Peptone and NVIDIA introduce PepTron-o, the Ensemble First Structure Prediction for the Disordered Proteome

Peptone and NVIDIA present at GTC Paris 2025

Intrinsically disordered regions (IDRs) constitute more than 30% of the human proteome; however, contemporary leading structural predictors remain primarily tailored for well-folded domains. These predictors generally reduce IDRs to a singular, low-confidence representation often referred to as “spaghetti.”

PepTron-o addresses this issue. Originating from a collaborative effort between Peptone and NVIDIA, it represents the inaugural sequence-to-ensemble model that directly acquires the conformational diversity of disordered proteins and provides an automated approach to enhance any existing ensemble.

Radically lowering the entry bar to Structural Ensemble Generation

  • We developed a synthetic dataset comprising tens of thousands of proteins containing intrinsically disordered regions (IDRs), which were generated using our Oppenheimer platform and reweighted according to high-quality experimental observables such as nuclear magnetic resonance (NMR) chemical shifts.
  • PepTron-o is developed based on NVIDIA BioNeMo. It utilizes various forms of parallelism, fused attention kernels, and mixed-precision arithmetic acquired from large-scale pretraining throughout the processes of fine-tuning to inference, while significantly reducing GPU requirements and facilitating the deployment of this class of protein structure generation models.

A Universal Reweighting Framework to Improve All Ensemble Predictions

Current flagship models possess the capability to generate ensembles, but those mix physically plausible and implausible conformations. We hereby introduce a reweighting procedure that:

  1. Evaluates each structure based on high-quality physical observables,
  2. Reweights the ensemble utilizing experimental readouts and back-propagates the weight updates, thereby enabling the base model to internalize the correction signal,
  3. Enhances the prediction of ensemble generation.

Because this procedure is model‑agnostic, you can drop PepTron-o’s reweighting head on top of any ensemble generator and watch the precision climb.

Introducing a Robust Benchmark Metric for Evaluating Generative Ensemble Quality

Single‑structure accuracy metrics (RMSD, LDDT) are not descriptive on IDRs. We, therefore, propose an ensemble consistency metric that penalizes ensembles that either

  • cannot be reweighted to satisfy the experiment
  • require extreme weight redistribution to do so

Watch the talk live at NVIDIA Stage

Join our CTO Carlo Fisicaro and Cedric Steenbeke – HCLS Startups Developer Relations Lead EMEA at NVIDIA on Friday the 13th at 11.20 am CET at VivaTech 2025, Paris, where he’ll present:

“Redefining how we generate and evaluate multi-domain protein structural ensembles.”

About Peptone

Peptone is a biotechnology company developing novel first-in-class small molecule therapeutics targeting intrinsically disordered proteins (IDPs) to treat diseases through a pioneering experimental approach that starts with studying the structural dynamics of IDPs. Peptone’s technology accurately translates bespoke experimental data into dynamic models of disordered proteins regardless of protein size or complexity with the necessary speed and insights to design therapeutics capable of affecting IDP function. The company’s initial focus is on oncology targets areas of highly unmet need with strong disease hypotheses and biological validation. For more information about Peptone, visit https://peptone.io.