SynthGuard: Biosecurity Screening for AI-Designed Sequences
AIxBio Hackathon 2026 — Track 1: DNA Screening & Synthesis Controls
BLAST-based screening misses 99.6% of AI-designed variants (ProteinMPNN/RFdiffusion). SynthGuard detects functional hazard — not sequence similarity.
| Track | Model | Key result |
|---|---|---|
| DNA (codon-optimized) | k-mer LightGBM | 90.7% recall vs 0.4% BLAST — 227× improvement |
| Protein (ProteinMPNN redesigns) | k-mer + ESM-2 v4 | 100% recall, AUROC 1.000 on 255 variants (5 toxin families) |
0.1 0.9
0.1 0.99
0.3 0.95
Examples
DNA model: 1,364-dim k-mer (k=3–6) + codon-usage features → LightGBM. Protein model v4: 906 features (426 k-mer/dipeptide + 480 ESM-2 mean-pooled) → LightGBM. ESM-2 captures protein fold rather than composition alone — enabling generalization to unseen toxin families.
Seyomi/synthguard-kmer | Seyomi/synthscreen-dataset | GitHub