Review Reports - Machine Learning in Single-Molecule Tracking Analysis of Superresolution Optical Microscopy Data

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

General comments

For a general readership, the conceptual structure and rationale of the review by Saaavedra and Barrantes remain unclear. In particular, the manuscript introduces single-molecule localization microscopy (SMLM; e.g. PALM, STORM) early on in close connection with tracking and molecular dynamics, without sufficiently clarifying that classical SMLM is predominantly applied to fixed samples.

This is potentially misleading, as SMLM is inherently difficult to apply to dynamic systems: resolving spatial structures requires acquisition times that are typically much longer than the timescales of molecular motion. This fundamental limitation is not clearly stated and should be made explicit to properly frame the subsequent discussion.

Closely related, the review frequently mixes SMLM with single-molecule localization and tracking. High-precision localization of single emitters substantially predates SMLM, and for many readers, single-molecule tracking would not be considered an SMLM approach per se. A clearer conceptual distinction between structural SMLM imaging and single-molecule localization/tracking would improve clarity and accessibility.

While the literature coverage is extensive, the review misses recent conceptual frameworks that explicitly and rigorously link localization and tracking, such as the statistical approaches developed by Steve Presse and coworkers (e.g. Nature Methods, 2024, https://www.nature.com/articles/s41592-024-02349-9). Including such work would strengthen the discussion beyond heuristic localization-linking strategies.

Finally, although machine-learning–based approaches are discussed at length, it remains unclear which specific limitations of classical localization or tracking methods these approaches are intended to address, and in which regimes they offer clear advantages over non-ML algorithms. A more explicit comparison would help readers assess the added value of ML-based methods.

Author Response

Comments and Suggestions for Authors:
General comments

For a general readership, the conceptual structure and rationale of the review by Saaavedra and Barrantes remain unclear. In particular, the manuscript introduces single-molecule localization microscopy (SMLM; e.g. PALM, STORM) early on in close connection with tracking and molecular dynamics, without sufficiently clarifying that classical SMLM is predominantly applied to fixed samples.

Reply: The rationale of the review is outlined in the last paragraph of the Introduction: Machine learning in the analysis of single-molecule tracking. The reviewer is correct in stating that the text does not address the chronology of SMLM techniques in a historical perspective, describing their initial application to fixed (biological) samples or inorganic matter with the sole purpose of obtaining structural information. We have now amended this deficiency and elaborated on this topic, despite the fact that the initial version of the manuscript outlined them as beyond the scope of the review.

This is potentially misleading, as SMLM is inherently difficult to apply to dynamic systems: resolving spatial structures requires acquisition times that are typically much longer than the timescales of molecular motion. This fundamental limitation is not clearly stated and should be made explicit to properly frame the subsequent discussion.
Closely related, the review frequently mixes SMLM with single-molecule localization and tracking. High-precision localization of single emitters substantially predates SMLM, and for many readers, single-molecule tracking would not be considered an SMLM approach per se. A clearer conceptual distinction between structural SMLM imaging and single-molecule localization/tracking would improve clarity and accessibility.

Reply: The reviewer’s statement is historically correct, but the distinction is currently no longer valid. Current approaches to single-molecule localization and tracking require, and achieve, high-precision (sub-nanometer) localization for tracking individual molecules with high time- (microsecond) and spatial- (below the nm) windows, making this distinction conceptually superseded. A paragraph is now included in the revised version to discuss this issue. In addition, that SMLM is “difficult to apply to dynamic systems” is not a valid critique. Studying the dynamics of molecules in living cells is faced with multiple technical difficulties, as reflected in the relatively scarce number of publications on the subject compared to purely “static” observations on fixed specimens, but these challenges do not preclude their application to live cells under restricted experimental conditions. Our laboratory has applied camera-based STORM to live cells for short periods and obtained dynamic data (Mosqueira et al., 2018, 2020) that has been recently confirmed by MINFLUX microscopy (Reina et al., 2025).

While the literature coverage is extensive, the review misses recent conceptual frameworks that explicitly and rigorously link localization and tracking, such as the statistical approaches developed by Steve Presse and coworkers (e.g. Nature Methods, 2024, https://www.nature.com/articles/s41592-024-02349-9). Including such work would strengthen the discussion beyond heuristic localization-linking strategies.

Reply: The paragraph in p. 4 of the Introduction alluding to SMLM explicitly directs the reader to relevant literature sources (protocols, reviews) on the foundations and implementation of the various superresolution techniques. Delving into each of these techniques is definitely beyond the scope of the review.
Single-molecule tracking is based on sequential single-molecule localizations as a function of time. Following the reviewer’s concern about the omission of the reference Presse et al. (BNP-Track), we have now included this citation together with other relevant works whose methods are not data-driven. See Tables 1 and 2 in the revised version.

Finally, although machine-learning–based approaches are discussed at length, it remains unclear which specific limitations of classical localization or tracking methods these approaches are intended to address, and in which regimes they offer clear advantages over non-ML algorithms. A more explicit comparison would help readers assess the added value of ML-based methods.

