← Latest papers
🧬 genetics

An Interactive Bidirectional De Novo Protein Design System

This paper introduces an open-source, interactive web server that integrates a structure-aware GNN for sequence design and an ESMFold-based predictor into a bidirectional loop, enabling researchers to visualize, quantify, and validate de novo protein designs without requiring software installation.

Original authors: Gammill, C. M., Tamil, L.

Published 2026-10-28
📖 5 min read🧠 Deep dive

Original authors: Gammill, C. M., Tamil, L.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the vast library of life, the instructions for building a protein are written in a chemical alphabet of just twenty letters. Yet, the shape a protein takes is what determines its function, acting as a key that fits into specific locks within the body. For decades, scientists have struggled with a two-way puzzle: predicting what shape a sequence of letters will form, and conversely, figuring out what sequence of letters is needed to build a specific shape. While computers have become remarkably good at guessing shapes from sequences, the reverse process—designing a new sequence to fit a custom-made shape—has often been a slow, technical ordeal reserved for experts with specialized software and powerful computers. This gap between high-level statistical success and the ability to visually inspect a single, custom design has left many researchers unable to easily test their own creative ideas about how proteins could be built from scratch.

A team of researchers at the University of Texas at Dallas has built a new tool to bridge this gap, creating a lightweight, interactive system that runs entirely in a web browser. Instead of requiring users to install complex software or write code, this system allows anyone to draw a protein skeleton, ask the computer to invent a sequence of amino acids that fits it, and then immediately see if that new sequence folds back into the intended shape. The researchers combined two powerful artificial intelligence models into a single loop: one model that designs sequences based on a shape, and another that predicts the shape a sequence will take. By connecting them, the tool lets a user construct a backbone, generate a sequence, fold it, and then compare the result to the original drawing, all within a single visual interface.

The system operates through two main modes that work together like a conversation between a designer and a builder. In the first mode, a user can construct a protein backbone from scratch. They can place individual building blocks, drag to draw a continuous chain, or insert idealized helices and strands, much like snapping together pieces of a model kit. They can also upload existing protein structures from public databases to use as a starting point. Once a shape is defined, the system asks its sequence-design model to invent a string of amino acids that should hold that shape together. The computer does not just offer one answer; it provides a list of possibilities along with a confidence score for each position, showing the user how sure the model is about its choices.

In the second mode, the user takes that newly generated sequence and asks the folding model to predict what three-dimensional structure it will actually form. This is where the true test happens. The system renders the predicted shape in 3D, coloring it to show which parts the model is confident about and which parts are uncertain. To make the comparison meaningful, the tool includes a built-in alignment feature that mathematically superimposes the predicted shape onto the original design. It calculates a precise measurement of how far apart the atoms are in the two structures, giving a single number that quantifies the success of the design. If the shapes do not match well, the user can tweak the original drawing or try a different sequence, iterating instantly without leaving the browser.

To demonstrate the tool's capabilities, the researchers tested it on a real protein chain imported from a database. They asked the system to design a sequence for that existing shape, then folded that new sequence back into a structure. When they aligned the result with the original, the atoms were separated by an average distance of 1.174 angstroms, a very small gap in molecular terms. To ensure this was a fair test, they also ran the original, natural sequence of that protein through the folding model. That natural sequence resulted in a slightly tighter match of 0.999 angstroms. This comparison suggests that the new design tool is performing nearly as well as nature itself for this specific case, successfully closing the loop between design and prediction.

The researchers emphasize that this tool is not a replacement for deep scientific study but rather a way to make the process of protein design accessible and visual. It allows a researcher to see exactly where a design fails, such as a kink in the wrong place or a mismatch in chemical properties, rather than just looking at a summary statistic. The code is open-source, meaning other scientists can use it with their own trained models or adapt it for their specific needs. While the system currently requires the user to run the software on their own computer with a powerful graphics card, the researchers hope this accessible interface will encourage more experimentation and hypothesis testing in the field of protein engineering, turning a complex computational task into a visual, interactive experience.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →