Skip to content

Latest commit

 

History

4 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

EvoGuard: Multi-Layer Screening for AI-Engineered Harmful Proteins

Overview

EvoGuard is a prototype biosecurity screening system designed to identify potentially harmful proteins — including novel or AI-engineered variants that evade traditional sequence-similarity checks.

Current biological screening pipelines primarily rely on sequence comparison tools such as BLAST to flag proteins similar to known toxins or pathogen components. However, advances in AI-driven protein design make it increasingly possible to generate proteins that retain dangerous biological functions while appearing dissimilar at the sequence level. These proteins may bypass existing screening methods.

EvoGuard demonstrates a multi-layer detection approach that goes beyond sequence similarity by incorporating functional similarity and evolutionary plausibility checks.

This project is a hackathon prototype illustrating how future biosecurity systems might detect potentially harmful proteins that traditional methods miss.


Motivation

The Biosecurity Gap

Organizations involved in biotechnology governance and screening (e.g., DNA synthesis providers and international biosecurity initiatives) currently use sequence similarity checks to prevent the synthesis of known pathogen sequences.

A simplified version of current screening looks like:

protein sequence
       ↓
sequence similarity search
       ↓
flag if similar to known pathogen/toxin

While effective for known threats, this approach has a critical weakness:

AI protein design tools can generate novel sequences with dangerous biological functions that do not closely resemble known proteins.

For example:

  • mutations can obscure similarity to known toxins
  • functional motifs can be rearranged
  • new sequences may be evolutionarily implausible yet biologically effective

This creates a potential screening blind spot.


Our Approach

EvoGuard introduces a three-layer screening system that attempts to detect these blind spots.

1. Sequence Similarity Layer

The first layer performs traditional sequence comparison against a database of known harmful proteins.

Purpose:

  • detect obvious matches
  • replicate current screening approaches

However, this layer alone can be bypassed through sequence mutation or AI-generated variants.


2. Functional Similarity Layer

The second layer attempts to detect functional similarity even when sequences differ.

This is done using protein representation models that convert protein sequences into embedding vectors capturing structural and functional information.

If a novel protein has a similar embedding to known toxins, it may share functional properties despite sequence differences.


3. Evolutionary Plausibility Layer

The third layer evaluates whether a protein sequence appears evolutionarily plausible.

Natural proteins tend to evolve through:

  • incremental mutations
  • phylogenetically consistent patterns
  • constrained mutation rates

Sequences with unusually large mutation distances or improbable motif combinations may indicate:

  • artificial design
  • directed evolution
  • synthetic modification

Risk Scoring

The outputs of the three layers are combined into a risk score that classifies proteins into categories such as:

  • Safe
  • Review Recommended
  • High Risk

The system also produces human-readable explanations for why a sequence was flagged.

Example output:

Risk Score: HIGH

Reasons:
• Low sequence similarity to known toxins (possible evasion)
• High functional similarity to toxin protein family
• Mutation pattern inconsistent with natural evolution

Example Scenario

EvoGuard is designed to demonstrate the following situation:

  1. A known toxin protein is mutated or AI-generated into a new variant.
  2. Traditional sequence similarity tools fail to detect it.
  3. EvoGuard identifies functional similarity and evolutionary anomalies.
  4. The system flags the protein as potentially dangerous.

System Architecture

Input Protein Sequence
        ↓
Sequence Similarity Analysis
        ↓
Functional Similarity Analysis
        ↓
Evolutionary Plausibility Check
        ↓
Risk Scoring & Explanation

Project Goals

This prototype aims to illustrate three ideas:

  1. Sequence similarity alone is insufficient for future biosecurity screening.
  2. Functional protein representations can detect hidden similarities.
  3. Evolutionary constraints can help identify artificially designed proteins.

The project is intended as a proof-of-concept for next-generation biosecurity screening tools in an era where AI-assisted biological design is becoming increasingly powerful.


Disclaimer

EvoGuard is a research prototype developed for a hackathon. It is not intended for real-world biosecurity deployment and does not perform comprehensive biological risk analysis.

The goal is to explore conceptual approaches for strengthening safeguards in biological research and biotechnology infrastructure.

About

No description, website, or topics provided.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages