Sculpting molecules in text-3D space: a flexible substructure aware framework for text-oriented molecular optimization

Table 3 The results of molecule optimization with structural constraints on the electrolyte additive dataset [49] were evaluated using the hit ratio (%) of property changes in each objective-oriented prompt. Statistically significant improvement (t-test over 5 different dataset splits, p-value< 0.05) is highlighted with bold text

Prompts	Validity (%)		Hit Ratio (%)
Prompts	GPT-3.5	3DToMolo	GPT-3.5	3DToMolo
This molecule has low HOMO (Highest occupied molecular orbital) value, which has high oxidation potential.	2.867	41.54	0.2797	10.91
This molecule has high HOMO (Highest occupied molecular orbital) value, which has low oxidation potential.	0.4196	42.73	0.06993	29.65
This molecule has low LUMO (Lowest Unoccupied Molecular Orbital) value, which has high reduction potential.	0	44.55	0	25.38
This molecule has high LUMO (Lowest Unoccupied Molecular Orbital) value, which has low reduction potential.	0.4895	42.45	0.4895	18.81
This molecule has high polarity.	1.399	42.03	1.119	30.35
This molecule has low polarity.	0.1399	43.15	00.00	12.17
This molecule is soluble in water.	2.448	55.24	1.678	36.50
This molecule is insoluble in water.	1.399	57.27	1.119	20.63

Note that the prompts we input into GPT\(-\)3.5 have subtle differences from those listed in the above table. To ensure a fair comparison, we formulated our request to GPT\(-\)3.5 as follows: “Optimize this molecule: <SMILES> to achieve <specified properties>. Optimization is restricted to hydrogen atoms only, while the rest of the input molecule remains fixed. The number of atoms (including hydrogen) added to the optimized molecule must be within the range of 1 to 10. Provide 10 different and valid results.” For the compound prompts, the hit ratio is calculated based on the satisfaction of the redox potential part. In most cases, GPT\(-\)3.5 fails to meet the requirements, as the generated molecules do not conform to either the specified structural criteria or the desired properties

ISSN: 1471-2105