Ternary classification accuracies of fine-tuned GPT-3 and trained GNN models for “unknown” molecules.

Conjugated fragment	Number of molecules	HOMO accuracy		LUMO accuracy
		GPT-3	GNN	GPT-3	GNN
Naphthalene (1)	475	0.94	0.95	0.88	0.91
Anthracene (2)	577	0.99	1.00	0.93	0.97
Tetracene (3)	72	0.96	1.00	0.90	0.99
Pyrene (4)	237	0.98	1.00	0.97	0.99
Perylene (5)	41	0.98	1.00	0.98	0.95
(1) + (2) + (3) + (4) + (5)a	1402	0.97	0.98	0.93	0.95
p-Benzoquinone (6)	295	0.83	0.91	0.87	0.87
1,4-Naphthoquinone (7)	282	0.82	0.91	0.94	0.96
9,10-Anthraquinone (8)	186	0.86	0.91	0.99	1.00
(1) + (2) + (3) + (4) + (5) + (6) + (7) + (8)b	2165	0.88	0.91	0.95	0.96
1,8-Naphthalimide (9)	85	0.86	0.93	1.00	1.00
Naphthalenetetracarboxylic diimide (10)	88	0.86	0.88	0.95	0.98
Perylenetetracarboxylic diimide (11)	76	0.85	0.89	0.97	0.97
(1) + (2) + (3) + (4) + (5) + (6) + (7) + (8) + (9) + (10) + (11)c	3177	0.88	0.88	0.97	0.97

^aAll five families of molecules were excluded from model training. The HOMO/LUMO prediction accuracies reported in this row were measured on these five families of molecules.

^bAll eight families of molecules were excluded from model training. The HOMO/LUMO prediction accuracies reported in this row were measured on the families 6–8 of molecules.

^cAll 11 families of molecules were excluded from model training. The HOMO/LUMO prediction accuracies reported in this row were measured on the families 9–11 of molecules.