Baktfold: Sensitive protein functional annotation across the microbial tree of life using structural information

Baktfold is a new, ultra-sensitive, and taxon-independent command-line tool that leverages protein structure information via Foldseek and ProstT5 to significantly improve the functional annotation of hypothetical proteins across the microbial tree of life compared to existing methods like Bakta and Prokka.

The Gist

Imagine the microbial world (bacteria, archaea, and tiny eukaryotes) as a massive, ancient library containing billions of books. Each "book" is a genome, and the "sentences" inside are proteins. For decades, scientists have been trying to read these books to understand what the microbes are doing.

However, there's a huge problem: a massive chunk of these sentences are written in a code we don't understand yet. In the scientific world, these are called "hypothetical proteins." It's like finding a page in a book that just says "Do something important here," but with no instructions on what that something is. For a long time, we've been stuck staring at these blank pages.

Enter Baktfold, a new digital tool designed by George Bouras and his team to finally translate these mysterious sentences.

Here is how Baktfold works, explained through simple analogies:

1. The Old Way: Looking at the Spelling (Sequence Homology)

Traditionally, scientists tried to understand a protein by comparing its "spelling" (its amino acid sequence) to other proteins they already knew.

• The Analogy: Imagine you find a word in a foreign language. You try to guess its meaning by looking for words that look similar. If the foreign word is "Cat" and you know "Cat" means a feline, you guess the new word means "Feline."
• The Problem: If the foreign word is spelled "Kyt," you might miss the connection. In biology, proteins can change their "spelling" over millions of years so much that they look completely different, even though they still do the exact same job. This is called the "Twilight Zone"—where the spelling is too different to recognize, but the meaning is still there.

2. The New Way: Looking at the Shape (Structure-Based)

Baktfold takes a different approach. It realizes that while the spelling of a protein might change, its 3D shape (its structure) stays very similar because the shape determines what the protein actually does.

• The Analogy: Imagine you find a strange, twisted piece of metal. You don't know what it is. But if you look at its shape, you realize it looks exactly like a key. Even if the metal is rusty and the key is made of plastic, the shape tells you it opens a door.
• The Magic: Baktfold uses a super-smart AI (called ProstT5) to instantly predict the 3D shape of these mysterious proteins. It then compares these shapes against a giant library of known shapes (like the AlphaFold database) to find matches.

3. The "Super-Spy" Search

Baktfold doesn't just look at one library; it acts like a super-spy checking four different databases at once:

1. Swiss-Prot: The "Gold Standard" library of perfectly curated, high-quality protein descriptions.
2. AlphaFold Database: A massive library of predicted shapes for almost every known protein.
3. PDB: The library of shapes that have been physically measured in labs.
4. CATH: A library organized by protein "families" and shapes.

It checks the shape of your mystery protein against all four. If it finds a shape match, it says, "Aha! This mystery protein looks just like a known enzyme that breaks down sugar!"

Why is this a Big Deal?

• It's a Speed Demon: Usually, predicting 3D shapes takes hours or days on powerful supercomputers. Baktfold uses a shortcut (the AI language model) to do this in minutes, making it fast enough to scan entire genomes instantly.
• It Solves the "Dark Matter" Problem: The paper tested Baktfold on hundreds of thousands of bacteria and archaea.
  • Bacteria: It successfully identified the function of 50% of the "mystery proteins" that the standard tools (like Bakta) couldn't solve.
  • Archaea: This is where it shines brightest. Archaea are weird, ancient microbes that are notoriously hard to study. Standard tools only understood about 36% of their proteins. Baktfold jumped that to 71.5%, effectively doubling our knowledge of these organisms.
• It Works Everywhere: It works on bacteria, archaea, plasmids (tiny DNA rings), and even tiny eukaryotes (like plankton).

The Bottom Line

Think of Baktfold as a universal translator for the microscopic world. Before, we could only read the easy, familiar sentences. Now, with Baktfold, we can look at the shape of the words to understand the difficult, ancient, and mysterious ones.

This doesn't just fill in gaps in a database; it allows scientists to finally ask, "What is this microbe actually doing?" which could lead to new antibiotics, better ways to clean up pollution, or a deeper understanding of how life evolved on Earth. It turns the "hypothetical" into the "known."

Technical Summary

1. Problem Statement

Despite significant advances in genome annotation, a substantial fraction of microbial proteins (estimated at ~30% in bacteria and even higher in archaea and micro-eukaryotes) remain classified as "hypothetical proteins."

• Limitations of Current Tools: State-of-the-art annotation pipelines (e.g., Bakta, Prokka, DFAST) rely primarily on sequence homology (BLAST, HMMs, profile alignments). These methods struggle in the "twilight zone" of sequence identity (20–35%), where sequence similarity is too low to infer function, even if the protein structure and function are conserved.
• Computational Barriers: While protein structure prediction (e.g., AlphaFold2, ESMFold) offers a solution by leveraging the fact that structure is more conserved than sequence, applying these tools at a genome or metagenome scale is computationally prohibitive due to high GPU and memory requirements.
• Gap: There is a lack of automated, scalable tools that integrate structural information into standard genome annotation workflows to resolve these "dark matter" proteins.

2. Methodology

Baktfold is a command-line software tool designed to bridge this gap by integrating protein language models (pLMs) and structural search algorithms into a high-throughput annotation pipeline.

