Currently, food products are formulated by a resource-intensive separation of agro-materials into pure components (isolates). These ingredients are subsequently combined again in food products. Alternatively, impure mildly refined ingredients require fewer resources but are difficult to use in product formulation. Machine learning can predict the properties of these impure ingredients, which facilitates their selection process. This study aims to formulate sustainable ingredient formulations based on predicted techno-functional properties instead of purity (composition). Extensively and mildly refined ingredients from yellow peas and lupine seeds are matched to a product portfolio based on gelation, viscosity, emulsifying stability, and foaming capacity. Formulations based on techno-functional properties include more milder refined ingredients and result in up to 70% reduced global warming potential, water usage, and raw materials compared to formulations based on composition and isolates only. This approach contributes to reducing the environmental impact of the food chain and should be extended to other ingredients or properties.