Bloomsbury Burger
Therapeutics
We are building a dual-target CAR-T therapy guided by single-cell genomics. This is not hormone modulation. Not symptom management. This is targeted cellular ablation of ectopic lesions — informed by a 54-patient transcriptomic atlas and a differentiable optimisation formulation we believe is novel.
Why endometriosis, why now
Endometriosis affects ~10% of people with a uterus worldwide. The current standard of care is surgical resection — cutting out the lesions — which has a high recurrence rate and is imprecise. Surgeons literally cannot reliably distinguish all pathological tissue from healthy tissue by visual inspection alone.
The field is chronically underfunded. A 1977 FDA mandate excluded women from early clinical trials and the field has been playing catch-up ever since. Femtech is currently 80% cycle trackers and 20% rebranded thermometers.
We think a targeted cellular therapy, informed by single-cell genomics, could be a genuinely better solution. CAR-T has already transformed blood cancer treatment. We're asking: what if we applied the same logic to solid tissue pathology?
Technical pipeline
Load & QC single-cell data
Load raw Cell Ranger output (raw_feature_bc_matrix.h5) for each sample. Filter out empty droplets and dying cells using gene count and mitochondrial read thresholds. Save filtered cells as .h5ad files.
Denoise with scVI-VAE
Single-cell data is noisy — genes drop out randomly. We run scVI (a variational autoencoder) to learn a clean latent representation of each cell's true expression profile. This enables stable binarisation for downstream combinatorial optimisation.
Differentiable marker optimisation
Instead of brute-force enumeration of ~4.5 million possible marker pairs, we relax the discrete combinatorial search into continuous space using a Gumbel-Softmax reparameterisation, optimise with gradient descent, then snap back to a hard discrete selection. We believe this specific formulation — combining a biomedical specificity objective with a whole-body safety penalty in a jointly differentiable system — is novel. → See full mathematics
Cross-reference Tabula Sapiens
Any surviving marker pairs get checked against the full human cell atlas. Infinite penalty for expression in heart, lung, brain or other critical organs. Only pairs that are truly lesion-specific survive this filter.
Output top candidate pairs
Gradient descent outputs a ranked top-5 list of dual marker combinations. These get handed off for benchtop feasibility assessment — checking whether scFvs (the targeting domains) exist or can be designed for each candidate.
Dataset — GSE213216
We're using the Human Endometriosis Cell Atlas published in Nature Genetics (2024). It contains single-cell RNA sequencing data from 54 patient samples across multiple tissue types: ectopic endometrial lesions (endometriomas + peritoneal lesions), eutopic endometrium, and unaffected control tissue.
Total size: ~15.7GB. Format: Cell Ranger output (10x Genomics). Each sample loads as a barcodes × genes matrix. Integration across samples is performed after QC and batch-aware modelling with scVI.