Table 1 Performance comparison between the trained transformers and random-forest models across several metrics

From: A chemical language model for molecular taste prediction

  

Unweighted average

 

Model

Accuracy

Precision

Recall

F1 Score

AUROC

Support

XGBoost: fingerprints (fp)

0.8988

0.8169

0.7400

0.7661

0.8520

100%

XGBoost: fp+descriptors

0.8962

0.8842

0.7402

0.7779

0.8513

100%

Balanced Random Forest: fp

0.7972

0.5845

0.7322

0.6014

0.8391

100%

Chemprop

0.8851

0.8037

0.7224

0.7499

0.8392

100%

FART

0.8860

0.7720

0.6873

0.7118

0.9639

100%

FART augmented

0.8940

0.8789

0.7388

0.7737

0.9744

100%

FART augmented + confidence

0.9155

0.9000

0.7617

0.7956

0.9806

94%

  1. All scores are given as unweighted averages across taste classes which emphasizes minority classes, in our case umami. Area under the receiver operating characteristic (AUROC) values are calculated as one-vs-rest for each taste class. Support refers to the number of compounds out of the test set which were assigned a prediction by the model. The best score is highlighted for each metric in bold.
Back to article page