Prediction of Ocular Toxicity of Prostaglandin F2α Analogs Based on Local Computational Models: Superiority over Global Models and Experimental Validation

Abstract

Drug-induced ocular toxicity is difficult to predict and evaluate, particularly for prostaglandin F2α (PGF2α) analogs used in the management of glaucoma. Traditional global predictive models, which are built with large and various datasets (n = 6187), provide systematic false-negative results for groups of structurally uniform compounds. To address this limitation, we developed special local classification models for PGF2α analogs. A structurally consistent local training dataset (n = 350) was assembled using Murcko scaffold filtering and Tanimoto similarity selection (≥0.5). Binary classification models were built using three different molecular fingerprints in combination with machine learning and two deep learning methods. Although two-dimensional molecular representations do not encode stereochemistry, computational predictions still apply to the shared two-dimensional scaffold of these compounds. These improved models were used to predict the ocular toxicity of latanoprost and its related impurities. The local models (n = 350) provided accurate toxicity predictions for latanoprost, latanoprost acid, 15(S)-latanoprost, and the trans-5,6-latanoprost isomer, whereas all 16 global computational models provided false-negative results. The experimental assessment performed with primary rabbit corneal epithelial cells (pRCECs) and a human corneal epithelial cell line (HCE-T) confirmed the cytotoxic effects, which were in agreement with the predictions made by the models. Among the tested compounds, 15(S)-latanoprost exhibited the highest cytotoxicity (IC50 = 86.22 μM, 95% CI: 82.90–89.56 μM), followed by trans-5,6-latanoprost (IC50 = 106.70 μM, 95% CI: 103.4–110.0 μM) and latanoprost (IC50 = 112.60 μM, 95% CI: 106.7–118.6 μM). At the standard therapeutic dosage (0.005%), no significant toxic response was observed. Virtual molecular docking was employed to explore the mechanism, and all analogs docked favorably into the quinone-binding channel of mitochondrial complex I and the catalytic cleft of SIRT3. These results show how a localized modeling approach is better able to capture structure–toxicity correlations among chemically similar compounds and highlight the critical need for rigorous impurity management in latanoprost products, particularly regarding 15(S)-latanoprost.