An AI pipeline that converts a farmer's pest report into ranked, safety-checked dsRNA candidates — in minutes, not months. Powered by local Llama 3.2 3B, PyTorch, and a 14-species safety panel.
Chemical pesticides are failing on three fronts simultaneously
600+ arthropod species have documented resistance. Brown planthopper alone destroys 30% of Asian rice yields in outbreak years. Farmers apply higher doses that accelerate resistance selection.
Neonicotinoids linked to pollinator collapse. EU banned outdoor use in 2018. Soil and water contamination persists for years after application. Regulators are tightening globally.
RNAi biopesticides are the new chemistry (EPA registered Ledprona in 2023), but dsRNA design takes 3-6 months per target gene. No integrated safety, fate, or regulatory layer exists.
Compress the dsRNA design loop from months to minutes
Select a pest target. The pipeline tiles pest transcripts into 200-nt dsRNA precursors, dices them into 21-nt siRNAs, scores each for efficacy, checks off-target risk against 14 non-target species, predicts environmental half-life, and generates safety cards + regulatory memos.
6-stage pipeline from pest selection to regulatory memo
Click a pest target card (7 species supported) or describe the problem in natural language.
PyTorch backend tiles pest transcripts into 200-nt dsRNA precursors, then dices into 21-nt siRNAs.
Dilated CNN (HyenaDNA-inspired) scores each siRNA. Caduceus-Ph-1 adapter available for SOTA accuracy.
K-mer index checks off-target risk against 14 non-target species (pollinators, livestock, aquatic, human).
Physics-Informed Neural Network predicts environmental half-life based on sequence and field conditions.
1000-trial Monte Carlo wet-lab simulation. Llama 3.2 3B generates safety cards + EPA-style regulatory memo.
PyTorch + ROCm-ready + local Ollama LLM
7 pest targets · 14-species safety panel covering every ecological role regulators evaluate
Brown planthopper · 180 siRNAs scored · results in 0.08 seconds
1000-trial Monte Carlo · 6-stage cellular knockdown pipeline · literature-informed kinetics
Lipofection efficiency for in vitro screens. Longer PINN half-life improves delivery stability.
Cellular uptake efficiency. GC-content dependent (optimal at 45% GC).
Dicer processes dsRNA into 21-nt siRNAs. Reynolds score dependent, repeats penalized.
Thermodynamic asymmetry determines guide vs passenger strand. Reynolds rules 3-7.
mRNA cleavage rate. CNN-predicted efficacy is the primary driver.
Soft-saturation dose-response. Reports mean KD, 95% CI, P(KD>70%).
Real measured data + AMD GPU training · architecture is production-ready, only training data is synthetic
| Component | Demo (synthetic) | Production (real data + AMD GPU) |
|---|---|---|
| Training set | 1,000 synthetic siRNAs | 30,000+ measured siRNAs (siRecords) |
| Safety panel | 14 synthetic transcriptomes | 14 real NCBI transcriptomes (~2 GB) |
| Model | Dilated CNN, 2 min on CPU | Caduceus fine-tune, 15 min on MI250 |
| Validation AUC | 1.00 (trivially separable) | 0.82-0.92 (realistic) |
| Off-target accuracy | 0% (random sequences) | Real per-species risk scores |
| Training cost | $0 (local CPU) | ~$1.50 (1 hour MI250) |
Codebase already supports production via three CLI flags: --source real · --device cuda · --use-caduceus
$84B market · EPA regulatory precedent set · no competitors with integrated AI pipeline
Only platform combining SOTA sequence models, PINN fate prediction, 14-species safety panel, and LLM-generated regulatory docs. 100-1000x faster than competitors.
From hackathon MVP to Series A in 12 months
The hackathon is the proving ground. The market is $84 billion. The regulatory pathway is open. The technology works. We are ready to build what's next on AMD.