Reply: We thank the reviewer for the suggestion. Two tables have been included in the revised version of the manuscript. The tables compare the relative weaknesses and strengths of the methods.
In the case of particle localization and trajectory linking, ML-based methods address time inference, user biases, and parameter selection that traditional methods barely address and/or require additional procedures to fine-tune parameters. In the case of single-molecule trajectory characterization, ML methods address the complexity of traditional approaches (which are primarily algorithmic). We clarify this issue in the introduction of each section.

Reviewer 2 Report

Comments and Suggestions for Authors

This review provides a timely and comprehensive overview of the application of machine learning (ML) and deep learning (DL) in single-molecule tracking (SMT) for membrane proteins. The authors cover a wide range of topics, from localization to trajectory analysis. However, while the technical descriptions are detailed, the manuscript lacks critical comparative analysis and, more importantly, fails to provide rigorous mechanistic evidence for the causal relationships proposed between specific factors and cellular phenotypes. Minor revisions are required before this manuscript can be considered for publication.

1.Lack of Comparative Summary Tables

The manuscript describes numerous ML architectures (CNNs, GNNs, Transformers) but lacks a systematic comparison. I suggest adding a table comparing these methods in terms of input requirements, computational cost, suitability for high-density tracking, and biological applicability. This is essential for a high-quality review in Cells.

2.Risk of Overfitting and Confounding Variables

In drawing causal inferences, the authors appear to underestimate the risk that the machine learning models may capture confounding variables or statistical noise rather than true biological drivers. For example, changes in membrane potential or variations in lipid composition may simultaneously influence the observed cellular phenotype, thereby confounding the model outputs. Therefore, a more critical and systematic analysis of how such latent variables might interfere with model interpretation and undermine the reliability of causal inference is necessary to enhance the rigor and credibility of the study’s conclusions.

3.Oversight of Biological Heterogeneity and Generalizability

The dynamics of membrane proteins are highly context-dependent and may vary substantially across different cell lines or under distinct physiological stress conditions. Therefore, the manuscript requires a more balanced discussion addressing whether the identified correlations remain valid across diverse biological contexts, rather than being presented as a single, universally applicable model.

4.Deficiency in Technical Specificity and Robustness Analysis

When discussing the roles predicted by the machine learning model, the manuscript does not provide quantitative assessments of uncertainty or statistical confidence. To ensure the robustness of the biological conclusions, it is necessary for the authors to include additional analyses of the model’s performance under non-ideal conditions.

Overall, the study is generally sound in terms of experimental execution; however, it remains clearly insufficient in mechanistic depth and the completeness of characterization. Substantial additional key evidence and strengthened theoretical and mechanistic interpretation are required.

Author Response

Comments and Suggestions for Authors:

This review provides a timely and comprehensive overview of the application of machine learning (ML) and deep learning (DL) in single-molecule tracking (SMT) for membrane proteins. The authors cover a wide range of topics, from localization to trajectory analysis. However, while the technical descriptions are detailed, the manuscript lacks critical comparative analysis and, more importantly, fails to provide rigorous mechanistic evidence for the causal relationships proposed between specific factors and cellular phenotypes. Minor revisions are required before this manuscript can be considered for publication.

1.Lack of Comparative Summary Tables
The manuscript describes numerous ML architectures (CNNs, GNNs, Transformers) but lacks a systematic comparison. I suggest adding a table comparing these methods in terms of input requirements, computational cost, suitability for high-density tracking, and biological applicability. This is essential for a high-quality review in Cells.

Reply: We thank the reviewer for the suggestion. The revised version includes two tables comparing methods for single-molecule localization, trajectory linking, and trajectory analysis. Each table includes the columns ‘Method’, ‘Input’, ‘Strengths’, and ‘Limitations’. Whenever possible, the tables indicate the suitability of each method for high-density conditions (see columns ‘Strengths’ and ‘Limitations’).
Including computational costs in terms of time of inference is not feasible since different authors carry out the tasks using different computer systems, making comparison potentially misleading. Instead, we discuss computational requirements in terms of memory usage and training requirements.
Regarding biological applicability, the papers discussed in the review validate their methods using different experimental criteria. The majority of the methods employed are applicable to the characterization of molecules present in cell membranes using superresolution microscopy. We clarify this in the Introduction. Re-stating this in a table would be redundant.

2.Risk of Overfitting and Confounding Variables
In drawing causal inferences, the authors appear to underestimate the risk that the machine learning models may capture confounding variables or statistical noise rather than true biological drivers. For example, changes in membrane potential or variations in lipid composition may simultaneously influence the observed cellular phenotype, thereby confounding the model outputs. Therefore, a more critical and systematic analysis of how such latent variables might interfere with model interpretation and undermine the reliability of causal inference is necessary to enhance the rigor and credibility of the study’s conclusions.

