FISH-Dist: An Automated Pipeline for 3D Genomic Spatial Distance Quantification in FISH Imaging

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

In this manuscript, Aigouy et al. developed a new bioinformatic pipeline for measuring short range genomic interaction distances using Oligo-FISH (FISH-Dist). Authors combine deep learning methods for segmentation, 3D Gaussian fitting for pixel localization as well as affine and linear chromatic aberration correction (ACC, LCC) approaches. However, throughout the paper, ACC and LCC approaches exhibited nearly identical results.

 

Using FISH-Dist, authors were able to demonstrate reduction in distance errors in control experiment using two distinct probes targeting the same control locus. Performance of the method was also validated using DNA origami nanorulers with defined lengths.

 

Afterward, authors went on to quantitate the relationship between target sequence length, composition and probe density with distance measurement resolution and sensitivity.

 

Overall, I believe this is a well written manuscript that will help the community for measuring short genomic distances (< 50 kb). The results and discussions are well written and include potential pitfalls for the approach, including sample to sample variability.

 

Here are some minor points that should be addressed:

 

1. The supplementary figures cannot be accessed. Please correct the link provided in the manuscript.

 

2. Reference is missing in the methods section for Affine Chromatic Correction: The transformation matrix is computed using an implementation adapted from a library by C. Gohlke [? ], which solves the resulting least-squares point-set alignment problem using a singular value decomposition (SVD)–based approach.

 

3. Line 252 is missing the supplementary figure number:

Analysis of residual chromatic displacements after affine correction revealed no obvious position-dependent variation (Figure S ??A,B)

 

4. Line 255 is missing the supplementary figure number:

in-plane orientation bias, whereas the axial (Z) component dominated (Figure S ??C).

 

5. While not mandatory, it would be helpful to include visual illustrations showing the designs of the constructs (R1, R6) with the spacer sequences.

Author Response

Comment 1: The supplementary figures cannot be accessed. Please correct the link provided in the manuscript.
Response 1: Thank you for noting the missing reference to the supplementary figure. We apologize for this oversight—the figure was uploaded separately but not properly linked in the manuscript. In the revised version, we have ensured that Supplementary Figure S1 is correctly included and referenced.

Comment 2: Reference is missing in the methods section for Affine Chromatic Correction: The transformation matrix is computed using an implementation adapted from a library by C. Gohlke [? ], which solves the resulting least-squares point-set alignment problem using a singular value decomposition (SVD)–based approach.
Response 2: We thank the reviewer for pointing out the missing reference in the 'Affine Chromatic Correction' subsection. This was due to a formatting error in our bibliography library, which has now been corrected. The reference to C. Gohlke’s library is properly included in the revised manuscript.

 

Comments 3&4: Line 252 is missing the supplementary figure number; Line 255 is missing the supplementary figure number

Response 3&4: we appologize for this see Response 1

Comment 5: While not mandatory, it would be helpful to include visual illustrations showing the designs of the constructs (R1, R6) with the spacer sequences.
Response 5: We have now included the complete sequences of all constructs and oligos in Supplementary File S1. This file contains the full nucleotide sequences for R1, R6, and all associated spacers, as well as the oligos used for targeting.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors investigated “FISH-Dist: An Automated Pipeline for Accurate 3D Genomic Spatial Distance Quantification in FISH Imaging.” The study is meaningful, and I do not have major concerns about the overall content; however, I have several comments.

 

The manuscript title positions the work as an “automated pipeline for accurate 3D genomic spatial distance quantification,” yet the main text and primary results are largely focused on  calibration and correction of optical chromatic aberration, validation of distance measurements using nanorulers and transgenes with defined spacers, and empirical evaluation of probe length and labeling density effects on measurement precision and detection yield. In contrast, for a pipeline- or software-centered manuscript, several essential elements are not presented in a sufficiently systematic manner, such as clear definitions of the inputs and outputs; an end-to-end workflow with justification for each key module and its necessity/novelty; details of the training data and the generalization limits of the deep-learning model; a quantitative assessment of how false positives/false negatives and cross-channel pairing errors bias the distance distributions; the applicability and failure modes of ACC and LCC under different imaging conditions; fair benchmarking against existing tools using the same dataset and evaluation metrics. Therefore, I suggest that the authors either substantially strengthen the methodological or software presentation and provide rigorous benchmarking to align with the current title, or revise the title and narrative structure to more accurately position the manuscript as an experimental methods paper.

