{"id":18004,"date":"2026-01-14T08:00:00","date_gmt":"2026-01-14T02:30:00","guid":{"rendered":"https:\/\/cbirt.net\/staging\/?p=18004"},"modified":"2026-10-22T14:02:23","modified_gmt":"2026-10-22T08:32:23","slug":"toposbio-unveils-topos-1-an-all-atom-foundation-model-for-intrinsically-disordered-proteins","status":"publish","type":"post","link":"https:\/\/cbirt.net\/staging\/toposbio-unveils-topos-1-an-all-atom-foundation-model-for-intrinsically-disordered-proteins\/","title":{"rendered":"ToposBio Unveils Topos-1: An All-Atom Foundation Model for Intrinsically Disordered Proteins"},"content":{"rendered":"\n

Intrinsically disordered proteins (IDPs) are proteins that are central to neurodegenerative diseases and aggressive cancers like prostate cancer, have been considered \u2018undruggable\u2019 as structures predicted by traditional tools to determine how they function and interact, and to be studied in medicine and biotechnology are static. ToposBio introduced Topos-1, an all-atom generative model, which addresses this gap by predicting realistic conformational ensembles, outperforming existing tools across global and local structural evaluations.<\/em><\/p>\n\n\n\n

Issues with \u2018Undruggable Proteins\u2019 and the Need for a Generative Model<\/h2>\n\n\n\n

Many proteins, especially intrinsically disordered proteins (IDPs), constantly shift between 3D structures, rather than existing in a single fixed state. Although this flexibility is important for binding, catalysis, and regulation, it also makes it a challenge to study and design therapeutic drugs tackling diseases linked to these proteins, making them \u2018Undruggable\u2019.<\/p>\n\n\n\n

As IDPs make up nearly a third of the human proteome, ignoring their flexibility leaves a major gap in drug discovery. Current machine learning models, like AlphaFold, are not designed to generate conformational ensembles, a collection of all possible shapes that a protein can adopt.<\/p>\n\n\n\n

Topos-1, an all-atom generative model, addresses this gap by predicting realistic conformational ensembles, meaning they aren\u2019t just random protein structures, but obey the physical and chemical rules of the molecular behaviour. This helps make sure the conformations align with experimental data and known biophysical principles.<\/p>\n\n\n\n

An Introduction to Topos1<\/strong><\/h2>\n\n\n\n

Topos-1 can be considered a breakthrough generative AI model specifically designed to handle the challenge of structure prediction<\/a> for highly flexible proteins. It is the largest IDP-focused simulation corpus to date.<\/p>\n\n\n\n

Instead of training mainly on structured proteins, such as those from the Protein Data Bank (PDB), Topos-1\u2019s architecture and training pipeline are tailored to dynamic motifs by combining in-house physics-based simulations with experimental IDP data.<\/p>\n\n\n\n

As a result of this, it produces physically realistic conformational ensembles that generalize well to novel IDPs and achieve best-in-class accuracy across both global (large-scale structural properties) and local properties (fine-grain details crucial for drug design). <\/p>\n\n\n\n

Topos-1\u2019s performance continues to improve predictably as model size, computation, and training data increase.<\/p>\n\n\n\n

Scaling Potential and Efficiency of Topos-1<\/h2>\n\n\n\n

MD simulations<\/a> are traditional gold standards for studying proteins that lack a stable 3D structure, but they are extremely computationally expensive, requiring days of GPU time.<\/p>\n\n\n\n

A major advantage of Topos-1 is its computational efficiency, as it can be parallelized across many GPUs, speeding up the generation of ensembles by 1000 times compared to molecular dynamics, making it not only accurate but also more practical for large-scale biomedical applications such as drug discovery.<\/p>\n\n\n\n

Evaluations of Topos-1: Setting a new standard for Global and Local Poperties<\/h2>\n\n\n\n

The model\u2019s performance was validated by comparing it against an evaluation dataset of experimental measurements from small-angle X-ray scattering (SAXS) measurements, restricted to IDPs with length <200 residues, as shorter sequences are computationally manageable. To ensure fairness, none of the proteins in the evaluation set was included in the training data. They also excluded any sequences with more than 50% similarity to training data proteins, using MMseqs2 linclust.<\/p>\n\n\n\n

The goal was to prevent the model from being tested on proteins too close to what it has already learned.<\/p>\n\n\n\n