Admetica

What is Admetica?

Admetica is Datagrok's open-source ADMET prediction toolkit for small molecules. In ProteinIQ, the current tool follows the active published runtime surface available in the published admetica==1.4.1 package and the published web runtime: you can predict 22 pharmacokinetic and toxicity properties from SMILES strings.

Admetica uses Chemprop graph neural networks trained per endpoint. Each model predicts one experimentally grounded ADMET property, so the tool is useful for early compound triage, lead optimization, and comparing candidates before synthesis or assay work.

How to use Admetica online

Paste one SMILES per line, upload an .sdf, .csv, .smi, .smiles, or .txt file, or fetch structures from PubChem. Tab-delimited name<TAB>SMILES input is supported if you want to preserve your own identifiers.

Input formats

Each run supports up to 1,000 molecules. Split larger libraries into separate jobs so prediction results and the native CSV remain within the safe delivery budget.

Property selection

ProteinIQ supports the native property subset selector. By default, all 22 available models are selected. You can deselect endpoints you do not need to reduce result columns and run only the models relevant to your screen.

Available properties

Output columns

Results are returned as a spreadsheet with one row per molecule. The published runtime lowercases its prediction column keys; ProteinIQ preserves those keys, their requested order, and the source endpoint definitions.

If you run only a subset of models, the output contains only ids, smiles, and the selected native property columns in the order you requested them.

The Files tab includes the exact CSV written by admetica_predict and a provenance JSON file containing the verified package version, source-distribution hash, checkpoint-bundle hash, and scientific runtime versions. The native CSV contains only the selected prediction columns; ids and smiles remain in the ProteinIQ spreadsheet for tracking.

Interpreting results

Admetica mixes regression and classification endpoints. Continuous outputs such as caco2, solubility, cl-hepa, cl-micro, half-life, and ld50 should be interpreted in the context of the published training data and endpoint definitions. Classification outputs such as the CYP, P-gp, and herg endpoints are probabilities for the model's positive class, not measured assay values.

Invalid SMILES remain in their submitted row with blank predictions, matching the source behavior. ProteinIQ reports the affected compound identifiers in a job warning instead of silently dropping them.

Use the predictions as triage signals rather than hard acceptance rules. Compounds outside the model training domain, including unusual chemotypes and larger non-drug-like structures, can produce less reliable estimates.

Native surface in ProteinIQ

This tool intentionally stays thin. ProteinIQ preserves the native runtime predictions and returns the currently available 22-model surface. The tool keeps ids and smiles for workflow tracking, but it does not add ProteinIQ-only descriptor columns to the native prediction CSV.

Limitations

• The tool reflects the published admetica==1.4.1 runtime surface rather than every asset present in the original repository.
• Only the 22 endpoints currently available by the native runtime are selectable here.
• Predictions are best suited to small-molecule chemical space similar to the published training data.
• This page does not add uncertainty estimates or re-score the native outputs.