• Core Workflow:

  1. Input: Accepts Bakta JSON outputs (containing predicted CDS) or raw protein FASTA files. It specifically targets unannotated "hypothetical proteins" to optimize resource usage.
  2. Structure Prediction via pLM: Instead of running full 3D structure prediction (which is slow), Baktfold uses the ProstT5 protein language model. ProstT5 predicts Foldseek 3Di tokens (a discrete representation of protein structure) directly from the amino acid sequence. This is orders of magnitude faster than AlphaFold2 or ESMFold.
  3. Structural Search: The predicted 3Di sequences are searched against four complementary structure databases using Foldseek (a fast structural alignment tool):
     • Swiss-Prot: Curated, high-quality annotations.
     • AlphaFold Database (AFDB) Clusters: Large-scale predicted structures (v6 release).
     • PDB: Experimentally resolved structures.
     • CATH: Domain-level structural classification.
  4. Prioritization & Output: Results are prioritized based on database quality (Custom > Swiss-Prot > AFDB > PDB > CATH). Outputs are generated in GFF3, INSDC-compliant flat files, and JSON, ensuring 100% interoperability with Bakta and compatibility with NCBI submission tools.

• Database Construction: The authors constructed a custom database containing ~3.1 million AFDB cluster representatives, alongside Swiss-Prot, PDB, and CATH, all indexed for Foldseek.

• Benchmarking Strategy:

  • Datasets: 305,825 bacterial/archaeal genomes (GlobDB), 8.8M plasmid proteins (IMG/PR), 241 protist genomes (Ensembl), and 713 eukaryotic plankton genomes (SMAG).
  • Comparators: Compared against Bakta, Prokka, eggNOG-mapper, and structural benchmarks using ColabFold/AlphaFold2 and ESMFold.
  • Hardware: Benchmarked on high-performance computing nodes (AMD MI250x and NVIDIA A100 GPUs).

3. Key Contributions

• Novel Integration: First tool to seamlessly integrate pLM-derived structural representations (3Di) into a standard microbial annotation pipeline, enabling structural homology detection at scale.
• Computational Efficiency: By using ProstT5 for 3Di token generation rather than full 3D structure prediction, Baktfold achieves near-AlphaFold2 sensitivity with a runtime of minutes per genome (vs. hours/days for full structure prediction).
• Taxonomic Breadth: Demonstrated efficacy not just in bacteria, but significantly in Archaea and Micro-eukaryotes, groups historically difficult to annotate due to low sequence conservation.
• Customizability: Supports user-defined custom databases, allowing researchers to curate specific protein libraries (e.g., archaeal-specific) to further boost sensitivity.

4. Key Results

• Bacterial Annotation:

  • Overall Rate: Baktfold achieved a median overall annotation rate of 87.8% across bacterial genomes, compared to 72.9% for Bakta.
  • Hypothetical Proteins: It annotated 50.1% of proteins that remained hypothetical after Bakta processing.
  • Plasmids: Improved annotation of plasmid proteins by 29.5% over Bakta, with significant gains for proteins >100 amino acids.

• Archaeal Annotation (Major Breakthrough):

  • Baktfold achieved a median annotation rate of 71.5%, vastly outperforming Prokka (35.8%) and Bakta (10.3%).
  • It annotated 68.0% of hypothetical archaeal proteins.
  • Even with a custom curated archaeal database, Baktfold maintained a lead over Prokka (85.6% vs. 79%), proving structure-based methods are superior to sequence alignment for archaea.

• Micro-Eukaryotes:

  • On protist genomes, Baktfold annotated 70.0% of CDS, surpassing the 60.7% rate of existing Gene Ontology (GO) annotations in reference databases.
  • On environmental plankton (SMAG), Baktfold achieved a 50.6% functional annotation rate vs. 39.6% for eggNOG-mapper.

• Performance vs. Full Structure Prediction:

  • When compared to ColabFold/AlphaFold2 and ESMFold, Baktfold (using ProstT5) showed comparable sensitivity (median hypothetical annotation rates within 1-2% of full structure methods) but with drastically reduced wall-clock time (30–450 seconds vs. hours).

• Twilight Zone Detection:

  • Case studies confirmed Baktfold's ability to detect homologs with <30% sequence identity (e.g., RecBCD enzymes, transport proteins) by identifying strong structural similarities (low RMSD) that sequence tools missed.

5. Significance

• Illuminating Microbial Dark Matter: Baktfold provides a scalable solution to functionally characterize the vast number of unannotated proteins in microbial genomes, particularly in under-studied domains like Archaea.
• Paradigm Shift: It moves the field from purely sequence-based annotation to structure-informed annotation as a standard, automated step in genomic pipelines.
• Accessibility: Being open-source (MIT license), available via Bioconda/PyPI, and compatible with existing workflows (Bakta), it lowers the barrier for researchers to utilize structural biology for functional genomics without needing massive GPU clusters for full structure prediction.
• Future Research: The high-confidence annotations generated by Baktfold can guide targeted in vitro experiments and hypothesis generation for protein function, accelerating the understanding of microbial ecology and evolution.

Availability:

• Software: GitHub (gbouras13/baktfold), Bioconda, PyPI.
• Databases: Zenodo and HuggingFace.
• Analysis Code: GitHub (gbouras13/baktfold-analysis).