X POST "https://api.cognichem.com/api/v1/jobs/submit" \
  -H "X-Api-Key: $COGNICHEM_API_KEY" \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{
  "job_name": "my-train-admet-custom-run",
  "job_type": "train-admet-custom",
  "payload": {
    "model_name": "admet-custom-tiny",
    "admet_endpoint": "admet_lipophilicity_astrazeneca",
    "input_data": [
      "CCO",
      "CCCO",
      "CCCCO",
      "CCCCCO",
      "CC(C)O",
      "… 35 more"
    ],
    "targets": [
      -0.001,
      0.389,
      0.779,
      1.169,
      0.387,
      "… 35 more"
    ],
    "methods": [
      "recal_linear",
      "recal_isotonic",
      "stack_linear"
    ],
    "features": [
      "morgan_counts"
    ],
    "morgan_bits": 512,
    "cv_folds": 3,
    "test_size": 0.2,
    "random_seed": 0,
    "unit": "logD"
  },
  "resource": "cpu"
}'

{
  "model_name": "admet-custom-tiny",
  "admet_endpoint": "admet_lipophilicity_astrazeneca",
  "input_data": [
    "CCO",
    "CCCO",
    "CCCCO",
    "CCCCCO",
    "CC(C)O",
    "… 35 more"
  ],
  "targets": [
    -0.001,
    0.389,
    0.779,
    1.169,
    0.387,
    "… 35 more"
  ],
  "methods": [
    "recal_linear",
    "recal_isotonic",
    "stack_linear"
  ],
  "features": [
    "morgan_counts"
  ],
  "morgan_bits": 512,
  "cv_folds": 3,
  "test_size": 0.2,
  "random_seed": 0,
  "unit": "logD"
}

Sample data from the job catalog; long values are shortened here. Each job_name must be unique among your jobs.

Workflow inputs

Table (list) · optionalCSV, TSV, JSON
Molecules (list) · optionalSMILES

Workflow outputs

ArchiveZIP
TableCSV
Trained modelOPAQUE

Run it from the API

Custom ADMET Model Training train-admet-custom

Custom ADMET Model Training `train-admet-custom`