AI-designed proteins could unlock a new generation of sweetness

Scientists at The Pennsylvania State University have used artificial intelligence protein design tools to invent brand-new proteins that could, in principle, taste sweet. In a study published in npj Science of Food, Saisai Ding and Yi Zhang of the Department of Food Science describe a complete computational pipeline that designs protein binders against the human sweet taste receptor, the molecular sensor on the tongue that detects sugar and signals sweetness to the brain. The work is entirely computational at this stage, but it lays out a rational strategy for creating a new generation of high-potency sweeteners that do not come from plants, bacteria, or any natural source at all.

The motivation behind the study is the growing global effort to reduce dietary sugar intake without sacrificing the sensory pleasure of sweetness. Natural sweet proteins, such as those found in certain berries and West African plants, offer extraordinary sweetness potency, meaning they can be hundreds to thousands of times sweeter than sucrose on a per-weight basis. Yet their usefulness is constrained by nature’s limited diversity. Only a handful of such proteins exist, and each brings its own baggage: difficulties in large-scale production, poor stability under heat or acidic conditions, and formulation challenges in real foods and beverages. Rather than waiting for nature to reveal another sweet protein, the Penn State team asked whether modern protein design software could simply write one from scratch.

Their target was not arbitrary. The human sweet taste receptor is a heteromeric complex composed of the TAS1R2 and TAS1R3 proteins, and the researchers focused on the extracellular domain of TAS1R2, the subunit that carries the Venus flytrap domain where sweet molecules lodge. This domain, named for the hinged clamshell shape it shares with related receptors, opens and closes around sweet ligands, and the adjacent cysteine-rich domain helps transmit that binding event into the cellular signal that the brain perceives as sweet. Any protein that grips these structurally relevant regions has at least a chance of mimicking what sucrose, advantame, and natural sweet proteins do.

The design workflow combined several of the most powerful tools in contemporary computational biology. First, RFdiffusion generated candidate protein backbones, the three-dimensional scaffolds of amino acid chains, shaped to complement the receptor surface. Then ProteinMPNN assigned amino acid sequences to those backbones, choosing residues likely to fold into the intended shape and form stable contacts with TAS1R2. Boltz-1, a structure prediction model, evaluated whether each designed sequence would actually adopt the intended fold and dock correctly against the receptor. Finally, a static GBSA energetic assessment estimated the binding energetics of each candidate, and molecular dynamics simulations tested how the complexes would behave over time in a realistic, fluctuating molecular environment. This layered approach allowed the team to prioritize a small set of designs from a vast theoretical design space.

Three designed binders emerged as the leading candidates, and each told a different story. Binder2 showed the strongest structural confidence in the Boltz-1 evaluation, suggesting its predicted fold and receptor docking were the most self-consistent. Binder1, by contrast, delivered the most favorable static GBSA energy estimate, indicating the most favorable calculated binding energetics. Binder3 completed the trio as a comparative case. The fact that the two ranking metrics disagreed between candidates is itself informative, the authors note, because it demonstrates that structural plausibility and energetic favorability capture different aspects of binding quality and should be weighed together rather than in isolation.

To probe these differences in depth, the researchers ran 200-nanosecond molecular dynamics simulations on each binder-receptor complex, using a structure of TAS1R2 bound to the high-intensity sweetener advantame as a structurally defined reference for trajectory-based analyses. The simulations revealed distinct interaction profiles for each design. Binder1 formed the largest and most interaction-rich interface with the receptor, building an extensive web of contacts across the binding surface. Binder2 maintained a persistent mixed network of polar and nonpolar interactions, a balanced chemical handshake that persisted throughout the simulation. Binder3 showed greater conformational rearrangement and a more restricted interface, suggesting its grip on the receptor was less stable and its structure less rigidly maintained over time.

These interaction profiles matter because receptor activation is a dynamic process, not a static snapshot. A sweet molecule must hold the Venus flytrap domain in the correct conformation long enough to trigger the signaling cascade. A binder whose interface dissolves under thermal motion, or whose binding surface shifts and shrinks during simulation, is less likely to activate the receptor even if its initial docking pose looks convincing. By comparing each binder’s behavior against the advantame-bound reference, the team could rank candidates not just by how well they docked but by how their contacts endured, redistributed, and evolved over the simulated trajectory.

The authors are careful to frame the results as a demonstration of feasibility rather than a claim of sweetness. No designed binder has yet been synthesized, expressed, or tested on taste cells, and the study does not report receptor activation or sensory data. What the work does establish is that receptor-guided de novo design can generate structurally diverse proteins that bind plausible regions of TAS1R2, and that the computational framework supports rational candidate prioritization before any laboratory resources are spent. The next step, the researchers say, is experimental validation of receptor activation and sensory sweetness for the top-ranked designs, work that will determine whether any of these virtual proteins actually tastes sweet on the human tongue.

If the pipeline validates, the implications reach well beyond the laboratory. Sweet proteins designed to order could sidestep the production and stability problems that plague natural sweet proteins, since synthetic sequences can be engineered for heat tolerance, pH resilience, and clean taste profiles from the outset. They could also be optimized for potency, reducing the amount of protein needed per serving. The same receptor-guided strategy could be adapted to other taste receptors, opening a general route to designed proteins that modulate bitterness, umami, or saltiness. In the longer view, the study hints at a food science in which flavor molecules are not discovered but authored, with AI design tools writing protein sequences the way a composer writes a score.

The research also reflects a broader shift in how the food and health sciences approach molecular discovery. Tools like RFdiffusion, ProteinMPNN, and Boltz-1 were developed largely for drug discovery and structural biology, and their application to taste receptors shows how rapidly these methods are migrating into food science. The Penn State work, supported by the Rasmussen Endowment, the USDA National Institute of Food and Agriculture, and Hatch Appropriations, positions computational design as a first screen that can compress years of trial-and-error screening into weeks of simulation. For a field searching for alternatives to sugar, that compression could be transformative, turning the dream of a designed sweet protein from a speculative idea into a testable candidate list ready for the bench.

Subject of Research: De novo design of protein binders targeting the human sweet taste receptor as potential sweet proteins

Article Title: De novo design of protein binders targeting the human sweet taste receptor as potential sweet proteins

Article References: Ding, S., & Zhang, Y. (2026). De novo design of protein binders targeting the human sweet taste receptor as potential sweet proteins. npj Science of Food. https://doi.org/10.1038/s41538-026-01202-x

Image Credits: AI Generated

DOI: 10.1038/s41538-026-01202-x

Keywords: protein design, sweet taste receptor, food science, artificial intelligence, molecular dynamics, sweeteners, TAS1R2, computational biology, novo, design, protein, binders