Key points are not available for this paper at this time.
numerous sequences is time-consuming, labor-intensive, and costly. To address these limitations, computational approaches based on machine learning have been developed for rapid prediction of enzyme pHopt, supporting applications in protein engineering.Recent advances, exemplified by EpHod (enzyme pH optimum prediction with deep learning) (Gado et al., 2025), enable prediction of the enzyme pHopt directly from protein sequences. The EpHod model leverages embeddings from the protein language model (PLM) ESM-1v and achieves a root mean squared error (RMSE) of 1.25 pH units on the held-out test data (Gado et al., 2025). To further improve predictive performance, an increasing number of AI-driven tools have been developed.For instance, Zhang et al. introduced the model VENUS-DREAM, which employs the PLM ESM-2 and reduces the RMSE to 0.809 (Zhang et al., 2025). These AI-powered tools are revolutionizing enzyme discovery and design by enabling high-throughput prediction of pHopt.Similarly, studies of microbial adaptation to environmental pH frequently require knowledge of enzyme pHopt. This information can be used to investigate the underlying adaptive mechanisms. However, experimental determination of pHopt for large-scale enzyme sequences remains impractical due to high costs and low throughput. Fortunately, the high-throughput predictive capacity of these AI-driven (Gado et al., 2025), leverages embeddings from the ESM-1v protein language model (PLM) and can directly predict the enzyme pHopt from protein sequences. It achieves a root mean squared error (RMSE) of 1.25 pH units on the held-out test set (Gado et al., 2025), showcasing its superior predictive capabilities. Serendipitously, these tools can be employed to explore how organisms respond and adapt to pH fluctuations, potentially yielding breakthroughs in metabolic flexibility. In this context, we use differentially expressed isoenzymes as case studies to discuss how these AI-driven tools facilitate inferences about adaptive strategies microorganisms employ in response to pH variations.To illustrate the value of AI-powered enzyme pHopt prediction in understanding microbial pH adaptation, we selected three enzymes whose pHopt values were accurately predicted by EpHod. The first example is the acid protease Thermopsin (UniProt: P17118), purified from the culture supernatant of Sulfolobus acidocaldarius strain DSM 639 (Lin and Tang, 1990). S. acidocaldarius strain DSM 639 was isolated from Locomotive Spring in Yellowstone National Park, USA, where the spring has a pH of 2.4, and exhibits optimal growth at pH 2-3 (Brock et al., 1972). Thermopsin acts as a secreted enzyme that digests extracellular proteins to provide nutrients for the cell (Rawlings, 2013;Tang and Lin, 2013). A logically relevant question is whether secreted Thermopsin can function effectively under such acidic conditions and thereby contribute to the adaptation of strain DSM 639 to its acidic habitat.Notably, the pHopt of Thermopsin predicted by EpHod is consistent with experimental observations, which showed that this enzyme displays maximal proteolytic activity at pH 2 (Lin and Tang, 1990) -a value that agrees well with the in situ environmental pH. This consistency between the AI-predicted and experimentally determined pHopt supports that EpHod can help reveal microbial niche adaptation, which is further reinforced by the second example presented below.The second example is the acid-resistant membrane-bound amylopullulanase Apu (UniProt: Q4J9M2), which is also produced by S. acidocaldarius strain DSM 639.The purified Apu catalyzes the hydrolysis of starch, amylopectin, and pullulan into small oligomers for cellular uptake (Choi and Cha, 2015). EpHod again yielded an accurate pHopt prediction for Apu, which matched the experimentally determined pHopt value of approximately 3.0 for its amylolytic activity toward starch. The low pHopt values of both Thermopsin and Apu ensure their catalytic activity under the acidic growth conditions of strain DSM 639, conferring a distinct physiological advantage for strain DSM 639 thriving in its native habitat.The third example is an extracellular pectate lyase PelA (UniProt: D0VP31) from the alkaliphilic bacterium Bacillus sp. N16-5. The strain N16-5 was isolated from sediment of Wudunur Soda Lake in Inner Mongolia, China, and grows well within a pH range of 8.5-11.5 (Ma et al., 1991). The purified PelA from the culture broth of Bacillus sp. N16-5 efficiently depolymerizes polygalacturonate and pectin with digalacturonate and trigalacturonate as the main products (Li, 2010), and transcriptomic analysis of Bacillus sp. N16-5 grown on pectin indicated that these