primeforge

Prime editing design SDK: modern C++ core + thin Python API, with deterministic pegRNA / nicking-guide enumeration for PE2/PE3 and batch-friendly data models. v0.1 is CPU-first and pure logic; CUDA backends are available for high-throughput PAM scanning and will later host thermodynamics/ML scoring.

Status

• Core C++ library with explicit edit model (subs/ins/del), strand-aware design, PBS/RTT heuristics, and optional PE3 companion guides.
• Python bindings and dataclasses mirror the C++ API.
• CUDA PAM scanner available (opt-in); benchmarks included. Scoring/thermo GPU kernels come next.

Build (C++)

cmake -S . -B build -DPRIMEFORGE_ENABLE_CUDA=OFF
cmake --build build
ctest --test-dir build  # runs C++ (and Python if bindings + pytest present)

Python (dev)

pip install -r python/requirements-dev.txt
# build/install once bindings are implemented
# from repo root:
# cmake -S . -B build -DPRIMEFORGE_BUILD_PYTHON=ON -DPRIMEFORGE_ENABLE_CUDA=OFF
# cmake --build build
# pytest python/tests  (PYTHONPATH=build:$PYTHONPATH)

Benchmarks (optional)

PAM scan throughput (CPU by default, set DEVICE=cuda when built with CUDA):

cmake -S . -B build -DPRIMEFORGE_BUILD_BENCHMARKS=ON -DPRIMEFORGE_ENABLE_CUDA=OFF
cmake --build build --target bench_pam
./build/primeforge-core/benchmarks/bench_pam 5000000 NGG 3  # args: length motif iterations (first is warm-up)
# With CUDA (example for GTX 1060, compute 6.1):
# CUDACXX=/usr/bin/nvcc cmake -S . -B build -DPRIMEFORGE_ENABLE_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=61 -DPRIMEFORGE_BUILD_BENCHMARKS=ON
# cmake --build build
# DEVICE=cuda ./build/primeforge-core/benchmarks/bench_pam 5000000 NGG 3
# Recent run (GTX 1060, NGG, 5 Mb): CPU ~12 Mb/s, CUDA ~19 Mb/s (post warm-up)

To gate benchmarks in CI: add -DPRIMEFORGE_RUN_BENCH=ON (requires CUDA build) and ctest will run a short CUDA PAM check.

Features (v0.1)

• Typed edit specs (substitution/insertion/deletion) with strand awareness.
• pegRNA assembly: spacer, PAM cut logic, PBS/RTT enumeration, GC heuristics, distance flags.
• Optional PE3 companion nicking guide selection with configurable PAM motifs.
• Batch APIs for large edit sets; zero I/O in the core.
• CUDA hook points for PAM scanning (present) and future scoring/thermo kernels.

Roadmap

• GPU-optimized thermodynamics + secondary-structure heuristics.
• Plug-in scoring interface (EasyPrime/OPED/etc.).
• Library-scale workflows and genome-provider abstractions.

Design goals

• Explicit edit model (subs/ins/del) instead of magic strings.
• pegRNA anatomy first-class: spacer, PBS, RTT, nicking guides.
• Deterministic rule-based design; ML hooks added without entangling the core.
• Library-scale batch APIs; zero I/O in the C++ core.

CUDA roadmap (high level)

• GPU PAM scanning for genome-scale sweeps.
• Batched thermo/secondary-structure heuristics and ML scoring.
• Variant-aware masking for personalized genomes.

See docs/design_rules.md for the rule set and docs/api.md for the API sketch.