 

Specific points:

 

Lines 69–71: Which version or refs of the Drosophila melanogaster genome was used?

 

Lines 75–80: Were R1/R6 adopted from prior ref or designed/validated by the authors?

 

Line 234: The 1,917 spot pairs, how many images, how many wings, and how many flies do these pairs come from? What is the statistical unit of independence used in the analysis?

 

Line 252: “(Figure S ??A,B)” is not specified; similarly, line 255 has the same issue. Please cite the exact supplementary figure panels.

Author Response

Comment 1: In contrast, for a pipeline- or software-centered manuscript, several essential elements are not presented in a sufficiently systematic manner, such as clear definitions of the inputs and outputs
Response 1: We have added a detailed description of the inputs and outputs in the software documentation at https://fish-dist.readthedocs.io/en/latest/inputs_outputs/ to avoid overloading the main text.

Comment 2: an end-to-end workflow with justification for each key module and its necessity/novelty

Response 2: We have added a detailed end-to-end workflow with justification for each key module and its rationale in the software documentation at https://fish-dist.readthedocs.io/en/latest/detailed_end_to_end_overview/

 

Comment 3: details of the training data and the generalization limits of the deep-learning model
Response 3: We will provide our training scripts and data upon request (Lines 395-396) and data avilability (Lines 461-462). We discuss the limits of the model (Lines 389-396)

 

Comment 4: a quantitative assessment of how false positives/false negatives 
Response 4:  In the context of DNA FISH imaging, false positives and false negatives can be defined as follows: false positives are spots detected by the model that do not correspond to true genomic targets—primarily arising from non-specific probe hybridization or background noise mimicking FISH signals—while false negatives are actual genomic targets that go undetected due to weak signal or segmentation limitations. In our pipeline, the majority of false positives are automatically excluded by restricting analysis to spots within segmented nuclei, as off-nuclear signals are biologically irrelevant and non-specific for DNA FISH. However, quantifying false positives and negatives in biological samples remains fundamentally challenging due to the absence of ground truth: real FISH experiments exhibit inherent variability in probe binding, chromatin accessibility, and signal intensity, making it impossible to definitively label every spot as true or false. While DNA origami nanorulers provide precise ground-truth distances for validating spatial accuracy and chromatic aberration correction, they do not address false positives/negatives in biological contexts, as they lack the complexity of chromatin organization and probe hybridization dynamics.

Synthetic datasets could theoretically enable quantification of false positives/negatives, but their interpretation is complicated. For example, a "missed" spot in synthetic data might not necessarily indicate a pipeline failure—it could simply reflect a realistic scenario where weak signals, such as those caused by low probe density or chromatin inaccessibility, fall below detection thresholds. This ambiguity makes it difficult to distinguish between true false negatives and biologically plausible weak signals. Instead, we address these challenges through practical strategies: nuclear masking eliminates most false positives, systematic probe optimization (Figures 4–5) minimizes false negatives by ensuring sufficient signal intensity, and the consistency between our two independent chromatic aberration correction methods (ACC and LCC) further validates the robustness of our distance measurements (Figures 2–3). Thus, while we cannot report absolute false-positive/negative rates, our pipeline robustly minimizes their impact through a combination of nuclear masking, probe design optimization, orthogonal validation with nanorulers, and cross-method consistency, offering a rigorous framework for accurate distance quantification in real-world FISH experiments.

 

Comment 5: and cross-channel pairing errors bias the distance distributions

Response 5: The potential for cross-channel pairing bias to affect distance measurements and chromatic aberration registration is a valid concern. However, our pipeline incorporates multiple layers of safeguards to mitigate this issue. First, distance measurements are reported using the median, a statistic inherently robust to outliers and skewed distributions. This ensures that occasional mismatched pairs—whether due to false positives or stochastic errors in probe localization—do not disproportionately bias the results.

Second, for affine chromatic aberration correction (ACC), we initially observed that outliers—particularly pairs with large distances—could dominate the registration calculation, resulting in inaccurate alignment. To mitigate this, we exclude the upper and lower 25% of paired spot distances, focusing the analysis on the central 50% of the distribution. This ensures that the affine transformation is derived from the most reliable and representative pairs, minimizing the impact of mismatched or aberrant pairings. By prioritizing this subset of higher-confidence pairs, the algorithm aligns the channels based on the most consistent and biologically relevant signals.

