Preserve useful CAR expression across defined activation and manufacturing states.
Measure: % positive · MFI · normalized activityREGULATORY / DESIGNTF-P01
Synthetic regulatory design
Build an auditable candidate batch from grounded elements, then optimize T-cell activity and off-target leakage as separate objectives.
Treat malignant-cell leakage as an independent gate, not an optional annotation.
Measure: leakage · contrast · detection floorAn average score cannot hide a failed leakage threshold. Rank candidates only after hard gates remain visible.
Pareto rank · reproducibility · context coverageLighthouse workflow
A controlled design–build–test–learn cycle
The first cycle stays interpretable. Unconstrained de novo generation is admitted only after this baseline works.
Curated T-cell regulatory sequences, source evidence and rights
Candidate cores, TF motifs, sequence grammar and uncertainty
Fixed core promoter plus named enhancer modules and controls
Continuous activity and leakage in pre-registered cell states
Use every measurement and failure to choose the next batch
Candidate batch
Start with modules we can name, edit and test
REG-001TRAC / TCRα coreTCRα module ASingle-module activity baselineReferencegrounded
REG-002TRAC / TCRα coreCD3 module BIndependent T-cell regulatory baselineReferencegrounded
REG-003TRAC / TCRα coreA + BTest non-additive module grammarCombinationproposed
REG-004TRAC / TCRα coreA(mut) + BAttribute the nominated motifControlproposed
The learning signal is a contextual measurement, not a winner label
01On-target activity% positive · MFI · normalized reporter / CAR
02Off-target leakageNormal panel + malignant plasma/B-cell panel
03Specificity contrastON/OFF ratio with raw values retained
04Biological contextDonor · activation · disease state · time point
05Technical contextVector · copy number · integration · replicate
What each model is allowed to do
Human sequence-to-activity oracle + iterative proposal search
Research-readyCell-line MPRA; not a T-cell product oracleCNN activity prior + synthetic sequence design
BaselinePrimarily Drosophila enhancer assaysCell-type-conditioned enhancer design + motif grammar
Research pilotCross-context transfer requires calibrationConditional de novo sequence proposals
WatchPreprint; predominantly in-silico evidenceActivity–leakage prediction in our assay schema
MissingTrain only after matched continuous data existA motif hit proposes a causal experiment; it does not prove enhancer activity.
The disease-state counter-screen is a separate gate.
Model scores stay predictions until the exact construct and context are assayed.
Portable regulatory elements and endogenous locus control remain separate architecture arms.