Table 6.

Results for the AI models are reported for split 90–10. Each row represents a distinct AI architecture. The first of the three sections relates to the evaluation of the models on the external test set, the second assesses the model in training, and the third contains information on the encoder used.

	Test Set					Training Set
	Balance Accuracy	Precision	Sensitivity	Specificity	MCC	F1-Score	Balance Accuracy	Encoder
DNN Circular Fingerprint	0.746	0.527	0.672	0.819	0.454	0.591	0.870	Circular Fingerprint
DNN MACCS	0.700	0.446	0.638	0.762	0.358	0.525	0.737	MACCS fingerprint
DNN CDDD	0.808	0.539	0.828	0.788	0.542	0.653	0.836	Latent representation CDDD
DNN Molecular Descriptors	0.774	0.471	0.828	0.720	0.470	0.600	0.811	Molecular Descriptors
MPNN	0.746	0.527	0.672	0.819	0.454	0.591	0.741	Graph
NLP chars Embedding	0.780	0.551	0.741	0.819	0.510	0.632	0.753	Text Vectorization and character embedding
NLP chars Embedding Augmented	0.815	0.616	0.776	0.855	0.585	0.687	0.886	Text Vectorization and character embedding
Multimodal	0.808	0.592	0.777	0.839	0.564	0.672	0.830	All (no graph)
Extreme Gradient Boosting (best ML with oversampling methods)	0.742	0.605	0.600	0.883	0.485	0.602	0.921	Latent Description CDDD