🧬 CRISPR Precision Studio

Design, rank, explain, and validate CRISPR guides — in one lightweight platform.

Transparent guide prioritization: interpretable on-target scoring, both-strand off-target analysis, and prime-editing support — without GPUs, cloud dependencies, or black-box predictions.

_{CI runs the test suite and renders a live UI screenshot (downloadable as a build artifact) on every push.}

---

📸 The interface

_{Rendered automatically by CI on every push — one Score per guide, with a per-feature Details breakdown.}

---

✨ Highlights

	Feature	What it means
🎯	One Score	A single 0–100 prioritization number ranks each guide — no column soup.
🔍	Explainable	POST /api/explain shows the per-feature breakdown behind every score.
🧬	Both-strand off-targets	Vectorised NumPy scan + per-site CFD & MIT/Hsu + aggregate specificity.
🌟	Prime Editing Studio	PRIDICT2.0-informed pegRNA design (Spacer + RTT + PBS).
📚	Peer-reviewed scoring	CRISPRscan weights reproduced verbatim & unit-validated — zero downloads.
🔌	Pluggable models	onnx → trained-linear → heuristic, auto-selected and reported.
⚡	Lightweight	No GPU, no LLM keys, no DB — everything computed per request.

---

🚀 Quickstart

cd crispr_app
pip install -r requirements.txt
uvicorn main:app --reload

➡️ Open http://127.0.0.1:8000

---

🔄 Workflow

   DNA sequence (paste or FASTA)
            │
            ▼
   Guide discovery  ── both strands, multi-PAM (NGG/NAG/NG/TTTV)
            │
            ▼
   On-target scoring ── built-in model + CRISPRscan
            │
            ▼
   Off-target analysis ── CFD + MIT/Hsu + aggregate specificity
            │
            ▼
   Ranking ── one 0–100 Score per guide
            │
            ▼
   Explanation ── per-feature breakdown (/api/explain)

---

⚡ Example

curl -s -X POST http://127.0.0.1:8000/api/design \
  -H 'Content-Type: application/json' \
  -d '{"dna_sequence": "ATGGCCGAGTACAAGCCCACGGTGCGCCTCGCC...", "pam": "NGG"}'

Real output (288 bp input → 21 guides found; model: linear, the shipped default):

#	Guide (5′→3′)	PAM	Strand	GC%	Score
1	GATGTGGCGGTCCGGATCGA	CGG	−	65	74
2	AAGGTGTGGGTCGCGGACGA	CGG	+	65	74
3	ATCGACGGTGTGGCGCGTGG	CGG	−	70	69

POST /api/explain then returns the per-feature breakdown (GC, T_m, position-specific contributions) behind any guide's Score.

---

🎯 The score (production)

Each guide gets one Score, 0–100 (higher = better) — a relative prioritization score combining the on-target predictors, not a literal % editing rate. Color-coded:

🟢 High	🟡 Moderate	🔴 Low
≥ 60	40 – 59	< 40

Click Details on any guide to see why it scored that way (GC, Tm, position-specific features…). Component sub-scores stay in the API/CSV for power users — never on screen.

---

📊 Accuracy — measured, not asserted

Held-out Spearman ρ on real public datasets (full table + method in BENCHMARKS.md). The platform offers two tiers — a transparent built-in ranker (default) and an optional external deep-learning backend for maximum raw accuracy:

🪶 Built-in — lightweight, interpretable, zero setup

Model	ρ	Notes
Shipped trained (default)	0.22 – 0.41	pooled human SpCas9, leave-one-dataset-out
Trained on your own data	0.40 – 0.52	one command — train.py
Heuristic (always available)	~0.25	fully interpretable fallback
CRISPRscan (peer-reviewed, validated)	0.58	on its home dataset

🧠 Optional — external deep-learning backend

Model	ρ	Notes
ONNX (DeepSpCas9 / CRISPRon)	~0.85	bring your own export; auto-detected

The built-in tier optimises for transparency and speed — its job is to rank candidates well enough to prioritise, with every score explainable. For maximum raw correlation, drop in a deep model via ONNX.

⚠️ Honesty note. No predictor can exceed the ~0.71–0.77 reproducibility ceiling of the wet-lab data itself; published state-of-the-art tops out around ~0.85–0.88. Our scores are deterministic surrogates for ranking; wet-lab validation remains essential.

Train a stronger model (NumPy-only, no heavy ML stack)

cd crispr_app
python train.py dataset.csv          # columns: guide,measured[,ngg_context]
# → writes models/linear.json; the API auto-loads it and reports model="linear"

python benchmark.py data.context.tab # measure Spearman on a CRISPOR-format set

Position-specific dinucleotide features roughly double Spearman on datasets with signal (chari2015 0.20→0.40, morenoMateos 0.17→0.43); gradient boosting matched ridge to ±0.02, so we stay dependency-free.

