FlexRibbon

Protein foundation models have historically split into two families: sequence-only language models that learn from amino acid strings, and structure-based models that operate on three-dimensional coordinates. Each captures part of the picture, but neither natively unifies evolutionary sequence signal with geometric detail in a single pretrained representation. FlexRibbon is designed to close this gap by learning jointly from both modalities.

FlexRibbon is a pretrained protein model that learns simultaneously from amino acid sequences and 3D structures. Its pretraining combines masked language modeling over sequence with diffusion-based denoising over structure, giving a bidirectional objective that ties residue identity to geometry without relying on multiple sequence alignments at inference. The model is trained on a large structural corpus spanning both experimentally resolved structures and AlphaFold2 predictions, allowing it to capture global folds as well as the flexible conformations that matter for function.

Developed by researchers at Beijing Zhongguancun Academy and collaborators and released as a preprint in October 2025, FlexRibbon reports state-of-the-art results across twelve downstream tasks and has been accepted for ICLR 2026.

#Key Features

• Joint sequence-structure pretraining: A single model consumes both amino acid sequence and 3D structure, rather than treating them as separate specialized encoders.
• Dual self-supervised objective: Masked language modeling on sequence is combined with diffusion-based denoising on structure, coupling residue prediction with geometric reconstruction.
• MSA-free operation: The model captures evolutionary and structural signal without requiring multiple sequence alignments at inference, simplifying deployment.
• Broad benchmark coverage: FlexRibbon reports state-of-the-art performance across twelve protein modeling tasks spanning representation and generation.

#Technical Details

FlexRibbon is a 3-billion-parameter transformer-based protein foundation model trained directly on amino acid sequences and large-scale structural data. Pretraining unifies two self-supervised signals: masked language modeling reconstructs hidden residues from sequence context, while diffusion-based denoising reconstructs perturbed 3D coordinates, so the model must reconcile sequence and structure within a shared representation. The structural corpus combines experimentally determined structures with AlphaFold2 predictions, enabling coverage of both rigid global folds and flexible conformational states. Evaluated across twelve downstream tasks, FlexRibbon achieves state-of-the-art results reported in the accompanying preprint.

#Applications

FlexRibbon targets protein scientists who need a single backbone that serves both predictive and generative workflows. Its joint representation can be applied to structure-aware property prediction, functional annotation, and generative design, and its explicit modeling of flexibility makes it relevant for studying conformational variability that static single-structure predictors overlook.

#Impact

FlexRibbon advances the unification of sequence- and structure-based protein modeling into one pretrained foundation model, showing that a combined masked-modeling and diffusion objective can deliver leading performance across a wide task suite. Its acceptance at ICLR 2026 and breadth of benchmark coverage position it as a reference point for multimodal protein foundation models. As a preprint, its reported results await broader independent validation.