Multi-approach optimization of exopolysaccharide production from <i>Leuconostoc mesenteroides</i> ABNFT-1 using statistical and AI-based model

Exopolysaccharides (EPS) are high-molecular-weight polymers secreted by microorganisms into their extracellular environment. In this study, the isolate ABNFT-1 exhibited the highest EPS production (11.51 ± 0.35 g/L) after 48 h of incubation under static conditions. On the basis of cultural, morphological, and biochemical characterization, along with 16S rRNA sequencing, the isolate was identified as <i>Leuconostoc mesenteroides</i> ABNFT-1 (GenBank accession number: PV465880). The study aimed to enhance EPS production through a systematic, multi-step optimization strategy involving one-factor-at-a-time (OFAT) screening, response surface methodology (RSM), and artificial neural network (ANN) modeling. Critical process parameters such as inoculum size, sucrose concentration, potassium nitrate, and ferrous sulfate were initially screened using OFAT and then optimized using central composite design (CCD) under RSM. The optimal conditions-10% inoculum size, 20% sucrose, 4% potassium nitrate, and 0.04% ferrous sulfate led to a significantly improved EPS yield of 21.735 ± 0.76 g/L. Further modeling with ANN using the Levenberg-Marquardt Algorithm (LMA) predicted an even higher EPS yield of 23.66 g/L. Additional comparisons with the Scaled Conjugate Gradient Algorithm (SCGA) and Bayesian Regularization Algorithm (BRA) showed that SCGA achieved strong performance in training (<i>R</i> = 0.98) but weaker validation (<i>R</i> = 0.50), while BRA, despite offering better generalization and overfitting control, did not outperform LMA in prediction accuracy. These findings highlight the advantage of hybrid optimization strategies that integrate machine learning with classical statistical techniques for maximizing EPS production, offering promising applications for industrial biopolymer synthesis.