---

🆚 How it compares

Honest positioning — including where we're weaker. CRISPOR/CHOPCHOP are mature, genome-aware tools; our edge is transparent prioritization in a lightweight, API-first package.

Capability	CRISPR Precision Studio	CRISPOR	CHOPCHOP	Benchling
Single explainable prioritization score	✓	partial¹	partial¹	✗
Per-feature score breakdown (API)	✓	✗	✗	✗
Both-strand off-target (CFD + MIT)	✓	✓ (reference)	✓	✓
Genome-wide off-target search	✗ (background seq only)	✓	✓	✓
Prime-editing pegRNA design	✓	✗²	partial	✗
JSON API-first	✓	partial	✗	✓
Runs locally, no GPU / no keys	✓	✓³	✓³	✗ (SaaS)

_{¹ Report several separate scores rather than one explained number. ² CRISPOR targets Cas9/Cas12a guide design; pegRNA design is usually a separate tool (PrimeDesign / pegFinder). ³ Open-source but heavier to self-host. Marks reflect typical usage and may change as those tools evolve.}

Honest gap: genome-wide off-target scanning is the main capability CRISPOR/CHOPCHOP have that we don't — it's on the roadmap.

---

🏗️ Architecture

Browser (templates/index.html + static/app.js)
        │  JSON over fetch()
        ▼
FastAPI (main.py)  ──►  Pydantic validation + utils.validate_sequence
        │
        ▼
Science layer
   ├── scoring.py     on-target efficiency   (Doench RS2 / Azimuth-informed)
   ├── crisprscan.py  CRISPRscan             (Moreno-Mateos 2015, verbatim)
   ├── offtarget.py   CFD + MIT/Hsu + aggregate specificity
   ├── prime.py       pegRNA design          (PRIDICT2.0-informed)
   ├── features.py / models.py / train.py    pluggable + trainable models
   └── analysis.py    pipeline + vectorised both-strand off-target search
        │  pandas DataFrame → JSON
        ▼
Browser renders one ranked table

---

🔌 API reference

Method & route	Purpose
GET /health	liveness check
POST /api/design	ranked gRNAs with the ConsensusScore (the 0–100 Score)
POST /api/offtargets	per-site CFD/MIT hits + per-guide specificity summary
POST /api/simulate	protein / indel outcome of an edit
POST /api/prime-design	ranked pegRNAs (Spacer + RTT + PBS)
POST /api/explain	interpretable per-feature score breakdown
POST /api/upload-fasta	parse pasted FASTA / plain DNA
GET /api/models	active & available on-target backends

---

🌟 Prime editing — how pegRNAs are chosen

For a target base substitution, prime.py enumerates and ranks candidate pegRNAs using determinants from PRIDICT2.0 (Mathis 2024) and Anzalone 2019:

1. Spacer / nick. Scan NGG PAMs within ~30 nt of the target; place the Cas9 nick 3 bp 5′ of each PAM. Require the edit to fall 0–15 nt downstream of the nick.
2. PBS (primer-binding site). Enumerate lengths 8–17 nt; the PBS is the reverse complement of the sequence immediately 5′ of the nick. Its nearest-neighbour T_m is optimised toward ~37 °C (Gaussian reward), with mild length penalties favouring ~13 nt.
3. RTT (reverse-transcriptase template). Enumerate lengths 10–20 nt; the RTT encodes the edit and must retain ≥3 nt of 3′ homology past the edit for flap resolution. Penalties: RTT that begins with C (destabilises the edited flap) and RTT GC far from ~55%. Length term favours ~12 nt.
4. Ranking. A calibrated logistic score blends the PBS T_m, PBS/RTT length terms, 3′-homology constraint, RTT-starts-with-C penalty, and GC term into one 0–1 Score.

The pegRNA score is PRIDICT2.0-informed, not the trained PRIDICT2.0 network. It reproduces the published determinants for ranking; it has not yet been numerically benchmarked against a PRIDICT test set (on the roadmap). No secondary-structure (e.g. RNAfold) penalty is applied yet.

---

🔬 Scientific basis

Component	Model / source
On-target	Doench 2014/2016 Rule Set 2/Azimuth (Nat. Biotechnol. 34:184); CRISPRscan (Moreno-Mateos, Nat. Methods 2015)
Off-target (site)	CFD (Doench 2016) · MIT/Hsu (Hsu 2013, Nat. Biotechnol. 31:827)
Off-target (guide)	aggregate specificity 10000 / (100 + Σ scores) (CRISPOR convention)
Prime editing	PRIDICT2.0 (Mathis 2024, doi:10.1038/s41587-024-02268-2); Anzalone 2019 (Nature 576:149)

---

✅ Tests

pip install pytest
python -m pytest tests/ -q     # 35 passing

Covers on-target scoring, CFD/MIT scoring, aggregate specificity, both-strand off-target detection, pegRNA design, the model registry & trainer, CRISPRscan reference-vector validation, performance, and dependency hygiene.

---

MIT licensed · No API keys required · Wet-lab validation always essential