Access
Login/Register
Supplier Login
How will a new active ingredient affect skin cells? What biological responses might two ingredients produce when used together? Before committing to complex experiments, can researchers identify which approaches are most worth testing?Unilever and its partners hope to provide new tools for addressing these questions through a virtual skin model.
On November 5, during Guangzhou International Beauty Week, Unilever, the Guangzhou Institute of Quality Inspection, the School of Mathematical Sciences at Shanghai Jiao Tong University, and mathAI jointly unveiled DermCosmos and launched collaborative research into AI-driven cosmetics safety and efficacy evaluation. The partners describe DermCosmos as “the world’s first skin-specific virtual cell world model,” designed to connect skin biology, computational prediction and experimental validation.
For the beauty and personal care industry, the initiative’s potential lies in helping researchers assess earlier how an ingredient might work and which experiments should take priority.
From Mapping Skin to Predicting How It Changes
Single-cell and spatial omics are giving researchers a more detailed understanding of skin: which cells are present, what states they are in, and how different cells may interact.Building on these “maps of skin,” DermCosmos seeks to address a further question: how will cells change when exposed to a new intervention?
According to the research team, the model learns from large-scale human skin single-cell transcriptomic data. It brings together molecular cell states, skin biology and intervention conditions to establish connections between cell state, perturbation and response.
Here, a “world model” refers to a dynamic computational system that learns cell states and the patterns governing how they change. The research aims to predict the response to an intervention and explore how cell states may evolve under continued or additional interventions.
“The key to a world model is learning the states of a living system and how those states change in response to interventions,” said Professor Luonan Chen of the School of Mathematical Sciences at Shanghai Jiao Tong University. He also emphasized that model outputs must be tested through real experiments to provide reliable support for scientific research.
The Original Methods and Research Behind the Model
DermCosmos draws on longstanding research by Chen and his team in computational systems biology, causal inference and dynamic prediction. Chen is a chair professor at Shanghai Jiao Tong University’s School of Mathematical Sciences and School of Artificial Intelligence, a chief professor at the Hangzhou Institute for Advanced Study of the University of Chinese Academy of Sciences, and chief scientist at Shanghai-based mathAI. He previously served as executive director of the Chinese Academy of Sciences’ Key Laboratory of Systems Biology. According to Shanghai Jiao Tong University’s public profile, he has published more than 400 papers as a corresponding or co-corresponding author and developed original methodological frameworks, including dynamical network biomarkers (DNB).
In their UniCure study, published in Cancer Cell, the team and its collaborators combine cellular gene expression information with a drug’s molecular structure and dose to predict how cells change after treatment. This research provides a methodological foundation for DermCosmos to connect ingredient characteristics, initial cell states and responses to intervention, supporting the exploration of candidate ingredient screening across different cellular contexts.
Their CauTrigger study, published in Nature Methods, goes a step further by working backward from a target state to identify potential drivers, then using virtual perturbations to predict the responses that changing those factors might produce. For skin research, this offers an approach to starting with a desired improvement in cell state and working backward to identify potential targets and intervention strategies.
The team’s DNB methods also address the timing of intervention. By examining coordinated fluctuations in molecular networks, they identify early warning signals before a biological system undergoes a major state transition. Combining such methods with skin perturbation models could help researchers explore when to intervene and why the same ingredient may produce different responses at different stages. These studies provide methodological support for DermCosmos to further explore which ingredients to use, which biological processes to influence and when to intervene.
Benchmarking Against Representative International Models
According to benchmark results supplied by the research team, DermCosmos was compared with Geneformer, scGPT, UCE and State using the same skin test dataset and gene panel. These comparators represent several major approaches in single-cell AI research.
Geneformer learns relationships within gene networks through large-scale pretraining. scGPT applies generative pretraining to tasks such as cell annotation, data integration and perturbation prediction. UCE focuses on general-purpose cell representations across tissues and species. State, developed by Arc Institute and collaborators, directly predicts gene expression responses to genetic or chemical interventions. The associated studies have appeared in Nature, Nature Methods and Cell.
On both measures, DermCosmos ranked first among the models evaluated for agreement between predictions and experimental observations. It achieved a PCC of 0.9828, reflecting overall gene expression patterns, and a ΔPCC of 0.5754, reflecting changes caused by the intervention.
These results demonstrate the advantages of skin-specific modeling in supporting ingredient research.
Making Ingredient Screening and Experimental Design More Targeted
For cosmetics R&D teams, DermCosmos is intended to support the analysis of active ingredient mechanisms, prioritization of candidate ingredients and combination interventions, experimental design, and exploration of biological signals relevant to efficacy and safety.
When evaluating multiple candidate actives, for example, researchers hope to use the model to compare the cellular responses each might induce and identify the options most deserving of experimental resources. For combination interventions, the model could help teams develop more specific mechanistic hypotheses, which can then be tested experimentally.
The value of computational prediction therefore needs to be reflected in practical R&D decisions: whether it helps teams narrow the search, choose more informative experiments and better understand the results they observe.
“By bringing together skin science, predictive models and experimental validation, we hope to gain a deeper understanding of how ingredients affect skin and translate those insights into better products and consumer experiences,” said Erica Liang, R&D Strategy and Operations Lead at Unilever.
As part of the collaboration, Unilever will contribute expertise in skin and scalp biology, active ingredients and formulation science, together with experimental research and validation capabilities using cell-based and three-dimensional skin models. These contributions will support iterative model development and its application to real R&D challenges.
Supporting Regulatory Science Through Virtual Skin Research
A distinctive feature of the initiative is that it brings industrial applications, computational modeling, testing and evaluation into one collaborative framework, connecting industry R&D with regulatory science research.
Shanghai Jiao Tong University’s School of Mathematical Sciences and mathAI will jointly develop the model. Unilever will contribute industry research questions, scientific knowledge and experimental validation. The Guangzhou Institute of Quality Inspection will draw on its expertise in testing, evaluation and the development and revision of standards to explore how model predictions can connect with safety and efficacy evaluation.
From a regulatory science perspective, this arrangement extends the research to a practical question: how can AI-generated predictions become scientific evidence that can be understood, independently checked and used to study evaluation methods?
“Our focus is on how model predictions and experimental results can corroborate one another, and what scientific conclusions those results can support,” said Guoshan He, president of the Guangzhou Institute of Quality Inspection.
Under the collaboration plan, the team will work on validating predictions, connecting experimental and clinical data, and studying evaluation methods. It will also examine data quality, the interpretability of model outputs, reporting practices and the boundaries of model applicability. The aim is to build verifiable, reproducible evidence to support methodological research into AI-assisted cosmetics safety and efficacy evaluation, as well as the future study and refinement of relevant standards.
From Digital Predictions to Demonstrable R&D Value
Virtual cell models require ongoing validation. Their practical value will depend on whether they remain reliable across different skin cell types, biological contexts and intervention conditions, and whether their predictions improve experimental decisions.
The research approach proposed for DermCosmos establishes a continuous cycle between prediction and experiment. The model helps generate hypotheses and set validation priorities; experimental results test those predictions, identify the boundaries of applicability and feed back into model improvement.
For the cosmetics industry, this approach offers a direction worth watching: making fuller use of existing biological knowledge and data before experiments, then building a more systematic understanding of the connections between ingredients, cellular responses and product efficacy once results are available.
As the collaboration progresses, a key question for DermCosmos will be whether it can translate this potential into reproducible experimental results and more effective R&D decisions.