August 25, 2026

Closing the Egg Replacement Gap

Ingredion researchers used machine learning to narrow the formulation space for egg-reduced baked goods, but human sensory expertise remained essential to getting the texture right.

Person preparing food by beating eggs in a bowl

Food Technology Magazine

Replacing eggs in baked goods is not a matter of finding a single substitute. Eggs aerate, bind, emulsify, and gel during baking, while also contributing color, aroma, and flavor. That combination of functions makes high levels of egg replacement a particularly difficult formulation challenge.

At the IFT FIRST Annual Event and Expo, Liyi Yang, business scientist with Ingredion’s global applications team and technical lead for bakery and snack products, described a data-driven approach designed to make that challenge more manageable. The goal is not to ask a model to invent an egg replacer from scratch, but to use machine learning to map ingredient functionality and narrow the number of formulations food scientists need to test at the bench.

The team began by comparing full-egg reference products with egg-reduced pancakes and pound cakes. As egg levels dropped, pancakes lost volume and developed a coarser crumb with fewer, larger air bubbles. A 50% egg-reduced pound cake was softer and more cohesive, with sensory panelists also noting more chalkiness, tooth packing, and mouth coating.

Modeling the Texture Gap

Because egg proteins help set structure during baking, the researchers chose gelation as the first function to model. They screened more than eight proteins and more than nine starch and texturizer ingredients, measured their gelling properties, and used those data to build machine learning models tied to cake outcomes such as firmness, volume, cell size, and air bubble formation.

The relationships were not linear. Stronger gels generally correlated with firmer cakes, while viscosity changes during heating and cooling had a major influence on volume. The timing of gel formation affected cell size and the balance between expansion and structure setting.

Those relationships allowed the team to create a virtual formulation space and then select promising points for physical validation. Yang said exploring the same space without predictive modeling could require months of baking. With the model guiding where to look, the team moved formulations toward an acceptable range with about 20 runs.

Predictive models can narrow the formulation space, but human expertise is still required to judge the attributes the model cannot capture.

Where Human Expertise Takes Over

The model was much less successful when the researchers tried to predict the full sensory experience. It produced only a weak model for firmness, in part because gelation was only one of the functions being modeled. Aeration and emulsification were not included, and many sensory characteristics do not have a direct instrumental measure.

“Not every sensory attribute can be measured with an instrument and predicted with machine learning tools,” Yang said. “So we still rely on human expertise to fine tune and validate.”

That became clear in two protein case studies. One protein produced good firmness but poor volume; another delivered good volume but poor firmness. Machine learning helped identify how texturizers influenced each system differently, while sensory evaluation guided the formulation changes that followed.

The result is less a universal egg replacer than a method for navigating a complicated formulation problem. Predictive models can rule out less-promising options and help scientists focus on the functionality they need. Human expertise is still required to judge the attributes the model cannot capture—and ultimately decide whether the product delivers the desired eating experience.