# Lucid

- **Event:** [Zero to Agent: Vercel x Deepmind Hackathon NYC](https://cerebralvalley.ai/e/zero-to-agent-nyc)
- **When:** Sat, Mar 21 at 9:00 AM – 10:00 PM (EDT)
- **Where:** 220 W 26th St, New York, NY
- **Team:** [Evan Correa](https://cerebralvalley.ai/u/evancorrea), [Aarjav Jain](https://cerebralvalley.ai/u/aarjavjain)
- **GitHub:** https://github.com/Aarjav0210/lucid/tree/main
- **Demo video:** https://youtu.be/6mL2URpXO5w
- **Gallery:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery/29

Commercial protein synthesis has become significantly more accessible and affordable in recent years, expanding the number of providers and lowering the cost of custom protein orders. This accessibility has prompted regulatory attention: a 2024 executive order requires federally funded researchers to purchase synthetic nucleic acids and proteins only from providers adhering to established screening frameworks, and additional legislation is currently moving through the Senate to broaden these requirements.
Current screening infrastructure predominantly relies on sequence similarity search (BLAST) against databases of known pathogens and select agents. This approach has documented limitations. Sequences that have been redesigned to preserve function while reducing similarity to known threats can fall below detection thresholds. Split-order attacks, in which individually innocuous fragments are ordered separately and assembled post-delivery, are not addressed by per-order screening in isolation.
Lucid is a screening tool that applies multiple analysis layers to incoming synthesis orders. Given a protein sequence, it performs sequence similarity search against pathogen and toxin databases, functional domain annotation via InterProScan to identify threat-associated motifs independent of overall sequence identity, and structural homology search via Foldseek to detect conserved folds in sequences that have diverged beyond recognition by the other two methods. Each layer's output informs whether the next layer is invoked, following an escalation model where computationally cheaper checks run first and more intensive analysis is triggered only when earlier results are ambiguous.
In addition to whole-sequence analysis, Lucid decomposes each order into its constituent protein domains, screens them individually, and evaluates the combination for synergistic risk. This addresses the split-order problem by assessing whether components that pass screening in isolation could constitute a threat when assembled.
The output is a structured risk assessment that includes per-layer findings, an overall risk classification, reasoning traces, and recommended actions. Lucid is deployed as a single API endpoint, accepting a sequence and returning a report, intended to integrate into existing order intake systems at synthesis providers with minimal implementation overhead.

---

Markdown version of https://cerebralvalley.ai/e/zero-to-agent-nyc/hackathon/gallery/29. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt