<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> tangermeme addresses the post-training analysis of sequence-based machine learning models in genomics. It offers researchers robust, efficient, and assumption-free tools for applying models, dissecting learned features, and designing novel sequences, streamlining genomic sequence model workflows.

How It Works

The library implements atomic sequence operations (motif manipulation, shuffling) and tools for applying predictive models. Its "assumption-free" design supports arbitrary PyTorch models, multi-input/output architectures, and user-supplied loss functions, enabling frictionless integration. Operations are unit-tested and optimized for efficiency, extensible to any alphabet.

Quick Start & Requirements

• Installation: pip install tangermeme or uv add tangermeme.
• Prerequisites: PyTorch. Installation can take up to 10 minutes.
• Optional: tangermeme[docs] for local documentation builds.
• Development: Clone repo, use uv sync --extra dev.
• Claude Code Skill: Install via tangermeme-install-skills.
• Docs: Links to preprint, docs, tutorials, vignettes are mentioned.

Highlighted Details

• Assumption-Free Design: Accommodates flexible model architectures and user-defined loss functions.
• Sequence Manipulation: Functions for motif insertion, substitution, and shuffling in tangermeme.ersatz.
• Model Interpretation: Tools for prediction, attribution (DeepLIFT/SHAP), marginalization, ablation, and saturation mutagenesis.
• Sequence Design: Algorithms like greedy_substitution for generating sequences with desired model predictions.
• Seqlet Calling: Identifies functional genomic regions based on attribution scores, with optional motif annotation.
• Claude Code Integration: Simplifies tangermeme usage with the Claude Code AI assistant.

Maintenance & Community

The project roadmap indicates phased development, with core prediction and attribution functionalities complete. Future releases target command-line tools, iterative approaches, and motif discovery. No explicit community links were found.

Licensing & Compatibility

The license type and compatibility notes for commercial use are not specified in the provided README.

Limitations & Caveats

• FIMO/Tomtom command-line tools moved to memesuite-lite (PyTorch-free).
• High flexibility in model wrapping/broadcasting may lead to bugs; user reporting is encouraged.
• Users must supply custom loss functions when needed.