products could be subsequently imported into cells for metabolism, thereby supporting carbon acquisition (Song et al., 2013). EpHod also accurately predicted the pHopt of PelA, which was in good agreement with the experimentally determined value of 11.5 (measured at 50 ˚C using polygalacturonic acid as the substrate) (Li, 2010). The high pHopt of PelA enables its robust function in alkaline environment, thereby promoting the survival and adaptation of Bacillus sp. N16-5 in its ecological niche. modifications, including acetylation, phosphorylation, glycosylation, methylation, and ubiquitylation (Macek et al., 2019), and are ubiquitous within cells (Beltrao et al., 2012). Enzyme pH-activity profiles can be significantly modulated by chemical modifications (Xue et al., 2010;Li, 2014;Giri et al., 2021). Therefore, to improve the accuracy of pHopt prediction, it is essential to equip predictors with the ability to incorporate PTMs. A promising strategy for integrating PTM information into embedding processes is to use PTM-aware PLMs. A critical step in this direction is the precise mapping of PTM annotations to corresponding protein sequences.Recently, Peng et al. proposed PTM-Mamba, a PLM that integrates PTM tokens using bidirectional Mamba blocks fused with ESM-2 embeddings via a gating mechanism (Peng et al., 2025). PTM-Mamba has strong potential as a foundational PLM for various downstream PTM-aware tasks, with enzyme pHopt prediction being a prominent application.It should be noted that in addition to pH, other environmental factors, such as temperature or ionic strength, can also influence enzyme activity (Ishigami and Morita, 1977;Ivanov et al., 2026); thus, experimental validation remains indispensable. Direct experimental determination of pH optima provides solid and reliable results that cannot be replaced. On the other hand, as more experimentally measured pHopt values accumulate and are incorporated into training datasets, the performance of such AI predictors will continue to improve.Microorganisms have evolved diverse strategies to adapt to environmental pH. As essential biological catalysts, enzymes play a pivotal role in biochemical reactions.Eukaryotic microbes also exploit isoenzyme regulation for pH adaptation. Take the yeast Komagataella phaffii as an example. When exposed to alkaline stress, it exhibits differential expression patterns across multiple pairs of genes encoding isoenzymes (Albacar et al., 2023). For instance, genes responsible for β-1,3glucanosyltransferases (required for cell wall assembly) and those encoding peroxisomal 2,4-dienoyl-CoA reductase (an auxiliary enzymes in fatty acid beta-oxidation) exhibit markedly inverse transcriptional responses. By forecasting the pHopt for isoenzyme activity, these pHopt predictors empower researchers to link pHopt to transcriptional responses and physiological functions and conduct hypothesis-driven explorations of isoenzyme roles under different pH conditions. Moreover, the advanced pHopt predictors enable us to infer how biological communities respond to pH variations at the molecular level. The gut microbiome, for example, inhabits a dynamic environment where pH can fluctuate due to factors such as diet, disease, and medication. These predictors can help explain how the gut microbiome meticulously adjusts the expression of isoenzymes to cope with these pH changes.Similarly, for the plant-rhizosphere microbial system, pHopt predictors can illuminate how the rhizosphere microbial communities react to pH changes by fine-tune expression. Furthermore, by utilizing pHopt prediction tools, we can better hypothesize the underlying mechanisms that marine ecosystems rely on to buffer the impact of acidification by fine-tuning expression of isoenzyme with different pHopt.Line spacing: Double Formattedhockeytown: Knowledge of enzyme pHopt is a key to understanding the dynamic interplay between enzymes, pH, and biological adaptation. Investigating pHopt is therefore highly valuable for uncovering the mechanisms of microbial adaptation. However, experimental determination of pHopt for large-scale enzyme sequences (e.g. those annotated from genomes or metagenomes) is highly time-consuming, labor-intensive, and expensive. Using these AI-powered prediction tools enables time-efficient, costeffective, and high-throughput prediction of pHopt values. Here we discuss the application potential of advanced AI-driven pHopt predictors in studying the adaptive strategies employed by microorganisms to cope with environmental pH.High-throughput predictions facilitate researchers in inferring the functional roles of enzymes in pH adaptation and conducting hypothesis-driven studies to explore the adaptive strategies of microorganisms. For