Aligning AI Toxicity Predictions with Wet-Lab Biology for PFOA Toxicity in SH-SY5Y Cells

Authors : Didem Oral

Pages : 110-116

View : 48 | Download : 118

Publication Date : 2025-12-23

Article Type : Research Paper

Abstract :Perfluorooctanoic acid (PFOA) is a highly persistent per- and polyfluoroalkyl substance (PFAS) widely detected in the environment and biological systems. Its resistance to degradation and bioaccumulative behavior make it a critical toxicological and public health concern. The present study investigates whether probability based artificial intelligence (AI) toxicity predictions align with experimental in vitro findings in human SH-SY5Y neuroblastoma cells. Cells were exposed to PFOA at concentrations ranging from 0 to 2000 µM for 24, 48, and 72 hours, and cell viability was determined using the MTT assay. The resulting IC₅₀ values419.52 µM, 174.97 µM, and 104.64 µM, respectively demonstrated a clear time-dependent increase in apparent cytotoxic potency (~4.01-fold from 24 to 72 h). These empirical data were compared against AI-derived toxicity probabilities from two external platforms: ProTox and CompTox/invitrodb. Calibration between predicted probabilities and observed biological outcomes was assessed using the Brier score. ProTox showed good calibration (Brier = 0.102), whereas CompTox/invitrodb yielded poor alignment (Brier = 0.537), highlighting the importance of endpoint- and time-matched probabilities. The results emphasize that AI models lacking temporal or biological context may underestimate toxicity, particularly when effects manifest gradually over prolonged exposures. This study presents a reproducible, curve-free workflow for integrating AI predictions with time-resolved in vitro toxicity data, providing a framework to enhance biological realism in computational toxicology and guide future PFAS risk assessments.
Keywords : PFOA, SH-SY5Y, Yapay Zekâ, Sitotoksisite

ORIGINAL ARTICLE URL