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

Table 4 The results of molecule optimization based on specified editing positions on the small molecule drug dataset, with the evaluation measured by the hit ratio (%) of the 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

Drug Molecule	Objectives	Prompts	Validity (%)		Hit Ratio (%)
Drug Molecule			GPT-3.5	Ours	GPT-3.5	Ours
Apixaban	Electron-donating substitution	High HOMO energy	55	73	1	18
Empagliflozin	More hydrophilic	Large polarity	0	15	0	8
Ibrutinib	Higher permeability	High permeability	19	33	0	2
Dapagliflozin	Large aqueous solubility	More soluble	0	35	0	13
Osimertinib	Electron-withdrawing substitution	Low HOMO energy	30	52	0	39
Olaparib	Electron-withdrawing substitution	Low HOMO energy	34	30	1	2
Abemaciclib	More hydrogen bonding interactions	More hydrogen bond acceptors	20	42	0	21
		More hydrogen bond donors	0	45	0	25
Pomalidomide	Electron-withdrawing substitution	Low HOMO energy	24	64	11	10
Tafamidis	More hydrogen bonding interactions	More hydrogen bond acceptors	46	61	0	13
Tafamidis		More hydrogen bond donors	10	48	0	5
Lisdexamfetamine	To control the hydrolysis rate	High HOMO energy	0	11	0	8
Lisdexamfetamine		Low HOMO energy	0	19	0	5

Note that the prompts shown in the table have been abbreviated. The prompts actually used in 3DToMolo can be found in Table 2, while we formulated our request to GPT\(-\)3.5 as follows: “Optimize the following molecule: <SMILES> to achieve the desired properties: <specified properties>, while ensuring that the specified substructure <substructure> remains unchanged. Provide 10 different and valid results. Note in this experimental setting, we employed a multi-run approach to facilitate the generation of a total of 100 optimized results for each drug molecule. Specifically, each iteration obtained 10 distinct outputs from GPT\(-\)3.5, with a total of 10 iterations conducted. 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