Suite2p: beyond 10,000 neurons with standard two-photon microscopy

Marius Pachitariu, Carsen Stringer, Mario Dipoppa, Sylvia Schröder, L. Federico Rossi, Henry Dalgleish, Matteo Carandini, Kenneth D. HarrisbioRxiv 2016 · DOI

TL;DR

This paper introduces Suite2p, a fast, accurate, and complete computational pipeline for processing two-photon calcium imaging data. It enables the routine detection and analysis of over 10,000 neurons simultaneously from standard microscopes, overcoming previous limitations in speed and interpretability.

What Problem Does This Paper Address?

Two-photon microscopy has revolutionized neuroscience by allowing researchers to image the activity of large populations of neurons. However, the computational pipelines used to process the resulting massive datasets have lagged behind, being slow, inefficient, and often producing results that are difficult to interpret or validate. This bottleneck prevents researchers from fully leveraging the power of modern microscopes to understand the complex dynamics of neural circuits. Specifically, existing methods struggle with:

  • Speed: Processing large datasets takes an impractically long time, often requiring specialized computing clusters.
  • Accuracy and Completeness: They miss a significant number of active neurons, especially those with lower fluorescence signals or sparse activity.
  • Interpretability: The processing steps can be opaque, making it hard to identify and correct errors or biases.
  • Neuropil Contamination: Signals from surrounding non-neuronal tissue (neuropil) can obscure true neuronal activity, and existing methods don't adequately address this.
  • Generalizability: Pipelines often fail to perform well across different cell types, brain regions, or imaging conditions.

Methods

Suite2p is an end-to-end pipeline designed to address these challenges through a modular approach, with each stage optimized for speed and accuracy:

  1. Image Registration: Corrects for brain movement during imaging using an enhanced phase correlation method with sub-pixel accuracy and optional non-rigid alignment. This is crucial for maintaining the spatial integrity of neuronal signals over time.
  2. ROI Detection: Employs a novel source-extraction algorithm that models neuronal activity and neuropil contamination. It distinguishes between cellular signals and diffuse neuropil signals, a key improvement over methods that treat neuropil as uniform noise.
  3. Quality Control: Features an automated classifier trained on human-curated data to distinguish true cells from artifacts. A user-friendly GUI allows for manual review and correction, which then retrains the classifier, creating a rapid and intuitive semi-automated process. A super-resolution technique is also incorporated to improve manual curation.
  4. Activity Extraction and Spike Deconvolution: Extracts neuropil-corrected calcium traces and optionally infers spike times. This stage uses a generative model that simultaneously estimates neuropil contamination coefficients and spike trains, providing a robust way to remove neuropil signals.

Datasets

The paper utilizes several datasets for evaluation:

  • Standard-size datasets: Obtained from mouse visual cortex using GCaMP6m, with 5-plane imaging and a 500µm field of view.
  • Large-scale datasets: Acquired from transgenic mice expressing GCaMP6s in excitatory neurons, involving 11 planes and a larger 900µm field of view, yielding over 1,000 cells per plane.
  • Simulated data: Used to test the image registration algorithm under controlled conditions with known translation and noise.
  • Neurofinder benchmark: A set of 9 test datasets from various labs, used to compare Suite2p's performance against other algorithms.

Key Insights and Ideas

  • Neuropil Modeling is Crucial: The paper strongly emphasizes that neuropil contamination is not random noise but a significant signal that needs to be explicitly modeled and accounted for, not just subtracted. This is key to distinguishing dimmer cells and improving ROI detection.
  • Modularity Enhances Flexibility: By breaking down the processing into independent stages, Suite2p allows users to adapt or replace individual components, making it more versatile than monolithic pipelines.
  • Speed and Scalability: The development of efficient algorithms, particularly the use of SVD for accelerating cell detection, allows Suite2p to handle datasets of unprecedented size (over 10,000 neurons) on standard workstations.
  • Semi-Automated Quality Control: The iterative retraining of the classifier based on human curation significantly speeds up the validation process, making it practical for large datasets.
  • Subcellular Analysis is Feasible: Suite2p's spatial constraints enable the detection and analysis of activity from smaller structures like dendritic spines and axonal boutons, opening new avenues for studying synaptic transmission and dendritic integration.

Key Results

  • Speed: Suite2p operates faster than real-time, with a single iteration of optimization taking ~7 seconds compared to ~8 minutes for CNMF on subsampled data.
  • Cell Yield: Suite2p recovers approximately twice as many neurons as previous state-of-the-art methods like CNMF. On large-scale datasets, it routinely detects over 10,000 neurons.
  • Accuracy: ROC analysis shows a reliably lower fraction of false-positive results with Suite2p compared to CNMF.
  • Computational Resources: Suite2p runs on standard workstations (e.g., $1,000 workstation with 16-64GB RAM and a mid-range GPU), unlike some methods requiring large compute clusters.
  • Subcellular Detection: Successfully detects activity from axonal boutons, dendrites, and spines, even when they are highly correlated.

Limitations and Future Work

  • Neuropil Contamination: While Suite2p significantly improves neuropil correction, the paper acknowledges that it's an ongoing challenge, especially in densely labeled preparations.
  • Manual Curation: Despite automation, manual curation is still required for optimal results, though Suite2p drastically reduces the time needed.
  • Generalizability: While designed to be general, performance might still vary across vastly different imaging conditions or brain regions, requiring potential parameter tuning.
  • Future Work: The paper implies that further refinement of the neuropil modeling and classifier customization will continue to improve performance. The ability to analyze subcellular structures suggests future work could focus on detailed synaptic and dendritic integration studies.

Summary updated

Abstract

Abstract Two-photon microscopy of calcium-dependent sensors has enabled unprecedented recordings from vast populations of neurons. While the sensors and microscopes have matured over several generations of development, computational methods to process the resulting movies remain inefficient and can give results that are hard to interpret. Here we introduce Suite2p: a fast, accurate and complete pipeline that registers raw movies, detects active cells, extracts their calcium traces and infers their spike times. Suite2p runs on standard workstations, operates faster than real time, and recovers ~2 times more cells than the previous state-of-the-art method. Its low computational load allows routine detection of ~10,000 cells simultaneously with standard two-photon resonant-scanning microscopes. Recordings at this scale promise to reveal the fine structure of activity in large populations of neurons or large populations of subcellular structures such as synaptic boutons.
BibTeX
@misc{pachitariu2016suite,
  author = {Marius Pachitariu and Carsen Stringer and Mario Dipoppa and Sylvia Schröder and L. Federico Rossi and Henry Dalgleish and Matteo Carandini and Kenneth D. Harris},
  title = {{Suite2p: beyond 10,000 neurons with standard two-photon microscopy}},
  year = {2016},
  doi = {10.1101/061507},
  howpublished = {bioRxiv},
}
Figure