PINNAL
Genomic-Foundation-Model RNA Therapeutic Candidate Ranker · Murai Labs
What It Is
A modular pipeline that ranks RNA therapeutic interventions — antisense oligos, splice-switching oligos, base edits — using an ensemble of genomic foundation models (Evo 2, a 4-replicate Borzoi ensemble, Enformer, ViennaRNA). You bring a gene, an exon, and a tissue; PINNAL tells you which of the ~200–5,000 possible binding sites to wet-test first, and shows its work. It is a candidate ranker — not a target-discovery tool — that closes the loop between computational prediction and wet-lab experimentation.
ASO development is fragmented across variant interpretation, off-target screening, structural analysis, and protocol prep — and most of that cost lands in the wet lab. PINNAL consolidates those steps into a single ranked shortlist with an audit-traceable decision trail, moving expensive experimental work upstream to computation. Every prioritization shows why it was made, so a reviewer can interrogate the ranking instead of trusting it.
What I Built
- Nine-module pipeline (INGEST → PREDICT → MAP BUILD → STRUCT → ENUM → SIM → RANK → OUTPUT) with an experimental FEEDBACK loop — each module usable standalone — plus PINNAL-θ, the proprietary trained ranking head
- Commercial-licensing-aware ensemble — Enformer + 4-replicate Borzoi + Evo 2 + ViennaRNA — with AlphaGenome (CC-BY-NC) hard-excluded from the commercial build via a PINNAL_COMMERCIAL_MODE enforcement path
- Pre-registered falsification test on SMN2 / ISS-N1 against a real NVIDIA Evo2-40B NIM — genuine model prediction, not a curated lookup
- Opentrons protocol generation from ranked output, plus an experimental-result ingest path that re-ranks candidates from wet-lab feedback
- CLI-first workflow (pinnal rank / protocol / ingest / bench) running locally on Apple Silicon (MPS) or CUDA — chemistry-agnostic by design (PINNAL outputs sequences; the synthesizing lab picks the chemistry)
- Provenance discipline throughout — every result carries config, inputs, and a verification command; validation records committed under docs/validation/
Tech Stack
Pre-alpha, demo-ready. In the SMN2 / ISS-N1 falsification test, PINNAL's commercial ensemble surfaced all 3 tiles of the clinically-validated ISS-N1 site in its top 10 while Evo 2 used as a ranker surfaced none (enrichment@10 4.10 vs 0.00; bootstrap 95% CI [2.23, 5.55]). The honest claim is consistency across splicing mechanisms — PINNAL is the robust generalist, not the per-mechanism peak. Apache 2.0 core; PINNAL-θ weights proprietary.