Inference

mpnn Property prediction with a trained MPNN (Chemprop-style) model: a public model or one you trained with the train-mpnn job. predictions has one value per molecule: the value (regression; the first target of a multi-target model, with every target in predictions_by_target), P(class 1) for binary models (labels at 0.5), or the predicted class's probability for multiclass (labels, and probabilities in output_spec.classes order). output_spec.task says which. Where the model has them (#908 Phase 3): uncertainty (an ensemble's member spread), intervals (90% conformal prediction interval of a regression model; intervals_by_target for several targets), ad_similarity (max Tanimoto similarity to the training set) and in_domain (inside the model's calibrated applicability domain). With return_contributions (#908 Phase 4): atom_contributions, one signed value per atom in input atom order (positive atoms raise the prediction, negative atoms lower it; Integrated Gradients through the whole model, ensembles averaged), attribution_base (a molecule's values add up to its prediction minus this baseline), attribution_method, and atom_contributions_by_target / attribution_base_by_target for several targets.

qsar Property prediction with a classical QSAR model you trained with the train-qsar job: the value (regression) or the probability of class 1 (binary), a 90% conformal interval or standard deviation where the model has one, the nearest-neighbour Tanimoto similarity to the training set, and optional signed per-atom attributions.

reaction_mpnn Reaction property prediction (yield, selectivity, barrier or rate constant) with a reaction model you trained with the train-reaction-mpnn job. input_data holds atom-mapped reaction SMILES (map them with reaction-map); a model trained with condition components (ligand, base, solvent, ...) also needs components with one SMILES per reaction under the same names. predictions has one value per reaction: the value (regression, clipped to the model's target_bounds; the first target of a multi-target model, with every target in predictions_by_target), P(class 1) for binary models (labels at 0.5), or the predicted class's probability for multiclass (labels, and probabilities in output_spec.classes order). Where the model has them: uncertainty (an ensemble's member spread), intervals (90% conformal prediction interval of a regression model), ad_similarity (max Tanimoto similarity of the reaction-plus-components fingerprint to the training set) and in_domain (inside the model's calibrated applicability domain).

thermo_hybrid Property prediction with a physics-hybrid model you trained with the train-thermo-hybrid job: for each corrected property, the group-contribution estimate (Joback or Stefanis-Panayiotou), the learned correction, their sum in the model's unit with a 90% conformal interval, and whether the molecule is in the applicability domain. Outside the domain the estimate is returned and flagged unless fallback is none.

thermo_mixture Solution and mixture property prediction with a model you trained with the train-thermo-mixture job: every solute (input_data) with every solvent (or row by row), at every combination of the condition values (e.g. a temperature grid), so one request gives a solvent screen or a property-versus-temperature curve. Each prediction carries its unit, a 90% conformal interval, and whether the solute is similar to training, the solvent was seen in training and the conditions lie in the training range.