example, from a pHopt perspective, the significance of isoenzymes that catalyze identical reactions but exhibit distinct pHopt values can be systematically evaluated.Moreover, such high-throughput predictions allow researchers to investigate the competitive advantages of different organisms under specific environmental pH conditions and their implications for species succession. For instance, elevated pH differentially affects the growth and survival of various marine phytoplankton species (Hansen, 2002;Zepernick et al., 2021). Based on pHopt prediction, the molecular mechanisms underlying the competitive advantages of alkalitolerant bloom-forming algal species under elevated water pH can be analyzed, improving the mechanistic understanding of algal bloom outbreaks. Furthermore, ocean acidification poses a severe threat to marine ecosystems, especially to calcifying organisms (Orr et al., 2005). Coral reefs, among the most vulnerable marine ecosystems, are being severely impacted (Hoegh-Guldberg et al., 2007). Using AI-predicted enzyme pHopt values, we can analyze the match between enzyme pHopt and fluctuating seawater pH, thereby inferring species adaptability and the adaptive succession of biological communities under ocean acidification.In summary, enzyme pHopt prediction tools enable a deeper understanding of the mechanisms by which organisms cope with environmental pH. As the field of AI-driven pHopt prediction continues to advance, novel insights will be gained to facilitate the elucidation of pH adaptation mechanisms, potentially leading to important scientific breakthroughs.Microorganisms across the spectrum from prokaryotes to eukaryotes are exposed to fluctuating pH levels, which influence the catalytic activity of enzymes as well as their regulation. Studies on adaptation strategies to pH stress reveal that certain isoenzymes exhibit opposing expression patterns, suggesting their regulation as a mechanism for pH adaptation. Beyond revolutionizing enzyme engineering, recent advancements in the field of AI-driven pHopt prediction is also poised to deepen our understanding of the dynamic interplay among enzymes, pH, and biological adaptation. Furthermore, the advanced pHopt predictors enable us to ecologically infer how microbial communities respond to pH variations at the molecular level. The gut microbiome, for example, inhabits a dynamic environment where pH can fluctuate due to factors such as diet, disease, and medication.These predictors can help explain how the gut microbiome meticulously adjusts the expression of isoenzymes to cope with these pH changes. Similarly, as for plant-rhizosphere microbial system, pHopt predictors can illuminate how the plant and its associated rhizosphere microbial communities react to pH changes by fine-tune expression. Furthermore, ocean acidification poses a severe threat to marine ecosystems, particularly calcifying organisms (Orr et al., 2005). Coral reefs, as one of the most vulnerable marine ecosystems, are suffering greatly (Hoegh-Guldberg et al, 2007). By utilizing pHopt prediction tools, we can better hypothesize the underlying mechanisms that marine ecosystems rely on to buffer the impact of acidification, such as the fine-tuning of isoenzyme expression in marine microorganisms.Enhancing predictive accuracy will facilitate the use of pHopt predictors, enabling them to achieve broader and more effective applications. For instance, chemical modifications to enzymes can be integrated during the embedding process. The pH-activity profile of an enzyme can be altered through chemical modifications (Giri et al., 2021;Xue et al., 2010;Li, 2014). To fulfill essential biological functions, enzymes often undergo post-translational modifications (PTMs). These PTMs encompass a wide range of chemical changes, including acetylation, phosphorylation, glycosylation, methylation, and ubiquitylation (Macek et al., 2019), and PTMs are ubiquitous within cells (Beltrao et al., 2012). Given this, when striving to improve the accuracy of pHopt predictions, it becomes imperative to equip predictors with the capability to account for chemically modified residues. One promising approach is to employ PTM-aware PLMs. For instance, a key step in this process is to map PTM annotations precisely to their corresponding protein sequences. Recently, Peng et al. made a significant contribution in this area by introducing PTM-Mamba (Peng et al., 2025). It integrates PTM tokens using bidirectional Mamba blocks fused with ESM-2 PLM embeddings via a gating mechanism. PTM-Mamba has the potential to serve as a foundational tool for a variety of downstream PTM-aware tasks, with pHopt prediction being a Deletedhockeytown:
Zhou et al. (Tue,) studied this question.