Researchers from Southern Medical University have successfully identified four FDA-approved drugs with previously unknown lipid-lowering properties using an innovative artificial intelligence approach, potentially offering new treatment options for patients with hyperlipidemia who struggle with statin intolerance or inadequate response to conventional therapies.
The study, published in Acta Pharmacologica Sinica, employed a comprehensive machine learning framework to analyze 3,430 FDA-approved compounds, including 176 known lipid-lowering agents and 3,254 control drugs. The research team extracted molecular descriptors and fingerprints from SMILES codes and physicochemical data, then trained 68 different machine learning models including random forest, support vector machine, gradient boosting, and elastic net combinations.
Machine Learning Models Achieve High Accuracy
The top-performing AI models reached an area under the curve (AUC) of approximately 0.886 and accuracy of 0.888. Drugs predicted positive in at least 8 of the top 10 models were flagged as repurposing candidates, yielding 29 potential lipid-lowering drugs for further investigation.
"We've established a paradigm for AI-driven drug repositioning. By integrating computational predictions with clinical and experimental validation, we bypass decades of traditional drug development—offering clinicians new tools faster and cheaper," said Dr. Peng Luo, the study's senior author.
Clinical Validation Confirms Four Promising Candidates
To validate the AI predictions, researchers analyzed medical records spanning nearly 25 years (June 1998–May 2024) from Zhujiang Hospital. Four agents demonstrated statistically significant lipid-lowering effects:
Argatroban showed the most dramatic results among 63 patients, with LDL cholesterol falling by 33% (from 2.96 mmol/L to 1.98 mmol/L), total cholesterol decreasing by 25% (from 4.68 to 3.51 mmol/L), and triglycerides also declining significantly (all P < 1 × 10⁻⁸).
Levoxyl (levothyroxine) users (n = 87) experienced LDL and total cholesterol reductions of 16% and 12%, respectively. Oseltamivir and thiamine also exhibited moderate but statistically significant lipid effects.
Animal Studies Reveal Diverse Lipid Effects
Sixteen drugs selected from computational predictions and clinical evidence were tested in male C57BL/6 mice to further validate the findings. The results showed varied but significant lipid modulation effects:
- Argatroban and Promega reduced total cholesterol by approximately 10%
- Levoxyl and sulfaphenazole each lowered triglycerides by 27-29%
- Prasterone, alpha-tocopherol acetate, sorafenib, Cedazuridine, and Promega significantly increased HDL levels, with Prasterone producing the largest rise of approximately 24%
Unexpectedly, some drugs showed modest LDL increases compared to controls, including procarbazine, dimenhydrinate, and Promega, highlighting the complex nature of lipid metabolism modulation.
Molecular Mechanisms Revealed Through Docking Studies
To understand the underlying mechanisms, researchers performed molecular docking and dynamics simulations on seven promising drugs against 12 lipid metabolism targets, including HMG-CoA reductase, coagulation factor X, serotonin receptors, thyroid hormone receptors, and others.
The analysis revealed distinct pathways for each candidate:
- Argatroban bound tightly to coagulation factor X (approximately -7.6 kcal/mol), forming stable hydrophobic interactions and hydrogen bonds
- Levoxyl showed high affinity for thyroid hormone receptor alpha (TRα)
- Sulfaphenazole bound to serotonin receptor subtypes
- Prasterone engaged RXRα and COX-2
- Sorafenib demonstrated affinity to HMG-CoA reductase
Clinical Implications and Future Directions
This research addresses critical gaps in hyperlipidemia treatment, where many patients struggle with intolerance or inadequate response to statins and other standard therapies. The study's approach integrates computational predictions with clinical and experimental validation, potentially offering alternative treatment options for patients requiring individualized therapy approaches.
The researchers acknowledge several limitations, including reliance on retrospective single-center clinical data and murine models that may not fully translate to human physiology. Randomized controlled trials and additional biochemical studies will be necessary to confirm efficacy and safety in human populations.
The heterogeneity of positive control drugs suggests that stratified predictive models may increase precision in future applications. The established framework could potentially be applied to drug repurposing efforts in other therapeutic areas, representing a significant advancement in AI-driven pharmaceutical research.
