lattice-sub 1.3.1

Lattice subtraction for cryo-EM micrographs - removes periodic crystal signals to reveal non-periodic features

Lattice Subtraction for Cryo-EM

.__          __    __  .__                                   ___.    
|  | _____ _/  |__/  |_|__| ____  ____             ________ _\_ |__  
|  | \__  \\   __\   __\  |/ ___\/ __ \   ______  /  ___/  |  \ __ \ 
|  |__/ __ \|  |  |  | |  \  \__\  ___/  /_____/  \___ \|  |  / \_\ \
|____(____  /__|  |__| |__|\___  >___  >         /____  >____/|___  /
          \/                   \/    \/               \/          \/ 

Remove periodic lattice patterns from cryo-EM micrographs to reveal non-periodic features.

---

Installation

pip install lattice-sub

That's it! GPU acceleration works automatically if you have an NVIDIA GPU.

Note: Requires Python 3.11+ and an NVIDIA GPU (RTX 20/30/40 series, A100, etc.) for best performance. CPU fallback is available but slower.

---

Quick Start

Process a Single Image

lattice-sub process your_image.mrc -o output.mrc --pixel-size 0.56

Pixel size: Use your detector's actual pixel size (e.g., K3=0.56Å, Falcon=1.14Å, K2=1.32Å)

Process a Folder of Images

lattice-sub batch input_folder/ output_folder/ --pixel-size 0.56

Generate Comparison Images

lattice-sub batch input_folder/ output_folder/ --pixel-size 0.56 --vis comparisons/

This creates 4-panel PNG comparison images for each micrograph showing:

1. Original - Input micrograph
2. Subtracted - Lattice-removed result
3. Difference - What was removed (5x amplified)
4. Threshold Curve - Threshold vs lattice removal efficacy

Limit the number of visualizations:

# Generate visualizations for first 10 images only
lattice-sub batch input_folder/ output_folder/ -p 0.56 --vis comparisons/ -n 10

---

Common Options

Option	Description
-p, --pixel-size	Required. Pixel size in Ångstroms
-o, --output	Output file path (default: sub_<input>)
-t, --threshold	Peak detection sensitivity (default: auto - optimized per image)
--vis DIR	Generate 4-panel comparison PNGs in DIR
-n, --num-vis N	Limit visualizations to first N images
--cpu	Force CPU processing (GPU is used by default)
-q, --quiet	Hide the banner and progress messages
-v, --verbose	Show detailed processing information

Example with Options

# Process with auto-optimized threshold (default - recommended)
lattice-sub process image.mrc -o cleaned.mrc -p 0.56 -v

# Override with a specific threshold if needed
lattice-sub process image.mrc -o cleaned.mrc -p 0.56 -t 1.5 -v

# Batch process, force CPU with 8 parallel workers
lattice-sub batch raw/ processed/ -p 0.56 --cpu -j 8

---

Using a Config File (Optional)

For most users, the defaults work great — just use -p for pixel size:

lattice-sub batch input/ output/ -p 0.56

Config files are useful for reproducibility or non-standard samples:

# params.yaml - only include what you want to override
pixel_ang: 0.56
unit_cell_ang: 120      # Default: 116 (nucleosome). Change for other crystals.
inside_radius_ang: 80   # Default: 90. Protect different resolution range.
# threshold: 1.45       # Uncomment to override auto-optimization

Then use it:

lattice-sub process image.mrc -o output.mrc --config params.yaml

Create a starter config file:

lattice-sub init-config params.yaml --pixel-size 0.56

---

GPU Troubleshooting

GPU should work automatically. If it doesn't:

# Check GPU status
lattice-sub setup-gpu

# Or verify manually
python -c "import torch; print(torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'No GPU')"

Requirements: NVIDIA GPU with driver 525+ (RTX 20/30/40 series, A100, etc.)

---

Multi-GPU Support

When processing batches on systems with multiple GPUs, files are automatically distributed across all available GPUs for faster processing. No extra flags needed!

