Ankh is an optimized, general-purpose protein language model designed to achieve state-of-the-art performance with significantly fewer parameters, aiming to democratize AI-driven protein engineering research. It provides researchers and the biotech community with accessible, high-performance models for advancing innovation in the field.

How It Works

Ankh models are trained on Google's TPU-V4 infrastructure, enabling efficient development of powerful language representations for protein sequences. The core approach focuses on parameter efficiency, delivering competitive or superior results compared to larger models, thereby reducing computational resource requirements for users. It serves as a foundational model for feature extraction and downstream predictive tasks.

Quick Start & Requirements

• Primary install command: python -m pip install ankh
• Dependencies: PyTorch is implied for model usage. No specific hardware (GPU/TPU) requirements are listed for inference, though training utilized TPU-V4.
• Relevant pages: Models and datasets are available via Huggingface.

Highlighted Details

• Achieves state-of-the-art performance with dramatically fewer parameters, promoting accessibility.
• Demonstrates strong benchmark results across various protein tasks, including Secondary Structure Prediction (e.g., 84.40% Q3 on CASP12 with Ankh3 XLarge NLU), Contact Prediction, Localization, and Remote Homology.
• Offers multiple pre-trained model variants: Ankh Base, Ankh Large, Ankh3 Large, and Ankh3 XL.
• Provides easy access to standard protein datasets through Huggingface integration.

Maintenance & Community

The project is a collaboration involving researchers from the Technical University of Munich, Proteinea, and the University of Columbia. It is sponsored by Google Cloud. Contributions are welcomed, and community interaction is facilitated through the GitHub issues page for questions and bug reporting.

Licensing & Compatibility

Ankh pretrained models are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license. This license restricts usage to non-commercial purposes, posing a significant limitation for commercial applications or integration into proprietary software.

Limitations & Caveats

The primary limitation is the non-commercial clause in its license, which prohibits use in commercial products or services. Additionally, some benchmark results are still marked as "In Progress," indicating ongoing development or validation for specific tasks and model versions.