Multi-scale machine learning reveals the effects of sulfur species on sludge anaerobic fermentation: Bridging macroscopic performance and microscopic molecular mechanisms.

Sulfur-mediated chemical regulation has emerged as an attractive strategy for enhancing sludge anaerobic fermentation (AF). However, the non-linear influence of diverse sulfur species on AF performance, and their cross-scale interactions with microbiomes and functional genes, remain poorly understood. Therefore, this study developed a multi-scale machine learning (ML) framework that integrates MacroML for AF performance prediction and MicroML for mechanistic microbiome-gene interpretation, using thiosulfate-mediated AF as a representative sulfur-regulated system. At the macroscopic level, support vector regression (SVR) exhibited superior predictive robustness for short-chain fatty acids (SCFAs) (R2-test = 0.94) and biogas (R2-test = 0.89) using multiple sulfur species and operational conditions as inputs. MacroML interpretation analysis identified H2S (>3000 ppm), thiosulfate (∼650 mg S/L), sulfite (3.0-3.5 mg S/L), sulfide (120-140 mg S/L), and sulfate (20-30 mg S/L) as determinants for increasing SCFA production, whereas biogas production required low thiosulfate (<200 mg S/L), sulfate (< 10 mg S/L), sulfide (10-20 mg S/L), sulfite (< 2.0 mg S/L), and H2S (< 2500 ppm). At the microscopic scale, a graph embedding-enhanced random forest (GE-RF) model (R2-test = 0.72-0.89) revealed that sulfur metabolism was associated with a metabolic shift toward acidogenesis, as indicated by the enrichment of core taxa (e.g., Proteiniphilum, Aminobacterium), higher expression of acidogenic genes (e.g., buk, ptb), and increased abundance of electron transfer chain components (e.g., aprA/B, COQ4). Concurrently, sulfur exposure (e.g., thiosulfate, sulfide, H2S) was associated with inhibition of terminal methanogenic enzymes (i.e., mcr and mtr) and a potential weakening of Methanothrix-centered functional modules, which may collectively contribute to reduced biogas production. This work provides a robust, data-driven roadmap for deciphering sulfur-mediated biochemical pathways and facilitates precision operational control in sludge AF systems.