AI-guided enzyme search yields two new cellulases and an eightfold secretion boost in Bacillus subtilis

Researchers have linked the discovery of two previously uncharacterised GH5 endoglucanases to experimental assessment of computational nominations, lead-enzyme development and extracellular production in Bacillus subtilis, according to a study published in Biotechnology for Biofuels and Bioproducts. Cellulases of glycoside hydrolase family 5 are central to lignocellulose deconstruction, but identifying additional GH5 enzymes remains slow. Deep-learning predictors have proliferated in recent years, yet few studies carry their predictions through experimental validation and testing in a production host. The new work addresses that gap by taking sequence-diversity-guided candidates all the way to measurable enzyme activity, mutagenesis and secretory recovery.

The team began with 25,148 initial UniProt entries, which they clustered and filtered using CatPred-based selection combined with homology searches seeded by EglS, a known GH5 endoglucanase, supplemented by two comparative screening routes. This produced a 23-gene panel for experimental testing. Twenty-two candidates were expressed in Escherichia coli and assayed on carboxymethylcellulose (CMC), with EglS retained as a separate control. Two enzymes, AQSP_CL and MACA_CL, for which the authors found no prior biochemical characterisation, were purified alongside EglS and the thermophilic benchmark Cel5A, making them genuinely new entries in the GH5 catalogue.

Characterisation of the purified enzymes showed that both new candidates have short-assay temperature optima of 70 °C. AQSP_CL was most active at pH 5.0 and gave an apparent mass-based CMC kinetic-efficiency estimate of 0.201 U·mL·mg⁻¹, close to the 0.231 U·mL·mg⁻¹ measured for EglS. The thermophilic benchmark Cel5A remained the more thermostable enzyme of the set. These apparent efficiency values are based on CMC rather than purified substrates, a distinction the authors preserve when reporting enzyme properties.

The computational predictions met a mixed reception in the experimental data. CatPred predictions of activity on cellotetraose showed no detected correlation with lysate CMC activity across the eight assayed candidates with reconciled identities, with a Spearman correlation of ρ = 0.10 and p = 0.82, a cautionary result for anyone relying on substrate-level model outputs without wet-lab confirmation. Ensemble docking offered partial discrimination: AQSP_CL had the highest fraction of poses meeting the operational distance criterion, at 52.0% compared with 28.7–39.0% for the other enzymes, although the differences were not resolved at the conformer level and Vina scores did not distinguish the enzymes.

Protein engineering proceeded in two rounds of mutagenesis. In first-round screening, the M65Q variant had the highest measured lysate activity among the tested single substitutions, at 1.61-fold wild type. A few-shot learning model trained on the first-round data then nominated the second-round library, but the additive predictions did not generalise to combinations, underscoring the difficulty of extrapolating from single substitutions to multi-mutant combinations with limited training data.

The final stage moved production into Bacillus subtilis. Screening of signal peptides identified SP3 as the best performer, increasing the extracellular activity of parental AQSP_CL approximately eightfold over the signal-peptide-free construct. The authors are explicit about scope: the CMC and flask-scale measurements establish enzyme properties and extracellular recovery, and AQSP_CL provides a new scaffold for further biomass-deconstruction research, but biomass-conversion performance remains to be evaluated on real lignocellulosic substrates.

Subject of Research: Discovery, engineering, and secretory production of GH5 endoglucanase enzymes for lignocellulose deconstruction

Article Title: Sequence-diversity-guided discovery, model-guided mutagenesis, and secretory production of GH5 endoglucanases in Bacillus subtilis

Article References: Jeong, S.-J., Yoo, M., Seo, H.-D., Song, N.-E., Cho, C.-W., Hong, H.-D., Rhee, Y. K., & Choi, J. W. (2026). Sequence-diversity-guided discovery, model-guided mutagenesis, and secretory production of GH5 endoglucanases in Bacillus subtilis. Biotechnology for Biofuels and Bioproducts. https://doi.org/10.1186/s13068-026-02838-3

Image Credits: AI Generated

DOI: 10.1186/s13068-026-02838-3

Keywords: GH5 endoglucanase, cellulase, lignocellulose, Bacillus subtilis, CatPred, machine learning, enzyme discovery, signal peptide, protein engineering, carboxymethylcellulose, secretory production, UniProt