VirtualSkin integrates skin-cell identity with intervention-response prediction

Technical Summary: VirtualSkin

Problem Statement
Human skin atlases have successfully resolved diverse cellular states across developmental stages, anatomical sites, and disease contexts. However, a critical gap remains between describing a cell's current state and predicting how that state will change in response to specific interventions (e.g., drugs, genetic perturbations, or physical treatments). Existing single-cell foundation models often struggle to maintain cell identity across independently sampled tissues while simultaneously supporting intervention-conditioned response prediction. Furthermore, public skin datasets present challenges regarding heterogeneity in cell capture, annotation resolution, and the distinction between biological replicates and repeated measurements from the same donor. There is a need for a unified framework that connects cutaneous cell identity, biological language, and intervention response without relying on experimental shortcuts or failing to generalize across study boundaries.

Methodology
The authors present VirtualSkin, a cell–language model designed to align skin transcriptomes with biological descriptions and predict intervention responses.

• Architecture: VirtualSkin employs a dual-branch transformer architecture. The molecular branch encodes gene identities and expression values jointly, while the text branch processes biological descriptions. Both branches utilize norm-first transformer layers with separate parameters, projecting cell and text representations into a shared, length-normalized space.
• Training Objectives: The model is trained using a composite objective combining:
  1. Molecular Reconstruction: Masked expression reconstruction to retain measured molecular variation.
  2. Ontology Classification: Connecting representations to declared cell identities.
  3. Symmetric Cell–Text Alignment: Linking cells to their paired descriptions.
• Robustness Mechanisms: To handle variability in paired descriptions, the model incorporates pair reliability weighting (based on provenance and quality), an ontology-aware mask (to prevent biologically related labels from acting as in-batch negatives), and a FIFO queue to extend the negative set beyond the current batch. Retrieval augmentation supplies contextual descriptions from a fixed, training-eligible document index.
• Intervention Prediction: For response prediction, the model conditions on the pre-exposure cell state, intervention identity, and measured dose or elapsed time. The output predicts the exposure-associated change rather than a static cell label.
• Evaluation Framework: The authors constructed a rigorous, cohort-resolved evaluation framework spanning healthy epidermis, ageing, inflammatory diseases, fibrosis, lymphoma, and carcinoma. Crucially, the evaluation design withholds specific donors, studies, perturbations, or intervention targets during training to test generalization. Comparisons are made against strong baselines (e.g., Geneformer, scGPT, LangCell, GEARS, CPA) using matched inputs and partitions.

Key Contributions and Results

1. Preservation of Cell Identity and Study Integration:

   • VirtualSkin demonstrated superior ability to preserve distinct skin cell classes (basal/spinous keratinocytes, fibroblasts, T cells) while reducing study-associated batch effects compared to Geneformer, scGPT, and LangCell.
   • In inflammatory skin annotation, VirtualSkin improved macro-F1 by 0.056 points over the strongest comparator.
   • In few-shot settings (1–10 labeled cells per class) for cutaneous T-cell lymphoma, it achieved higher macro-F1 scores, with a normalized area under the support-size curve of 0.598 vs. 0.501 for the best baseline.
   • In zero-shot transfer to vitiligo (fine labels withheld from training), it achieved a hierarchical macro-F1 of 0.549, outperforming LangCell (0.392) and scMulan (0.423).

2. Cell–Language Alignment and Robustness:

   • VirtualSkin achieved a Recall@10 of 0.73 for cell-to-text retrieval in held-out inflammatory skin, a 0.12 point improvement over scMulan.
   • The model showed resilience to incomplete molecular inputs, retaining 88% of its unmasked score at 50% gene masking, compared to 78% for the strongest comparator.
   • It maintained 82% of baseline performance even when 50% of cell–description pairings were disrupted, indicating robustness to imperfect pairing.

3. External Transfer:

   • The model successfully transferred across disease states (systemic sclerosis, lymphoma), anatomical sites (palm, sole, hip), and ageing contexts, consistently outperforming comparators in donor-level AUROC and site-specific macro-F1 metrics.
   • Spatial localization in basal-cell carcinoma assays correctly assigned stromal, tumor-core, interface, and epithelial states to distinct tissue compartments.

4. Intervention-Response Prediction:

   • Across five held-out intervention cohorts (genetic screens, drug exposure, photodynamic treatment, mechanical expansion), VirtualSkin improved response prediction scores by 0.03–0.09 points over the best baselines.
   • It achieved a mean average precision of 0.559 for perturbation ranking (vs. 0.489 for the best baseline) and a sign accuracy of 0.794.
   • Molecular fidelity was confirmed through high Pearson correlation (0.877) between predicted and measured pseudobulk log fold-changes and strong recovery of top differentially expressed genes.
   • Paired donor trajectories in tissue interventions showed that the model correctly reproduced the direction of immune remodeling and regeneration scores over time.

5. Reliability and Uncertainty:

   • VirtualSkin's predictive intervals showed better calibration (mean absolute coverage error of 0.017 vs. 0.045 for comparators).
   • Uncertainty estimates were highly correlated with absolute error (Spearman correlation of 0.912), enabling selective prediction: restricting predictions to the least uncertain units increased the response score from 0.59 to 0.78 at 55% coverage.

Significance and Claims
The paper claims that VirtualSkin establishes a connection between skin-cell recognition and intervention-response prediction within a common evaluation framework. Its primary contribution is not a single score but a unified approach that validates representation learning, annotation, and response prediction under strict, independent evaluation conditions.

The authors emphasize that the model's ability to recover changes absent from the baseline state is distinct from its ability to annotate existing cells. They assert that the results define the conditions under which predictions can inform experimental prioritization, particularly through the use of uncertainty estimates to withhold predictions under distribution shift. The paper modestly notes that while the model supports hypothesis generation and prioritization, it does not establish causal pathway mediation or clinical efficacy, and its scope is currently limited by the availability of non-malignant skin intervention data and the specific anatomical sites covered. The work provides an integrated route from cutaneous molecular identity to testable response hypotheses, supported by matched controls and uncertainty analyses.