Problem: gene therapy is too expensive, not accurate enough Goal: predict AAV capside x promoter transduction in specific input cell types

We want to:

1. Predict if RNA/DNA edits reach intended cells for given vector + context
2. Convert all wet lab steps to be computational
3. Do so while being as quantitatively accurate as what we expect in the wet lab (evals per step)

This is inspired by the paper that creates a new system on vector delivery into reinal + brain cells. I have written more about this on .

python -m venv venv
source venv/bin/activate
pip install --upgrade pip
pip install -r requirements.txt

Inputs from the scAAVengr experiment (Öztürk et al., eLife 2021)

• make_inputs_from_paper.py: script to build data/capsids.fasta and data/promoters.tsv.
• data/promoters.tsv: prewritten example (strengths are placeholders; paper does not report numeric promoter strengths).

To generate capsids.fasta with AAV1/2/5/6/8/9 (UniProt) and AAV2 engineered variants (K91, K912, K916, K94), run:

python make_inputs_from_paper.py

• AAVrh10: included in the study, but I didn’t hard-code an accession to avoid guessing, you can add it if you want it in capsids.fasta (the RefSeq entries exist, or you can map to UniProt for your preferred isolate)
• 4YF/4YFTV/2YF mutants: listed in the paper, but the exact residue map depends on numbering conventions. I left them out by default to avoid silent misannotation (avoid future debugging), can be added after mutation map