Table 1.

Comparison of yield predictions between the SEMG-MIGNN model with other SOTA models

Data splitting	Yield-BERT	DRFP	MFF	SEMG-MIGNN
Random 90/10	5.20 ± 0.500	5.09 ± 0.500	6.34 ± 0.500	4.79 ± 0.500
Random 70/30	5.82 ± 0.400	6.28 ± 0.300	6.77 ± 0.300	4.81 ± 0.400
Random 50/50	7.62 ± 0.500	7.36 ± 0.300	8.55 ± 0.300	6.83 ± 0.500
Random 30/70	9.41 ± 0.500	8.67 ± 0.500	10.09 ± 0.500	8.79 ± 0.700
Aryl Halidea	26.04 ± 0.300	26.19 ± 0.200	22.04 ± 0.200	19.34 ± 0.400
Additivea	21.29 ± 0.200	22.43 ± 0.200	21.66 ± 0.200	10.36 ± 0.200
Liganda	20.04 ± 0.200	18.35 ± 0.200	18.85 ± 0.200	11.02 ± 0.200
Basea	19.40 ± 0.200	19.90 ± 0.200	20.66 ± 0.200	14.52 ± 0.200

Note: The best performance of each task is shown in bold. ^aThese data splitting tasks refer to the extrapolative predictions based on the scaffold splitting of the reaction components. Details are elaborated in Supplementary Fig. 20. RMSEs are in %.