Additionally, the close agreement between ACC and linear chromatic correction (LCC)—two conceptually distinct methods—further validates that our pairing and correction strategies are not introducing systematic biases. If pairing errors were significant, we would expect discrepancies between the two correction approaches, particularly in the distribution of residual distances. The fact that both methods yield nearly identical results (Figures 2–3) strongly suggests that pairing bias is effectively controlled.

Finally, our use of nuclear masking ensures that only spots within biologically relevant regions are considered for pairing, reducing the likelihood of non-specific or artifactual signals influencing the results. While no method can entirely eliminate the theoretical risk of pairing bias, our combination of outlier-resistant statistics (e.g., median-based measurements), stringent pair filtering (excluding the upper and lower 25% of distances), cross-method validation (agreement between ACC and LCC), and biological context constraints (nuclear masking) provides a robust framework for accurate and unbiased distance quantification.

 

Comment 6: applicability and failure modes of ACC and LCC under different imaging conditions
Response 6: Both affine chromatic correction (ACC) and linear chromatic correction (LCC) apply global corrections across the entire image, making them inherently unable to address local variations in chromatic aberration or imaging artifacts. For instance, we have observed on one of our microscopes a sudden transient XY pixel shifts during acquisition—even within single frames—followed by realignment to the original position. Such localized instabilities fall outside our pipeline's correction capabilities and can significantly compromise distance measurements and registration accuracy, as both methods assume a stable, globally consistent aberration profile.

Both methods require large numbers of paired spots and are sensitive to pairing bias, though to different degrees. ACC, relying on 3D affine transformations, is particularly vulnerable to mismatched pairs or outliers, which is why we retain only the most reliable 50% of pairs for correction.

ACC requires a colocalization dataset to compute its affine transformation matrix, making it highly sensitive to temporal or session-to-session changes in chromatic aberration. If the aberration profile shifts (e.g., due to changes in the optical path), the precomputed transformation becomes invalid, leading to significant errors in distance measurements.

In contrast, LCC is computed internally and does not rely on a precomputed transformation matrix, making it more robust to such shifts. However, LCC still requires a colocalization dataset to establish the baseline resolution.

Despite their different approaches, both methods ultimately depend on high-quality colocalization data acquired under the same chromatic aberration conditions to ensure accurate correction, distance measurement, and resolution estimation. Importantly, close agreement between ACC and LCC results provides internal validation of the correction’s reliability. If the two methods yield divergent results, this likely indicates changes in chromatic aberration characteristics, necessitating acquisition of a new colocalization dataset for recalibration.

Comment 7: fair benchmarking against existing tools using the same dataset and evaluation metrics
Response 7: Direct benchmarking of FISH-Dist against existing tools is fundamentally challenging because most current pipelines (e.g., https://github.com/embl-cba/DNA-FISH) lack chromatic aberration correction—a critical component for accurate distance measurements, especially at the short genomic scales we investigate. As shown in Figure 2, chromatic aberration introduces systematic offsets that completely dominate measurements at these distances. Without correction for these optical artifacts, existing tools would inherently produce inaccurate results.

While synthetic datasets could in theory provide a controlled benchmarking environment, creating them presents substantial difficulties. Realistic FISH simulations would need to faithfully reproduce not only chromatic aberration patterns but also complex biological factors like probe hybridization dynamics, noise profiles, and chromatin accessibility variations. Moreover, performance on synthetic data—often devoid of the full spectrum of real-world noise—may not translate to biological samples. For example, methods optimized for noise-free synthetic conditions could become overly sensitive, mistakenly capturing noise as signal in real, noisy datasets like ours. Hence, unless synthetic datasets are carefully designed to include realistic noise and biological variability, strong performance on them would not necessarily guarantee robustness in experimental conditions.

 

Comment 8: Therefore, I suggest that the authors either substantially strengthen the methodological or software presentation and provide rigorous benchmarking to align with the current title, or revise the title and narrative structure to more accurately position the manuscript as an experimental methods paper.

Response  8: In line with the reviewer’s suggestion, we revised the manuscript to moderate claims of “accuracy” and “precision” and to better emphasize that FISH-Dist is a validated computational framework supporting experimental FISH workflows. Across the manuscript, we replaced terms like “precise” or “accurate” with “quantitative” or “reproducible,” clarified how validation was performed, and adjusted sentences and headings to reflect the methodological focus rather than absolute measurement claims.