# Automatically uses all available GPUs
lattice-sub batch input_folder/ output_folder/ -p 0.56

Example with 2 GPUs and 100 images:

• GPU 0: processes images 1-50
• GPU 1: processes images 51-100
• Single progress bar shows combined progress

This provides near-linear speedup with additional GPUs.

---

HPC Example (CU Boulder Alpine)

Using Open OnDemand Core Desktop with 2× RTX 8000 GPUs:

# Create environment
module load anaconda
conda create -n lattice_test python=3.11 -y
conda activate lattice_test
pip install lattice-sub

# Process 100 micrographs
lattice-sub batch input/ output/ -p 0.56

Output:

Phase-preserving FFT inpainting for cryo-EM  |  v1.3.0

  Configuration
  -------------
    Pixel size:  0.56 A
    Threshold:   auto
    Backend:     Auto → GPU (Quadro RTX 8000)

  Batch Processing
  ----------------
    Files:    100
    Output:   /projects/user/output
    Workers:  1

✓ Using 2 GPUs: GPU 0, GPU 1

  ✓ GPU 0: Quadro RTX 8000
  ✓ GPU 1: Quadro RTX 8000

  Processing: 100%|█████████████████████████| 100/100 [05:12<00:00,  3.13s/file]

  [OK] Batch complete (312.9s)

100 images processed in ~5 minutes with automatic multi-GPU distribution.

For compute-focused workloads, use Alpine's GPU partitions (A100, L40, MI100):

# Request 3 GPUs for 1 hour
sinteractive --partition=aa100 --gres=gpu:3 --ntasks=16 --nodes=1 --time=01:00:00 --qos=normal

---

Python API

from lattice_subtraction import LatticeSubtractor, Config

# Configure and process
config = Config(pixel_ang=0.56)
subtractor = LatticeSubtractor(config)

result = subtractor.process("input.mrc")
result.save("output.mrc")

Batch Processing

from lattice_subtraction import BatchProcessor, Config

config = Config(pixel_ang=0.56)
processor = BatchProcessor(config)

stats = processor.process_directory("raw/", "processed/")
print(f"Processed {stats.successful}/{stats.total} files")

---

What It Does

This tool removes the periodic "lattice" pattern from 2D crystal cryo-EM images:

1. Finds lattice peaks in the Fourier transform (the repeating pattern)
2. Replaces them with averaged local values (inpainting)
3. Preserves phase so the image stays accurate
4. Returns the cleaned image with non-periodic features visible

Key Parameters Explained

Parameter	What it controls
pixel_ang	Your detector's pixel size (check your microscope settings)
threshold	How aggressively to detect peaks. Higher = fewer peaks removed
inside_radius_ang	Protect low-resolution features (default 90Å is good for nucleosomes)
unit_cell_ang	Crystal repeat distance (116Å for nucleosome arrays)

---

Need Help?

# See all commands
lattice-sub --help

# See options for a specific command
lattice-sub process --help
lattice-sub batch --help

---

Citation

@software{lattice_sub,
  title = {Lattice Subtraction for Cryo-EM Micrographs},
  author = {Stephenson, George and Kasinath, Vignesh},
  year = {2026},
  url = {https://github.com/gsstephenson/cryoem-lattice-subtraction}
}

Original MATLAB algorithm: Vignesh Kasinath
Python package: George Stephenson
Kasinath Lab (Dept of Biochemistry) & BioFrontiers Institute, University of Colorado Boulder

---

License

MIT License - see LICENSE for details.

---

📚 Advanced Topics (click to expand)

v1.1.0 Optimizations

Version 1.1.0 introduces two major optimizations that make the tool both faster and smarter:

1. Adaptive Per-Image Threshold Optimization

Problem (v1.0.x): A fixed threshold (1.42) was used for all images, but optimal thresholds vary by image quality, ice thickness, and lattice order.

