The achievement is notable because of what the protein lacks. Most well-studied sweet proteins carry stabilizing features that honey truffle sweetener simply does not have. Thaumatin, mabinlin, and brazzein contain multiple disulfide bonds and mixtures of alpha-helices and beta-sheets; neoculin is reinforced by extensive disulfide crosslinks; even monellin, which is largely beta-rich, retains a short stabilizing helix. Honey truffle sweetener, by contrast, is cysteine-free and helix-free—a compact bundle of beta-strands held together by sheet hydrogen bonding and loop geometry. Comparisons of residue-contact density showed the fold is not uniformly loose; its overall packing is comparable to thaumatin and denser than brazzein. But cavity analysis revealed a larger pocket volume per residue than reference proteins, and the absence of classic stabilizing elements left the protein vulnerable to heat.

The first step was to see the molecule in atomic detail. The team crystallized the protein and determined its structure at 1.58 angstrom resolution using synchrotron radiation at the Shanghai Synchrotron Radiation Facility, solving the structure by molecular replacement with an AlphaFold-predicted model as the search template. The structure revealed a characteristic arrangement of seven major beta-strands forming a slightly twisted antiparallel sheet, flanked by three shorter strands—a so-called 7-plus-3 topology—connected by solvent-exposed loops, including an extended and apparently flexible loop between strands three and four. This high-resolution map became the common reference for every subsequent design decision, from mutation evaluation to disulfide-bond planning.

With the structure in hand, the researchers screened every one of the 121 positions for potentially stabilizing substitutions using two complementary computational methods. Rosetta, a physics-based framework, estimated the energetic effect of each mutation after side-chain repacking and constrained relaxation, while ThermoMPNN, a deep-learning model, judged whether each amino acid substitution would be compatible with the local structural environment. From the overlapping and divergent predictions, the team selected 72 single-point variants for experimental testing. The results validated the strategy: 41 of the 72 variants showed a positive shift in melting temperature, and 27 exceeded a one-degree gain. The single best mutations, Q37A and S55G, raised the melting temperature by 3.91 and 4.23 degrees respectively.

But the raw computational scores turned out to be only modestly predictive. Correlations between predicted scores and measured stability changes across the 72 variants were weak, with Pearson coefficients of just 0.206 for Rosetta and 0.248 for ThermoMPNN. That gap is precisely where the study’s most innovative element comes in: a machine-learning framework the authors call SMRC-Net, which learns from the experimental data itself. The system combines a baseline melting-temperature estimate derived from a pretrained protein language model, ESM2, with a residual correction built from molecular dynamics descriptors—flexibility changes, solvent accessibility, residue contacts, and global perturbation measures calculated from simulations of the wild-type protein and 65 mutants.

Under repeated strict nested cross-validation, the sequence-only baseline explained about 44 percent of the variance in melting temperature, with a mean absolute error of 1.35 degrees. Adding the molecular dynamics residual correction lifted the explained variance to nearly 60 percent, cut the error to 1.15 degrees, and improved the rank correlation of variants from 0.540 to 0.698. When the frozen model was tested prospectively on ten Rosetta-designed mutants it had never seen, it correctly predicted the direction of the stability change for nine of the ten—a 90 percent directional accuracy. The single miss, T115I, was predicted to stabilize the protein but produced a small experimental decrease of 0.67 degrees.

Armed with validated stabilizing mutations, the team assembled them stepwise rather than all at once, a modular strategy that kept each combination experimentally testable. Ten substitutions were grouped into four modules of two or three mutations, each verified to raise the melting temperature before further assembly. Combining the three-residue modules produced a six-site variant at 67.77 degrees, and the gains proved nearly additive—the measured improvement differed from the sum of individual contributions by only 0.12 degrees. Two parallel assembly routes then produced eight-site intermediates, and their union yielded the ten-site variant Mut10-1 at 73.03 degrees, 18.43 degrees above the wild type.

The final flourish was structural rather than chemical: an engineered disulfide bond. Screening candidate cysteine pairs by geometric criteria with Rosetta’s DisulfidizeMover identified Q37C/D46C as the most favorable linkage. Introducing that single bond added another 6.95 degrees on its own, and when combined with the ten mutations it produced the final variant, FM, with a melting temperature of 82.07 degrees. Crucially, sensory evaluation by eight trained panelists showed that both Mut10-1 and FM retained sweetness thresholds similar to the wild-type protein—evidence that the stabilizing changes did not disrupt the receptor-binding surface responsible for the protein’s intense taste.

Practical stress tests underscored the difference. After one hour at 80 degrees Celsius, the wild-type protein had largely precipitated out of solution, while FM remained predominantly soluble; even after four hours, FM retained a substantial soluble fraction. Circular dichroism spectroscopy showed that FM’s secondary structure was largely preserved through the prolonged heating, and the engineered protein also stayed soluble and structurally intact across acidic, neutral, and alkaline conditions at pH 3, 6, and 9. Molecular dynamics simulations suggested why: the mutations did not rigidify the protein wholesale but redistributed local flexibility, modestly remodeled the residue-contact network, and stabilized a single dominant folded ensemble. An experimental test supported one long-range interaction in particular—disrupting the H16-F99 contact with a double mutant lowered the melting temperature by 5.2 degrees, confirming its role in tertiary stability.

The authors are careful about the limits. Sweetness was tested only on untreated samples, so whether the engineered protein survives heating and pH shifts while still tasting sweet in real foods remains unverified, and sensory testing in complex food matrices is the necessary next step. The SMRC-Net predictor is also calibrated specifically to this scaffold, though the authors argue the residual-learning architecture could be retrained for other compact proteins where molecular dynamics remains computationally feasible. Still, the demonstration is striking: a cysteine-free, helix-free beta protein with no natural stabilizing armor was transformed into a heat-tolerant, sweetness-preserving molecule through a coordinated workflow of crystallography, complementary computation, machine learning, and disciplined stepwise assembly. For a protein once undone by a warm afternoon, that is a remarkable turnaround—and a template for engineering the next generation of sugar substitutes.

Subject of Research: Computation-guided protein engineering to improve the thermal stability of the fungal sweet protein honey truffle sweetener

Article Title: Integrated Computation-Guided Thermostabilization of the Fungal Sweet Protein Honey Truffle Sweetener

Article References: Wang, Z., Wang, W., Zhu, F., Zhang, Y., Li, Y., Zhu, Z., Lu, Z., Huang, A., Yu, M., & Liu, S. (2026). Integrated computation-guided thermostabilization of the fungal sweet protein Honey Truffle Sweetener. Current Research in Food Science, 13, Article 101584. https://doi.org/10.1016/j.crfs.2026.101584

Image Credits: AI Generated

DOI: 10.1016/j.crfs.2026.101584

Keywords: sweet proteins, honey truffle sweetener, protein engineering, thermostability, crystal structure, Rosetta, ThermoMPNN, molecular dynamics, machine learning, disulfide bond, sugar substitute, food science