The Materials That Should Exist
Biomaterials are shaped by interacting chemical, structural and biological factors that are difficult to optimise one at a time. We use experimental data and computational models to map these relationships, predict outcomes and identify candidate formulations for experimental testing.
Designing biomaterials across interacting variables.
A biomaterial is not defined by composition alone. Its chemistry, structure and biological environment interact over time, and changes in one feature can affect several outcomes.
Our workflow treats these relationships as a connected design problem. Experimental material and biological data are used to identify influential variables, build predictive models and search for formulations that are more likely to meet predefined targets.
Each round of experiments provides new information about the design space and helps determine which measurements, formulations or conditions should be examined next.
A data-guided workflow for biomaterials design.
The workflow moves through four linked stages. MAP identifies influential variables in the design space. PREDICT uses measured data to estimate later or unmeasured outcomes. PRESCRIBE searches for formulations that match predefined targets, and VALIDATE compares those predictions with new experimental results.
Reducing the experimental search space.
Testing several formulation variables at multiple levels quickly generates a large number of possible combinations. Orthogonal arrays provide a structured way to sample that space and estimate selected effects with fewer runs than a full factorial design. In the example below, the objective is to examine how formulation variables influence setting time and compressive strength.
Linking material features to experimental outcomes.
Composition, porosity, architecture and interfaces can be represented as measurable features and analysed alongside material properties and cell-response outcomes. This creates a common data structure in which relationships between material features and performance can be examined.
Models trained on these relationships can estimate outcomes for untested combinations and rank candidate formulations. These predictions should remain within the range supported by the available experimental data.
Selecting variables that improve prediction.
Adding more measurements does not necessarily improve a predictive model. For a given outcome, a subset of variables may contain most of the useful predictive information, while additional inputs contribute little. This module shows how predictive performance changes as variables are introduced one at a time.
Choose an application, then add factors to see how each one changes predictive performance.
Predicting 28-day pH from early measurements.
As calcium-silicate cement hydrates, hydroxyl ion release raises the pH of the surrounding environment. Early pH measurements capture part of that developing alkalinity. Here, measurements at 3 h and 24 h, together with specimen surface area, are used to estimate the subsequent 28-day pH profile and compare it with experimental measurements.
The pH target at 28 days. The shaded band marks the target range.
Optimising formulations under competing property targets.
Material properties do not always improve in the same direction. Increasing radiopacifier content can improve radiopacity while reducing strength, while increasing water can improve flow but delay setting. The genetic algorithm therefore evaluates candidate formulations against several selected property targets at the same time.
Each candidate formulation is scored against the target profile. Better-matching candidates contribute to subsequent generations, shifting the population toward regions of the design space that better satisfy the selected constraints. The result is a search for the best overall compromise rather than a formulation that maximises every property independently.
Set the property targets, then press PLAY to evolve the candidate formulations.
Extending the design workflow toward translation.
MAP, PREDICT, PRESCRIBE and VALIDATE describe the current research workflow. New experimental results can be returned to the dataset to refine subsequent predictions and design decisions. TRANSLATE represents the longer-term extension of this process, in which validated materials and models are evaluated under increasingly clinically relevant conditions.
04 STAGES + TRANSLATE
Design space. The cube shows three factors at three levels; the nine amber points are the runs actually tested.
Structured experimental designs identify variables associated with material performance.
Integrating multiple dimensions of cell response.
A material can perform differently across proliferation, migration and mineralisation assays, so no single endpoint provides a complete description of its in vitro biological response. The Digital BioScore combines these measurements into an integrated profile while preserving the contribution of each individual domain. The holographic representation is used to visualise how these domains vary together across materials.
Future directions: patient-specific prediction and automated discovery.
The next stage of the programme is to connect material design with patient-level information and increasingly automated experimentation. The longer-term objective is to move from formulations developed for average conditions toward candidates selected according to the biological and clinical constraints of individual cases.
Models that combine material characteristics with patient-level data to estimate how candidate materials may perform under the conditions of a specific case.
Experimental systems that can propose, test and refine candidate formulations iteratively, while scientists define the biological question, objectives and constraints.
Published evidence supporting the workflow.
The current workflow builds on three published studies that address different stages of biomaterials design. One used structured experimental design to tune material properties, another used early measurements to predict 28-day alkalinity, and a third used data-driven optimisation to identify formulations matching predefined property targets.
The role of AI in the experimental workflow.
Across the programme, AI and statistical modelling are used to organise complex design spaces, identify informative variables, estimate outcomes and search for candidate formulations. The scientific questions, experimental constraints and interpretation remain defined by researchers, and computational predictions are evaluated against laboratory evidence.