Available Inputs & UQ Methods

Below, you can find an overview of the input types, UQ methods, and error models implemented in UNIQUE.

List of available UQ methods & objects in UNIQUE.

Type	Name	Description	Reference(s)
Input Type	FeaturesInputType	Data-based inputs - i.e., features that can be directly computed from/linked to the data. Features can be provided as a single value or as an array of values/features for each datapoint. Numerical features can contain integer-only (binary included) or real-valued (floats) values. Check out Input Types Specification.
Input Type	ModelInputType	Model-based inputs - i.e., outputs associated with the original predictive model. Depending on the task (problem_type, either “classification” or “regression”), two types of model-based inputs can be provided: for classification tasks, the predicted main class probabilities (as a single value per datapoint); for regression tasks, either the individual ensemble member’s predictions (as an array) or the pre-computed ensemble variance (as a single value) for each datapoint. Check out Input Types Specification.
UQ Method	EnsembleVariance	Computes the variance of the ensemble’s predictions. Either the individual ensemble member’s predictions (as an array) or the pre-computed variance (as a single value) for each datapoint can be provided.
UQ Method	Probability	Returns the predicted primary class probability. Expects the predicted main class probability value as input, not the ensemble’s (class) predictions.
UQ Method	ManhattanDistance	Returns the k-nearest neighbors from the training set in the corresponding feature(s) space using the Manhattan distance metric.
UQ Method	EuclideanDistance	Returns the k-nearest neighbors from the training set in the corresponding feature(s) space using the Euclidean distance metric.
UQ Method	TanimotoDistance	Returns the k-nearest neighbors from the training set in the corresponding feature(s) space using the Tanimoto distance metric.
UQ Method	GaussianEuclideanKDE	Returns the kernel density estimation from the training set in the corresponding feature(s) space using the gaussian kernel and Euclidean distance metric.
UQ Method	GaussianManhattanKDE	Returns the kernel density estimation from the training set in the corresponding feature(s) space using the gaussian kernel and Manhattan distance metric.
UQ Method	ExponentialManhattanKDE	Returns the kernel density estimation from the training set in the corresponding feature(s) space using the exponential kernel and Manhattan distance metric.
“Transformed” UQ Method	SumOfVariances	Computes the sum of (computed) variances and distances converted to variances using the Calibrated Negative Log-Likelihood (CNLL) method.	Hirschfeld et al. (2020) - Eq. 11 & 12[1]
“Transformed” UQ Method	DiffkNN	Computes the absolute mean difference in predicted vs. target value for the k-nearest neighbors from the training set in the corresponding feature(s) space.	Sheridan et al. (2022)[2]
Error Model/”Transformed” UQ Method	UniqueRandomForestRegressor	Builds and trains a Random Forest regressor that predicts the pointwise prediction error.	Adapted from Lahlou et al. (2021)[3]
Error Model/”Transformed” UQ Method	UniqueLASSO	Builds and trains a LASSO regressor that predicts the pointwise prediction error.	Adapted from Lahlou et al. (2021)[3]

See also

Check out UQ Methods for more details about the difference between base and transformed UQ methods.

References

[1]

Hirschfeld L., et al. (2020). Uncertainty Quantification Using Neural Networks for Molecular Property Prediction. Journal of Chemical Information and Modeling, 60, 3770-3780. https://doi.org/10.1021/acs.jcim.0c00502

[2]

Sheridan R.P., et al. (2022). Prediction Accuracy of Production ADMET Models as a Function of Version: Activity Cliffs Rule. Journal of Chemical Information and Modeling, 62(14), 3275-3280. https://doi.org/10.1021/acs.jcim.2c00699

[3] (1,2)

Lahlou, S., et al. (2021). DEUP: Direct Epistemic Uncertainty Prediction. arXiv:2102.08501. https://arxiv.org/abs/2102.08501v4