CLOP-DiT: Structured-metadata-conditioned single-cell latent generation via contrastive language-omics pretraining and Diffusion Transformers

Single-cell RNA sequencing has transformed cellular heterogeneity analysis, but most single-cell generative models are conditioned on narrow categorical labels such as cell type, batch, or perturbation rather than structured biological descriptions. This limits controlled simulation and hypothesis-guided exploration, especially when biologically meaningful within-type heterogeneity must be preserved. We present CLOP-DiT, a three-stage framework for structured-metadata-conditioned single-cell latent generation. CLOP first aligns frozen BiomedBERT text embeddings and frozen scGPT cell embeddings in a shared 512-dimensional space. A conditional 1-D Diffusion Transformer then performs flow matching in this latent space under a fixed five-field prompt template covering cell type, tissue, organism, marker genes, and disease context. Finally, a frozen scGPT decoder maps generated latents back to gene expression for inspection. Across 69 deduplicated cell types from 80 GEO datasets (220,304 cells), a high-fidelity regime (CFG = 2.0) achieves 36.9% KNN accuracy, approximately 25 × random chance, and 81.0% steering. A high-diversity regime (CFG = 1.0) reaches a diversity ratio of 0.93 while retaining 80.7% steering. Conditioning-field ablation and swap-label tests identify marker genes as the strongest steering signal, with steering falling from 99.8% under the full prompt to 62.4% with metadata alone. We also report current limits: cross-dataset variance correlation remains near zero, discriminator AUC is 0.656, and rare-cell augmentation remains negative. CLOP-DiT therefore supports controllable structured-metadata generation while highlighting the remaining challenge of preserving within-type heterogeneity and downstream biological realism.