ProtBFN

ProtBFN is a state-of-the-art generative protein modelling tool available on the DeepChain platform, developed to help researchers explore the proteome and design novel, biologically plausible protein sequences. Built on a 650-million-parameter Bayesian Flow Network (BFN) architecture and trained on a curated dataset of 72 million biologically validated examples, ProtBFN is purpose-built for life sciences R&D teams seeking to accelerate protein discovery and therapeutic development.

Proteins are composed of 20 proteinogenic amino acids arranged in sequences that determine their three-dimensional structure and biological function. The space of possible protein sequences is astronomically large, making manual exploration impractical. ProtBFN addresses this challenge by learning to generate sequences that are both novel and biologically meaningful, equipping researchers with powerful tools to explore uncharted regions of the proteome and drive advancements in healthcare applications.

Why Generative AI Matters for Protein Sequencing

• Functional proteins can range from 10 to thousands of amino acids in length, creating a search space far beyond human comprehension.
• Generative AI focuses on biologically plausible sequences, identifying and creating candidate sequences with the most promise for further research.
• Protein generation falls into two key areas: unconditional (de novo) generation, where entirely new proteins are created without specific guidance, and conditional generation, where the model is directed to complete partial sequences or design proteins with specific properties.
• Traditional models such as GPT process data in a fixed left-to-right order, which fails to capture the non-sequential interactions within protein sequences, leading to less accurate outcomes.
• ProtBFN overcomes these limitations by generating meaningful and diverse protein sequences that are both novel and natural, demonstrating flexibility across both unconditional and conditional generation tasks.

How ProtBFN Works

• Bayesian Flow Networks (BFNs) model the underlying patterns and relationships within data rather than simply assembling sequences, making them well-suited to tasks where order is not fixed or essential.
• Like other foundational models, BFNs are pre-trained on vast datasets to identify general patterns and representations, but their unique approach enables a wide range of tasks with remarkable flexibility.
• ProtBFN supports zero-shot conditional generation, handling tasks without specific training, even in unfamiliar scenarios.
• When trained on data from the Observed Antibody Space (OAS), ProtBFN demonstrated exceptional performance in generating antibody variable heavy (VH) chain sequences.
• ProtBFN outperforms specialised inpainting models such as AntiBERTy and AbLang2 in antibody inpainting — predicting or generating missing parts of an antibody sequence from partially obstructed data.
• For antibody VH chains, ProtBFN can complete critical segments including Complementarity-Determining Regions (CDRs), which determine binding specificity, and Framework Regions (FRs), which provide structural stability.
• Notably, ProtBFN achieves this inpainting capability without having been specifically trained for such tasks, demonstrating the breadth of its learned understanding of antibody composition and sequence relationships.

Key Performance Characteristics

• 10,000 generated sequences from ProtBFN are matched to clusterings from UniRef50, with a coverage score measuring the ratio of unique clusters hit to the expected number of sequences drawn from the model's training distribution.
• Generated sequences with less than 100% identity to the best matching protein in the training data are considered novel sequences, confirming ProtBFN's ability to produce genuinely new proteins.
• ProtBFN generates discrete data, captures complex dependencies, and adapts to both unconditional and conditional tasks, overcoming the limitations of traditional generative models.

Future Direction and Vision

• Researchers have highlighted the formal connection between BFNs and Diffusion models, opening the door to advanced sampling methods that could further enhance the generative power of BFNs.
• The diffusion-like approach to data processing, combined with BFN versatility, highlights strong potential for continued advances in protein sequence modelling.
• The long-term vision extends beyond protein sequence modelling to modelling sequences alongside their full spectrum of associated metadata, building foundational models that capture the joint distribution of diverse scientific data for a more comprehensive view of the biological landscape.
• The goal is a model with superior performance by learning across multiple data types and sources, offering unparalleled flexibility for task-specific inference in the hands of scientists.

ProtBFN is available on the DeepChain platform, with an accompanying research paper, open-source training dataset, and model weights accessible on Hugging Face and GitHub. It is part of a broader suite of DeepChain AI tools for life sciences, including AbBFN for antibody design, Protein Folding models, InstaNovo for peptide sequencing, and Nucleotide Transformers for genomic analysis.