🤓 Latest Submissions
Biopesticide-AI
Chemical pesticides are failing — 600+ arthropod species have developed resistance, pollinators are collapsing, and regulators are pulling active ingredients faster than replacements arrive. RNAi biopesticides are species-specific and degrade in days rather than decades (the EPA registered the first one, Ledprona, in 2023), but designing dsRNA candidates is still a manual, expert-gated process that takes months per target gene. Biopesticide-AI compresses that design loop into an automated pipeline. It takes a natural-language pest and crop description, tiles the pest transcriptome into 200-nt dsRNA precursors, dices them into 21-nt siRNAs, and scores efficacy with a CNN model (with an optional Caduceus/Mamba-based variant). Candidates are then screened against 14 non-target species — pollinators, livestock, aquatic life, and humans — using mismatch-tolerant k-mer matching, and a physics-informed neural network estimates environmental soil half-life. A local LLM generates safety cards and EPA-style regulatory memo drafts for the top candidates. The entire pipeline runs end-to-end in under 4 minutes on a CPU, fully containerized with Docker, and requires no bioinformatics expertise to operate. The current release validates the architecture using synthetic and rule-derived training data. The codebase is built to swap in real assay data (siRecords, DRSC/TRiP, GenomeRNAi for efficacy; PPDB for environmental fate; NCBI for transcriptomes) via a single flag, with no architectural changes required — the next milestone is training on real measured knockdowns to move from a validated prototype to a field-ready design tool.
13 Jul 2026