 

Title Revision

Original: FISH-Dist: An Automated Pipeline for Accurate 3D Genomic Spatial Distance Quantification in FISH Imaging

Revised: FISH-Dist: A Computational Pipeline for 3D Genomic Distance Quantification in Confocal FISH Imaging

 

Abstract Revisions

Original: an automated computational pipeline for precise distance measurements

Revised: an automated computational pipeline for quantitative distance measurements

Original: We validated measurement accuracy using DNA origami nanorulers

Revised: We validated the pipeline by measuring the lengths of DNA origami nanorulers

Original:This enables robust quantification of spatial relationships in FISH datasets

Revised:This enables reproducible quantification of spatial relationships in 3D FISH datasets

 

Introduction Revisions

Original: Here we present FISH-Dist, an automated computational pipeline specifically designed for accurate 3D distance quantification in FISH imaging, with a focus on short-range genomic measurements.

Revised: Here we present FISH-Dist, an automated computational pipeline for quantitative 3D distance measurements in FISH imaging, with a focus on short-range genomic distances.

Original: By addressing technical limitations in standard confocal FISH, FISH-Dist enables robust and reproducible quantification of spatial relationships at scales most relevant to gene regulation.

Revised: By addressing technical limitations in standard confocal FISH, FISH-Dist provides a reproducible framework for quantifying spatial relationships at the short genomic distances most relevant to gene regulation.

 

Discussion Revisions

Original Subsection Title: FISH-Dist Enables Accurate Short-Range Genomic Distance Measurements

Revised Subsection Title: FISH-Dist Enables Reproducible 3D Measurements of Short-Range Genomic Distances

Original: optimized for precise 3D distance quantification

Revised: designed for quantitative 3D distance measurements

Original: accuracy validated

Revised: performance assessed

Original: More broadly, FISH-Dist provides a validated and automated framework for accurate 3D distance measurements in genome organization studies. By integrating chromatic aberration correction, robust spot detection, and quantitative distance analysis into a unified workflow, the pipeline lowers technical barriers to precise spatial measurements and enables reproducible, high-throughput analysis across diverse experimental systems.

Revised: More broadly, FISH-Dist provides a validated and automated framework for quantitative 3D distance measurements in genome organization studies. By integrating chromatic aberration correction, robust spot detection, and systematic distance analysis into a single workflow, the pipeline reduces technical barriers and enables reproducible, high-throughput spatial measurements across diverse experimental systems.

 

Comment 9: Lines 69–71: Which version or refs of the Drosophila melanogaster genome was used?
Response 9: The Drosophila melanogaster genome sequence used in this study corresponds to the FlyBase output based on genome Release 6 (R6). We have also updated the manuscript text to clarify this information (i.e. "genomic region69(X:6,760,094–6,770,369, Drosophila genome Release 6) adjacent to the attP18")

Comment 10: Lines 75–80: Were R1/R6 adopted from prior ref or designed/validated by the authors?

Response 10: The R1 and R6 synthetic reporter sequences were custom-designed by Daicel Arbor Biosciences specifically for this study as orthogonal, non-genomic sequences fully covered by 87 Oligopaint probes per construct (targeting both DNA strands) to maximize signal intensity and detection efficiency, while being completely absent from the Drosophila melanogaster genome to eliminate off-target hybridization. These de novo synthesized sequences serve as controlled, high-specificity targets for FISH experiments, with their design validated through colocalization experiments (Figure 2) and distance measurements (Figure 3), confirming their suitability for quantitative spatial analysis. Additionally, we now provide the sequences of R1 and R6 and of all the oligos used in this study in Supplementary File S1. Please also refer to Section 2.3 of the Materials and Methods for further details.

 

Comment 11: Line 234: The 1,917 spot pairs, how many images, how many wings, and how many flies do these pairs come from? What is the statistical unit of independence used in the analysis?
Response 11: The 1,917 spot pairs were derived from eight images of a single wing (one animal) to establish a calibration baseline. Reproducibility was confirmed using independent datasets from other animals, including acquisitions performed more than seven months apart (uncorrected median distance: 192.0 nm; ACC-corrected: 54.1 nm). The statistical unit is the individual spot pair. We have added nb of images/wings/animals to the figure legends where applicable.

Comment 12: Line 252: “(Figure S ??A,B)” is not specified; similarly, line 255 has the same issue. Please cite the exact supplementary figure panels.
Response 12: Thank you for noting the missing reference to the supplementary figure. We apologize for this oversight—the figure was uploaded separately but not properly linked in the manuscript. In the revised version, we have ensured that Supplementary Figure S1 is correctly included and referenced.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have addressed my concerns very well, and I do not have any further questions.