About This Project

Long-range biosensing is now possible: engineered microbes can be read from ~90 m by hyperspectral cameras. But the reporter signal is weak, buried in natural backgrounds (soil, leaves, water), and few molecules have measured spectra. HyperMix is an open-source, physics-informed deep-learning toolkit that jointly detects and unmixes engineered reporters in noisy remote cubes — with an open simulator, dataset and benchmark. MIT-licensed, built on clinical OCT methods.

Ask the Scientists

What is the context of this research?

Imagine spraying harmless engineered bacteria on a field and reading their chemistry from the sky, getting warned about pollution in the soil, missing nutrients in a crop, or cracks forming in concrete, all without collecting a single sample by hand.

That future just got closer. In a 2026 Nature Biotechnology study, scientists engineered bacteria that emit a specific color, then spotted it from 90 meters away with a hyperspectral camera, a sensor that sees hundreds of shades of light instead of just red, green, and blue.

The catch: the signal is very faint and hides inside the colors of soil, leaves, and water, which change everywhere you look. Today's standard detection methods work well only when the signal is clean and predictable, and it rarely is. My hypothesis is that a smarter, physics-guided algorithm can find these faint signals where the old methods fail. This project builds an open, free tool to test exactly that.

What is the significance of this project?

Better detection algorithms are pure software, so the return per dollar is high: every cheaper camera and every dataset the field funds becomes more useful when software can pull signal from clutter. This work is field-building infrastructure: an open toolkit, an open scene simulator, and an open spectral dataset that others can build on. It needs no wet lab, no new hardware, and no engineered-organism release, so it carries no regulatory friction and ships fast.

Its concrete output is the evidence itself: a public benchmark that quantifies, with confidence intervals, how well each detection method recovers faint engineered reporters against real backgrounds at controlled signal-to-noise, alongside an open dataset and a shared leaderboard. That gives hyperspectral biology a common yardstick it lacks today, so future sensors, reporters, and algorithms can be compared fairly and reproducibly rather than claim by claim.

What are the goals of the project?

This project will deliver an open, reproducible toolkit that detects engineered hyperspectral reporters against unknown backgrounds in noisy remote images. First, I will build a physics-based simulator that composites reporter signatures over real backgrounds with realistic illumination, atmosphere, and noise, producing images with known ground-truth labels. I will then reproduce matched-filter and unmixing baselines as the performance floor, and train a physics-informed network that jointly detects and unmixes, handles unknown backgrounds, and outputs detection maps with calibrated uncertainty. Finally, I will release everything under the MIT open-source license: a python library, Colab notebooks, an open spectral dataset, a public benchmark, and a standard scene format. Success means the benchmark measures whether and where the learned detector improves on classical baselines, and the pipeline runs end-to-end on a real, open hyperspectral cube.