Predictions, confidence intervals, feature importance, and model metrics from DeepChem.

You get back a predicted value for each molecule — a number for regression endpoints or a probability for classification — often with an uncertainty estimate and an applicability-domain flag. Read those together: a confident, in-domain prediction can drive triage directly, while an out-of-domain or high-uncertainty result is a signal to confirm experimentally before you act on it.

Output artifacts

Each completed DeepChem job returns property predictions, optional uncertainty estimates, and model metadata.

Primary output fields

Field	Type	Description
smiles	string	Input SMILES string
prediction	float	Predicted property value (regression) or probability (classification)
uncertainty	float	Prediction uncertainty (if uncertainty estimation enabled)
confidence_low	float	Lower bound of 95% confidence interval
confidence_high	float	Upper bound of 95% confidence interval
applicability_domain	boolean	Whether the molecule is within the model's applicability domain
model	string	Model architecture used
endpoint	string	Predicted property endpoint

Interpreting predictions

Regression endpoints

For continuous property predictions (solubility, logP, binding affinity):

Predicted value: The model's best estimate of the property value in the appropriate unit
Uncertainty: Standard deviation of the prediction (lower = more confident)
Confidence interval: 95% CI; narrower intervals indicate more reliable predictions

Endpoint	Unit	Typical Range	Good Value
Solubility (logS)	log mol/L	-10 to 2	> -4 (soluble)
Lipophilicity (logP)	unitless	-2 to 8	1-3 (drug-like)
IC50	nM	0.1 to 100,000	< 100 (potent)

Classification endpoints

For binary property predictions (hERG, AMES, BBB):

Probability: Value between 0 and 1 indicating predicted positive probability
Threshold: Default 0.5 for binary classification; adjust for sensitivity/specificity tradeoff
Confidence: Distance from the decision boundary; values near 0.5 are uncertain

Endpoint	Positive Class	Concern Threshold
hERG inhibition	Inhibitor	> 0.5 (potential cardiac risk)
AMES mutagenicity	Mutagenic	> 0.5 (potential genotoxicity)
BBB permeability	Permeable	> 0.5 (crosses blood-brain barrier)
CYP inhibition	Inhibitor	> 0.5 (potential drug interaction)

Uncertainty estimation

Ensemble uncertainty

Multiple models trained with different random seeds predict each molecule
Standard deviation across ensemble members provides uncertainty estimate
High disagreement between models indicates unreliable prediction

MC-dropout uncertainty

Multiple forward passes with active dropout layers
Mean prediction and standard deviation computed across samples
Efficient uncertainty estimation without training multiple models

Applicability domain

The applicability domain flag indicates whether the input molecule is similar enough to the training data for reliable predictions:

true: Molecule is within domain; prediction is expected to be reliable
false: Molecule is outside domain; treat the prediction with caution
Based on distance to nearest training set neighbors in feature space

Feature importance

When available, feature importance maps highlight which atoms or substructures contribute most to the prediction:

Atom-level importance: Color-coded molecular depiction showing contribution of each atom
Positive contributions: Atoms pushing the prediction toward higher values (or positive class)
Negative contributions: Atoms pushing the prediction toward lower values (or negative class)
Interpretation: Helps identify structural motifs responsible for the predicted property

Multi-task results

For multi-endpoint predictions, results include:

Separate prediction and uncertainty for each endpoint
Cross-endpoint correlation analysis
Combined scoring if multi-parameter optimization is configured
Endpoint-specific applicability domain flags

Model performance metrics (custom models)

When training custom models, the output includes performance metrics:

Metric	Type	Description
RMSE	Regression	Root mean squared error on test set
MAE	Regression	Mean absolute error on test set
R-squared	Regression	Coefficient of determination
AUC-ROC	Classification	Area under ROC curve
AUC-PRC	Classification	Area under precision-recall curve
Accuracy	Classification	Overall classification accuracy
Training loss	Both	Final training loss value
Validation loss	Both	Best validation loss (early stopping)

Downloading artifacts

CSV export: SMILES, predictions, uncertainties, and applicability domain flags
JSON export: Structured results with model metadata and performance metrics
Model checkpoint: Trained model weights for reuse (custom models only)
Feature importance PNG: Molecular depictions with atom-level importance coloring

Canvas visualization

When a DeepChem job completes, a result card appears on the canvas showing:

Prediction summary table with property values
Property distribution histograms for batch predictions
Feature importance visualization for selected molecules
Model performance metrics for custom training jobs

Quality checks

Verify that predictions are within physically reasonable ranges for the endpoint
Check applicability domain flags; out-of-domain predictions should be treated cautiously
Compare predictions with known experimental values for validation compounds
For classification, review the probability distribution to ensure the model is not uniformly uncertain
For custom models, check that test set metrics meet minimum quality thresholds (e.g., R-squared > 0.5, AUC > 0.7)

DeepChem outputs and interpretation