Table 5. Prediction Accuracies (R²) in the Additive-Out Ablation Study for the Test1–Test4 Datasets of the BHC Reaction^a.

	Test1	Test2	Test3	Test4
MPNN	0.93 (0.00)	0.90 (0.00)	0.71 (0.00)	0.52 (0.00)
MPNN(3.0m)	0.87 (0.00)	0.88 (0.00)	0.74 (0.00)	0.62 (0.00)
transformer	0.91 (0.00)	0.91 (0.00)	0.64 (0.01)	0.44 (0.00)
transformer(random)	0.56 (0.00)	0.61 (0.00)	0.21 (0.00)	0.29 (0.00)
MPNN-transformer	0.87 (0.01)	0.88 (0.01)	0.59 (0.03)	0.64 (0.01)
MPNN-transformer(3.0m)	0.87 (0.01)	0.90 (0.00)	0.74 (0.01)	0.61 (0.01)

^aFor each model and dataset, the average (standard deviation) of the R² value for the Test1–Test4 datasets of the BHC reaction using the five ensemble models is reported. Transformer: the transformer encoder layer. Transformer(random): the transformer encoder layer with random input vectors. MPNN: the MPNN part of the proposed model in Figure 2. MPNN(3.0m): MPNN with contrastive learning using 3.0 million compounds. For each test dataset, the highest R² value is highlighted in bold.