ISTDECO: In Situ Transcriptomics Decoding by Deconvolution

Axel Andersson, Ferran Diego, Fred A. Hamprecht, Carolina WählbybioRxiv 2021 · DOI

TL;DR

This paper introduces ISTDECO, a novel method that simultaneously decodes the location and type of mRNA molecules in tissue images. By combining spatial and spectral deconvolution, ISTDECO achieves state-of-the-art performance in challenging conditions like low signal-to-noise and high spot density, offering a more robust and efficient approach to in situ transcriptomics.

What Problem Does This Paper Address?

Understanding the spatial organization of cells within tissues is crucial for deciphering biological processes like development and disease. In situ transcriptomics (IST) techniques allow researchers to map gene expression directly within tissue samples. These methods typically generate a series of images where fluorescent spots, representing mRNA molecules, appear or disappear across different imaging rounds and color channels. This pattern acts as a "barcode" to identify the specific mRNA.

However, existing IST analysis pipelines usually separate the process into two distinct steps: first, locating the fluorescent spots, and then decoding their barcodes. This separation can lead to errors, especially in areas with weak signals or when many spots are clustered closely together (optical crowding). The authors argue that a unified approach, where localization and decoding happen simultaneously, can overcome these limitations and improve accuracy.

Methods

The core innovation of ISTDECO is its unified approach to decoding. Instead of separate localization and decoding steps, ISTDECO combines spectral and spatial deconvolution into a single algorithm. The method models the observed image data (Y) as a combination of the decoded barcode types (X), a known codebook (D) representing barcode patterns, a spatial blur kernel (G) modeling the microscope's point spread function, and background noise (b and ξ).

The observation model is expressed as: $Y = DXG + b + \xi$. The goal is to find the matrix X, which essentially represents a set of images where each channel corresponds to a specific barcode type, highlighting its location. ISTDECO uses a maximum likelihood estimation (MLE) approach, assuming Poisson-distributed noise, to solve for X. This involves minimizing a cost function that measures the difference between the observed data and the model's prediction, with the constraint that X must be non-negative. The paper derives an iterative update rule for X, which is efficient and can be implemented using vector operations suitable for GPU acceleration. For actual detection, a spatial non-maxima suppression is applied to the channels of X, and spots are selected based on intensity and a quality metric.

Datasets

The paper evaluates ISTDECO on two types of datasets:

  • Simulated Data: The authors generated synthetic datasets to rigorously test their method under controlled conditions. These simulations allowed them to vary parameters like signal intensity, spot density, and noise levels, enabling a systematic comparison with other methods.
  • Real IST Datasets:
    • MERFISH Dataset: A publicly available dataset from a MERFISH experiment. The authors compared ISTDECO's performance against the original decoding pipeline used for this dataset.
  • In Situ Sequencing (ISS) Dataset: Another publicly available dataset from an ISS experiment. Similar to MERFISH, ISTDECO was compared against the original decoding pipeline for this dataset.

The paper also mentions the inclusion of non-targeted barcodes in the codebooks for both real datasets to aid in false discovery rate (FDR) estimation and threshold tuning.

Key Insights and Ideas

  • Unified Decoding is Superior: The central insight is that jointly optimizing for spot localization and barcode decoding, rather than treating them as sequential steps, significantly improves robustness, especially in challenging imaging scenarios.
  • Deconvolution as a Powerful Framework: Applying deconvolution principles to the entire IST decoding problem provides a principled and mathematically grounded approach that can handle complex signal interactions.
  • Handling Noise and Crowding: ISTDECO's ability to maintain high precision and recall in low SNR and high spot density regions is a major advantage over methods that struggle with these conditions.
  • GPU Acceleration for Scalability: The implementation of ISTDECO using vector operations and its compatibility with GPUs make it computationally efficient, enabling analysis of large datasets.

Key Results

  • State-of-the-Art Performance: ISTDECO achieved state-of-the-art performance on both simulated and real datasets, outperforming existing methods in terms of accuracy (measured by F1 score).
  • Robustness to Noise and Density: The method demonstrated high precision and recall even with low signal-to-noise ratios and high spot densities, where traditional methods falter.
  • Strong Correlation with Reference Methods: On real datasets (MERFISH and ISS), ISTDECO showed a high correlation (Pearson's r ≈ 0.98) in the types and numbers of barcodes detected compared to established reference methods.
  • Potential for Novel Detections: Analysis of real data suggested that ISTDECO might detect barcodes missed by existing methods, potentially due to its more sensitive deconvolution approach.
  • Computational Efficiency: The GPU-accelerated implementation allows for rapid processing, with a 2048x2048 field of view processed in under 30 seconds on an 8GB laptop GPU.

Limitations and Future Work

  • Alignment Requirement: ISTDECO, like other IST methods, relies on well-aligned images across different rounds and channels. The authors suggest point-cloud based registration as an effective solution.
  • Hyperparameter Tuning: While the method is efficient, optimal performance still requires tuning of hyperparameters (e.g., spot quality thresholds, FDR levels), which was performed using grid search and hyperparameter optimization packages.
  • Future Work: The authors suggest that further speedups could be achieved by adopting down-sampling strategies. The potential for ISTDECO to identify true signals missed by other methods warrants further investigation.

Summary updated

Abstract

In Situ Transcriptomics (IST) is a set of image-based transcriptomics approaches that enables localisation of gene expression directly in tissue samples. IST techniques produce multiplexed image series in which fluorescent spots are either present or absent across imaging rounds and colour channels. A spot’s presence and absence form a type of barcoded pattern that labels a particular type of mRNA. Therefore, the expression of a gene can be determined by localising the fluorescent spots and decode the barcode that they form. Existing IST algorithms usually do this in two separate steps: spot localisation and barcode decoding. Although these algorithms are efficient, they are limited by strictly separating the localisation and decoding steps. This limitation becomes apparent in regions with low signal-to-noise ratio or high spot densities. We argue that an improved gene expression decoding can be obtained by combining these two steps into a single algorithm. This allows for an efficient decoding that is less sensitive to noise and optical crowding. We present IST Decoding by Deconvolution (ISTDECO), a principled decoding approach combining spectral and spatial deconvolution into a single algorithm. We evaluate ISTDECO on simulated data, as well as on two real IST datasets, and compare with state-of-the-art. ISTDECO achieves state-of-the-art performance despite high spot densities and low signal-to-noise ratios. It is easily implemented and runs efficiently using a GPU. ISTDECO implementation, datasets and demos are available online at: github.com/axanderssonuu/istdeco
BibTeX
@misc{andersson2021istdeco,
  author = {Axel Andersson and Ferran Diego and Fred A. Hamprecht and Carolina Wählby},
  title = {{ISTDECO: In Situ Transcriptomics Decoding by Deconvolution}},
  year = {2021},
  doi = {10.1101/2021.03.01.433040},
  howpublished = {bioRxiv},
}
Figure