AI‑Computational Chemistry Workflow Cuts Diabetic Wound‑Healing Drug R&D Time by 70%

Background

Diabetic foot ulcers (DFU) present a persistent clinical challenge because chronic hyperglycemia impedes wound closure and increases amputation risk. Traditional drug discovery for wound healing relies on large‑scale empirical screening of molecular libraries, which is time‑consuming and vulnerable to human bias.

AI‑CC Integrated Workflow

An AI‑Computational Chemistry (AI‑CC) workflow tightly couples large‑language‑model (LLM)‑driven literature mining (qualitative insight) with multi‑stage molecular simulations (quantitative validation) to create a closed‑loop system for analyzing drug‑protein nanoscale interactions.

Data Collection

Protein dataset : Queries of miRTarBase, PubMed Central, and Web of Science identified 26 miRNAs linked to diabetic wounds, yielding 20,334 miRNA‑protein records, 9,186 UniProt entries, and after deduplication 8,739 core proteins.

Drug dataset : DrugBank provided 4,487 drug records associated with the core proteins; ChEMBL supplied molecular structures and descriptors, resulting in 2,989 small‑molecule candidates.

LLM Evaluation

Four LLMs—LLaMA2‑Chat‑13B, PMC‑LLaMA‑13B, GPT‑3.5, and GPT‑4—were benchmarked on a manually annotated test set. GPT‑4 achieved the highest overall score (0.737) owing to superior zero‑shot and few‑shot performance and was selected for downstream analysis.

Qualitative Screening

Using the GPT‑4 literature‑mining module, 3,119 PubMed articles mentioning both diabetic wounds and target proteins were retrieved. The model extracted mechanistic clues, producing 756 drug‑protein pairs supported by 3,889 additional papers. Prompting GPT‑4 for mechanism‑of‑action (MOA) statements yielded 432 drug‑protein MOA entries.

Quantitative Multi‑Stage Molecular Modeling

For each drug‑protein pair the workflow applied:

Molecular docking

Molecular dynamics (MD) simulations

Quantum chemistry (QC) calculations

Quantum calculations for the folic‑acid–FGF complex used ORCA with the B97‑3c method (interaction energy = ‑78.126 kcal/mol) and Gaussian 16 with B3LYP‑D3/6‑31+G(d,p) (interaction energy = ‑86.20 kcal/mol).

Candidate Prioritization

Greedy set‑cover and chemical‑informatics clustering reduced the drug‑protein matrix to 50 key proteins. GPT‑4 analysis of the 2,989 drugs highlighted 30 top candidates; five expert‑suggested drugs (neomycin, mangiferin, mupirocin, metformin, sitagliptin) were added, yielding a final set of 35 candidates.

Experimental Validation

In vitro scratch assays with human dermal fibroblasts (HDFs) and HaCaT keratinocytes showed that folic acid significantly accelerated wound closure to 134.90 % of the untreated control (p < 0.001). Positive control mupirocin and negative control metformin behaved as expected. Acyclovir caused a slight delay, simvastatin was cytotoxic, cholic acid outperformed the positive control, and pyridoxal phosphate impeded healing.

Key Findings

The AI‑CC pipeline shortened the literature‑to‑experiment cycle by >70 % compared with conventional approaches.

Folic acid ranked first in combined qualitative (mechanistic) and quantitative (interaction energy) scores, confirming strong binding to fibroblast growth factor and therapeutic potential.

Integration of AI‑driven mechanistic insight with computational‑chemistry interaction metrics provides a complementary, iterative optimization loop.

Reference

Quantitative Computational Validation of Nanoscale Interactions between Drug Molecules and Diabetic Wound‑Related Proteins for Drug Discovery, ACS Nano Medicine. DOI: 10.1021/acsnanomed.5c00180