Solution (v1.1.0): Automatic per-image optimization using grid search:

• Tests 21 threshold values in range [1.40, 1.60] with 0.01 step
• Scores each using a quality function that balances:
  • Peak count (target: ~600 peaks for good lattice coverage)
  • Peak SNR (signal-to-noise of detected peaks)
  • Peak distribution (uniform hexagonal spacing preferred)
  • Coverage (adequate sampling across frequency space)
• Returns the threshold with highest quality score

Result: Each image gets its optimal threshold automatically.

2. GPU-Accelerated Background Subtraction (Kornia)

Problem (v1.0.x): Profiling revealed background subtraction (scipy median filter) consumed 94% of processing time — not FFT as expected.

Solution (v1.1.0): Replaced scipy's CPU median filter with Kornia's GPU implementation:

# Before (v1.0.x) - CPU, ~2.5s per image
from scipy.ndimage import median_filter
background = median_filter(log_power, size=51)

# After (v1.1.0) - GPU, ~0.05s per image  
from kornia.filters import median_blur
background = median_blur(log_power_tensor, (51, 51))

Result: 48x speedup on background subtraction alone.

Performance Comparison

Version	Threshold	Background Sub	Time/Image	Quality
v1.0.10	Fixed (1.42)	scipy CPU	~12s	Good
v1.1.0	Fixed	Kornia GPU	~1.0s	Good
v1.1.0	Auto	Kornia GPU	~2.6s	Optimal

Net result: 5x faster with better results per image.

Correlation Validation

Kornia GPU vs scipy CPU background subtraction correlation: 0.9976 (nearly identical output).

---

Algorithm Details

Image → Pad → FFT → Detect Peaks → Create Mask → Inpaint → iFFT → Crop → Output

The algorithm:

1. Pads the image to reduce edge artifacts
2. Applies 2D FFT to get frequency domain
3. Detects lattice peaks via thresholding on log-power spectrum
4. Creates a mask protecting inner (low-freq) and outer (high-freq) regions
5. Inpaints peaks with local average amplitude from 4 shifted neighbors
6. Preserves original phase information
7. Inverse FFT returns to real space

Full Parameter Reference

Parameter	Default	Description
pixel_ang	required	Pixel size in Ångstroms
threshold	auto	Peak detection threshold - auto-optimized per image (range 1.40-1.60)
inside_radius_ang	90	Inner resolution limit (Å) - protects structural info
outside_radius_ang	auto	Outer resolution limit (Å) - protects near Nyquist
expand_pixel	10	Morphological expansion of peak mask (pixels)
unit_cell_ang	116	Crystal unit cell for inpaint shift calculation (Å)
backend	auto	"auto", "numpy" (CPU), or "pytorch" (GPU)
use_kornia	true	Use Kornia for GPU-accelerated background subtraction (48x faster)

Supported Hardware

• NVIDIA GPUs: RTX 20/30/40 series, A100, H100
• CUDA versions: 11.8, 12.1, 12.4, 12.6, 12.8
• CPU fallback: Works on any system (just slower)

Development Setup

git clone https://github.com/gsstephenson/cryoem-lattice-subtraction.git
cd cryoem-lattice-subtraction

conda create -n lattice_sub python=3.11 -y
conda activate lattice_sub

pip install -e ".[dev]"
pytest tests/ -v  # Run tests

Code Structure

src/lattice_subtraction/
├── __init__.py           # Package exports
├── cli.py                # Command-line interface
├── core.py               # LatticeSubtractor main class
├── batch.py              # Parallel batch processing
├── config.py             # Configuration dataclass
├── io.py                 # MRC file I/O
├── masks.py              # FFT mask generation
├── processing.py         # FFT helpers + GPU background subtraction
├── threshold_optimizer.py # Auto-threshold optimization (NEW in v1.1)
├── ui.py                 # Terminal UI
└── visualization.py      # Comparison figures

Migration from MATLAB

This is a modernized rewrite of legacy MATLAB code (LAsub.m, bg_push_by_rot.m, etc.) with:

• 10-100× faster processing via GPU
• Pip-installable package
• YAML config files instead of hardcoded values
• Command-line interface
• Automated testing

Convert a legacy PARAMETER file:

lattice-sub convert-config PARAMETER params.yaml