AI for Disordered Proteins: IDPForge Models the Floppy Parts of Biology

AI tools like AlphaFold transformed biology by predicting the folded shapes of proteins. But many proteins, or parts of them, never settle into a single shape. They stay floppy and constantly moving. A new peer-reviewed study in Nature Communications introduces IDPForge, an open-source tool that brings AI for disordered proteins a step forward by generating realistic pictures of these shape-shifting molecules.

It’s a research tool for scientists rather than something you’ll see in a clinic, but it tackles one of the trickiest blind spots in protein science.

The problem with floppy proteins

Most of us picture proteins as neatly folded 3D shapes, like a key that fits a lock. That’s true for many of them. But some proteins, called intrinsically disordered proteins (IDPs), and some sections of otherwise folded proteins, called intrinsically disordered regions (IDRs), don’t have one fixed structure.

Instead, they’re best described as an ensemble: a large collection of different shapes the protein moves between.

Where today’s AI struggles

According to the researchers, popular structure-prediction tools such as AlphaFold and RoseTTAFold predict these disordered regions only with low confidence. Those tools were built to find a protein’s single folded state, and a floppy protein doesn’t have one.

How AI for disordered proteins works in IDPForge

IDPForge stands for Intrinsically Disordered Protein, FOlded and disordered Region GEnerator.

A protein language diffusion model

Under the hood, IDPForge uses a transformer protein language diffusion model. In plain terms, it combines two popular AI ideas:

• Protein language models, which learn patterns in protein sequences much like chatbots learn patterns in text.
• Diffusion models, the same broad approach behind many AI image generators, which build outputs step by step from noise.

Full detail, folded parts intact

IDPForge creates all-atom ensembles, meaning every atom is represented, not just a simplified sketch. For proteins that mix folded and floppy sections, it models the disordered regions while keeping the folded domains intact.

Fewer extra steps

The researchers highlight that IDPForge does not require:

• training a new model for each protein sequence
• converting from simplified, coarse-grained models back to full detail
• reweighting the ensemble afterward

That could make it simpler and faster for scientists to use.

Does it match reality?

The key test is whether the generated shapes agree with real lab measurements. The team compared IDPForge’s ensembles with two widely used experimental techniques:

• NMR (nuclear magnetic resonance), which probes protein structure and motion in solution
• SAXS (small-angle X-ray scattering), which reveals a molecule’s overall size and shape in solution

According to the paper, the ensembles showed good agreement with both. Scientists can also optionally guide the model with their own experimental data.

Where earlier AI tools shrugged at floppy proteins, IDPForge produces detailed shape collections that line up with real measurements.

Limitations to keep in mind

• A research tool, not a treatment. IDPForge helps scientists study proteins. Any benefit to patients would be indirect and likely years away.
• “Good agreement” isn’t perfection. Disordered proteins are hard to measure, and matching lab data well doesn’t mean every predicted shape is exactly right.
• Early days. Broader testing by other research groups on many more proteins will show how well it performs in practice.

Why it matters

Because disordered proteins don’t hold still, they’ve been hard to study with both experiments and AI. Better models can help researchers understand how these flexible molecules behave, which supports deeper structural studies of proteins that contain disorder.

Importantly, IDPForge is open source, so labs around the world can try it, test it and build on it.

The bottom line on AI for disordered proteins

IDPForge shows that AI for disordered proteins can go beyond low-confidence guesses, generating detailed, lab-consistent pictures of molecules that never sit still. It’s a valuable new tool for scientists exploring one of biology’s most flexible frontiers.

Read the study in Nature Communications: IDPForge: deep learning of proteins with global and local regions of disorder.