TCRForge/Operating Model
OPERATING MODEL / V0.1TF-P01

Promoter & enhancer design loop

How mu and jacky turn promoter and enhancer evidence into combinations, predictions, wet-lab results and a better model

Open current batch
Living documentv0.1 · 2026-09-06

Update after every admitted enhancer, changed assay, completed wet-lab batch or model revision

Inspect source evidence
North-star objective

Specificity first · Strength alone is not success

A candidate advances only when the underlying ON and OFF measurements remain visible

ONUseful expression in target T cells

Report positive fraction, MFI, state response and robustness

OFFLow leakage in every counter-screen

Normal and malignant B/plasma states are separate hard gates

ΔCalibrated uncertainty

The model must say when an experiment is needed to learn—not pretend certainty

Phase 1 operating loop

Evidence → combination → prediction → experiment → learning

Repeat until ranking becomes useful
  1. 1Evidence

    Papers, patents and scientific judgment

    jacky + muStructured evidence
  2. 2Enhancers

    Sequence, provenance, rights and intended cell state

    muAdmitted enhancer set
  3. 3Combine

    Enhancer × fixed core promoter with explicit controls

    jacky + AICandidate library
  4. 4Predict

    Activity, leakage and calibrated uncertainty

    jacky + AIScored candidates
  5. 5Select

    Balance best predictions, information gain, diversity and controls

    jacky + muSigned wet-lab batch
  6. 6Test

    Run the matched assay without dropping failures

    muRaw measurements
  7. 7Compare

    Compare predictions with values, ranks and failure modes

    jacky + muError analysis
  8. 8Learn

    Calibrate the model and choose the next informative batch

    jacky + AINew model revision
Feedback returns to the same evidence and model record; failed constructs are retained
Who does what

mu leads science and wet lab, jacky leads data, AI and platform

Science & wet lab

mu

  • Nominate promoter and enhancer candidates with biological rationale
  • Provide papers, patents, exact sequences and source provenance
  • Explain which motifs, cell states and mechanisms may drive specificity
  • Define target T-cell ON states and normal or malignant OFF states
  • Choose biological controls, assay conditions and pass thresholds
  • Review synthesis, vector and wet-lab feasibility before batch lock
  • Run matched experiments across replicates, donors and time points
  • Return raw measurements, QC, failures and biological interpretation
Data, AI & platform

jacky

  • Archive and structure papers, patents, sequences and usage rights
  • Create a versioned promoter and enhancer parts library
  • Generate combinations with explicit baselines and motif controls
  • Run model features and predict ON activity, OFF leakage and specificity
  • Report uncertainty and flag candidates that require experiments
  • Select a batch balancing top scores, information gain and diversity
  • Issue the construct and assay handoff with complete provenance
  • Compare predictions with raw results, update the model and choose the next batch
Decide togetherLock the objective, approve each wet-lab batch, review prediction disagreements and decide when Phase 2 can begin
Batch selection policy

Do not send only the top predictions

50%Best predicted
20%Uncertain
15%Diverse
15%Controls

This prevents selection bias and gives each wet-lab cycle enough information to correct the model

Wet-lab return contract

A ranking is useful; raw measurements are learnable

  • Every replicate, donor, batch and time point
  • Independent ON and each OFF-state result
  • Vector copy, transduction, viability and failures
  • Experimental ranking plus scientific interpretation
Learning roadmap

Earn the right to generate de novo

Evidence-grounded combinatorial optimization

Use supervised ranking, pairwise learning-to-rank, Bayesian optimization, active learning and uncertainty calibration · Reinforcement learning is not required at small batch sizes

Output: increasingly useful rankings in a controlled enhancer × promoter space
Current next actionAdmit the first enhancer set and lock the prediction + assay contract
Build combination batchReview assay matrix