EcoTox-QSAR Predictor
Model card
EcoTox-QSAR Predictor estimates aquatic or soil toxicity when a chemical–species–endpoint–condition combination lacks a matched measurement. W00 and M00 are fitted separately; each uses related records to inform predictions while retaining task-specific responses.
Extrapolation requires caution
S1/S2 assess conditional completion within the retained ECOTOX population. S3 canonical-parent and S4 nonempty-ring Murcko-scaffold holdouts are structural extrapolation stress tests. Low training support or a ring scaffold absent from the route-specific training reference warrants caution and consideration of experimental confirmation. These diagnostics are not calibrated error bounds or universal accept/reject rules.
Scientific use
Use the predictor to estimate laboratory toxicity for an unmeasured combination within an available task, then consider the estimate together with task-level performance and reference support. This can help prioritize which sparse chemical–species–condition combinations merit new measurements. Predictions remain distinct from observed values and should not replace experimental confirmation where decisions depend on them.
Model and inputs
Separate context-aware multitask models are used for aquatic W00 and soil M00 predictions. Molecular structure, selected species and taxonomy, endpoint and effect, effect level, and available exposure context inform a trained output head. A task without a trained head cannot be predicted. When model artifacts are available, the site serves four-seed ensembles; submitted inputs are not used for online training.
Users provide a canonical SMILES for an organic compound, medium, species, toxicity task, and optional exposure duration. The W00 target is -log10(mol/L); the M00 target is -log10(mol/kg). Back-transformed values are reported only when the conversion is defined, with soil mass concentration displayed in mg/kg.
Evidence and evaluation setting
The principal evaluations hold out condition keys (S1) and parent–species–task combinations (S2) within the retained ECOTOX population. These are not unseen-chemical validations. Canonical-parent (S3) and nonempty-ring Murcko-scaffold (S4) holdouts probe stronger structural extrapolation. Neither the principal evaluations nor the stress tests establish accuracy for every new query or independent study.
On this site, available task-level R2, RMSE, MAE, sample counts, and observed-versus-predicted plots describe the evaluation records shipped with the active model. Read the task and evaluation subset with each result; pooled performance cannot stand in for a weakly supported task. The plotted query marker is a prediction, not a new observed value.
Prediction Scope
Predictions are intended primarily for conditional completion within the modeled ECOTOX evidence space. An unmeasured chemical–species–endpoint–condition combination may be unsupported in an isolated task-specific dataset yet still draw on chemical, biological, and experimental information represented across related tasks. This does not establish unrestricted extrapolation to wholly unseen chemicals or species.
Training Support
Multidimensional training support describes the relative amount of input-side evidence available for a query. It is not a calibrated probability of prediction error, a binary reliability decision, or a universal applicability-domain threshold. Interpret support together with task-level performance and the requested prediction context. High training support does not override poor or insufficient task-level performance.
The chemical distance uses the site's available chemical reference and remains a descriptive measure. A separate ring-scaffold status is compared with the corresponding route's training reference when its provenance can be verified. Acyclic queries are marked explicitly and never called unseen-ring-scaffold queries. If the route reference is unavailable, the site reports “Scaffold status unavailable” instead of inferring novelty. These diagnostics do not reproduce the S4 test or the S3 support-stratification rule.
An unseen species/category can be encoded for an already trained output head, but this does not validate accuracy for unseen species. An output head without training is rejected.
Limits
The current soil input representation does not include soil pH, organic carbon, texture, or moisture. Predictions describe laboratory toxicity under the represented response definitions; they do not model site-specific exposure or ecological risk. The tool does not derive an SSD or HC5, and its outputs are not regulatory risk-assessment conclusions.