Machine Learning-Based Interpretation of Operating Conditions for Enriching Polyhydroxyalkanoate-Accumulating Microorganisms

Abstract

Polyhydroxyalkanoate (PHA) is a biodegradable polymer accumulated by microorganisms and is considered a promising alternative to petroleum-based plastics. This study investigated the operating conditions governing PHA accumulation and the enrichment of PHA-accumulating microorganisms in mixed microbial culture (MMC) systems. Data from four sequencing batch reactors (SBRs), operated in two experimental runs with three stages each, were integrated, including 1,733 days of process data, 400 PHA measurements, and 16S rRNA gene amplicon sequencing data from 48 samples. A leakage-free machine learning framework was applied to interpret the relationships among operating conditions, PHA accumulation, and microbial community dynamics. Under reactor-wise cross-validation, XGBoost explained PHA content from operating conditions alone (R2 = 0.38), and SHAP analysis identified settling strategy, nutrient decoupling, solids retention time (SRT), and feast/famine duration as key factors. PHA-accumulating genera, including Thauera, Paracoccus, and Azoarcus, were enriched compared with the inoculum, but their relative abundance was not significantly correlated with measured PHA content (Spearman’s ρ = −0.34, p = 0.06). In contrast, Thauera abundance was predicted from operating conditions in unseen reactors with relatively high accuracy (R2 = 0.61), whereas most other taxa were not predictable. These results suggest that operating control selectively shapes key PHA-associated microorganisms rather than the entire community. This study provides a data-driven basis for defining operating windows for MMC-based PHA enrichment and AI-assisted process optimization.