Project 03 · Computational Biology

HelixForge

Invents new antimicrobial peptides and screens them for activity and toxicity, then proves the numbers with homology-disjoint evaluation that most published work skips.

Scroll to watch how it works
How it works

Scene 1 / 6

Twenty letters.

Every protein is written in an alphabet of twenty amino acids.

Scene 2 / 6

A new peptide is written.

HelixForge generates new antimicrobial peptide sequences, letter by letter.

Scene 3 / 6

It folds into a helix.

Many antimicrobial peptides coil into a helix: one face oily, one face charged.

Scene 4 / 6

Screened twice.

Two classifiers score every candidate: one for activity against microbes, one for toxicity.

Scene 5 / 6

Tested on strangers.

Homology-disjoint evaluation: test peptides have no close relatives in training, so the score can't be memorised.

Scene 6 / 6

All in your browser.

The models run on your own device with ONNX Runtime Web and Transformers.js, at zero cost.

Open HelixForge
ARNDCQEGHILKMFPSTWYVActivity classifier ✓Toxicity classifier ✓TRAINING FAMILIESTEST FAMILIESNo close relatives across the split
huggingface.co/spaces/Shrugal/helixforge-demo
Runs on your device · zero cost

The details

Under the hood.

HelixForge invents new antimicrobial peptides and screens them for activity and toxicity, then proves the numbers mean something, using homology-disjoint evaluation that most published work skips. Everything runs in your browser at zero cost.

  • Homology-disjoint evaluation
  • Activity and toxicity scoring
  • Runs in the browser at zero cost
Peptide Generator
Dual Classifier Scoring
Leakage-Free Evaluation
Reward-Hacking Audit
3D Structure Prediction
In-Browser Inference

Built with

ESM-2PyTorchONNX Runtime WebTransformers.jsscikit-learnJavaScriptPythonHugging Face

Next project

STRATUM

Climate & Aviation AI