Reply: We thank the reviewer for the suggestion. The manuscript is a critical review of recent methodological developments in ML as applied to single-molecule tracking. We emphasize that most ML approaches in SMT are trained on simulated trajectories based on stochastic models (e.g., fBM and CTRW) as applied to molecules occurring the plasma membrane (Krapf, 2015). Hence, biological interpretability is limited by the selected simulation procedures. Machine learning models are not intended for causal inference per se. Moreover, researchers need to combine results from several methods (ML or not) to reach conclusions and find causes.
However, when FL-based methods are used, not all engineered variables should be included during analysis. Those variables that are "noisy" may provoke overfitting. Hence, a careful selection of parameters should be carried out to prevent poor generalization. We mention this in the section "FL or DL?" and, regarding reviewer's comment on confounding variables, we add a brief comment about how to interpret models’ results.
Krapf, D. (2015). Mechanisms underlying anomalous diffusion in the plasma membrane. Current Topics in Membranes, 75, 167–207. https://doi.org/10.1016/bs.ctm.2015.03.002

3.Oversight of Biological Heterogeneity and Generalizability
The dynamics of membrane proteins are highly context-dependent and may vary substantially across different cell lines or under distinct physiological stress conditions. Therefore, the manuscript requires a more balanced discussion addressing whether the identified correlations remain valid across diverse biological contexts, rather than being presented as a single, universally applicable model.

Reply: We thank the reviewer for the comment. Unfortunately, models are not directly applicable to any context. Depending on the application, models need fine-tuning, which is easily achieved by modifying the parameters used to synthesize simulation data and performing the training again. We mention this in the Conclusions section.

4.Deficiency in Technical Specificity and Robustness Analysis
When discussing the roles predicted by the machine learning model, the manuscript does not provide quantitative assessments of uncertainty or statistical confidence. To ensure the robustness of the biological conclusions, it is necessary for the authors to include additional analyses of the model’s performance under non-ideal conditions.

Reply: We thank the reviewer for the comment. Because this is review paper, we can only write statistical errors reported in papers and it is not appropriate to write new results. Because the number of samples used for validation in ML is, from a statistical point of view, high (n>1,000), statistical confidence of ML errors is not usually reported.
We decided to include the most relevant errors reported in the papers discussed. The section about ML methods for trajectory characterization already reports performance metrics. However, we noticed that trajectory linking and particle localizations sections seldom include information about performance in high-density conditions, a scenario where localizing particles is rather difficult due to PSF overlap. In the current version of the ms such sections include performance metrics reported in the referenced studies, particularly for high-density conditions.

Reviewer 3 Report

Comments and Suggestions for Authors

This manuscript presents a comprehensive and timely review of the application of machine learning (ML) techniques to the analysis of single-molecule tracking (SMT) data obtained via super-resolution optical microscopy. The authors provide a detailed and well-structured overview, covering key areas such as particle localization, trajectory linking, and the characterization of molecular dynamics using both feature-based learning and deep learning approaches. The integration of findings from important community challenges like the Andi Challenge adds significant value and context. They bridges the gap between advanced computational methods and experimental biophysics. The topic is of high relevance to the fields of biophysics, nanotechnology, and computational microscopy. The review is clearly written and, in principle, suitable for publication. However, several issues need to be addressed to improve clarity and consistency before the manuscript can be considered for publication.

(1) The formatting of the bibliographic entries in the References section is inconsistent. For example, the journal name appears as "Nature Methods" in some entries (e.g., Ref. 10) and as "Nat. Methods" in others (e.g., Ref. 22). I recommend that the authors adopt a single, uniform citation style throughout the reference list.

(2) The authors use the terms "single-molecule tracking (SMT)" and "single-particle tracking (SPT)" interchangeably throughout the text, even listing both as separate abbreviations. For clarity and consistency, I suggest choosing one primary term (e.g., SMT) for the main narrative and using it uniformly, while clarifying the relationship or equivalence to SPT in a footnote or upon first use.

(3) In the Introduction, the authors cite a large block of foundational super-resolution microscopy references "[9-13]". To improve the flow and provide better context for readers less familiar with the technical evolution, I suggest the authors disperse some of these citations throughout the introductory paragraphs, pairing them with brief explanations of the specific methodological advances each reference represents.

(4) The authors should briefly discuss the impact of external environments (non-Markovian or Markovian) on their results as the review primarily considers closed systems PRA 101, 013826 (2020); PRA 98, 023856 (2018); PRL 103, 210401 (2009)

(5) The authors should briefly outline how the ML-based analysis of SMT data, particularly for complex dynamics like state transition detection, can be realized and validated in current experimental configurations using techniques like MINFLUX or DNA-PAINT, supported by relevant literature.

Author Response

Comments and Suggestions for Authors:
This manuscript presents a comprehensive and timely review of the application of machine learning (ML) techniques to the analysis of single-molecule tracking (SMT) data obtained via super-resolution optical microscopy. The authors provide a detailed and well-structured overview, covering key areas such as particle localization, trajectory linking, and the characterization of molecular dynamics using both feature-based learning and deep learning approaches. The integration of findings from important community challenges like the Andi Challenge adds significant value and context. They bridges the gap between advanced computational methods and experimental biophysics. The topic is of high relevance to the fields of biophysics, nanotechnology, and computational microscopy. The review is clearly written and, in principle, suitable for publication. However, several issues need to be addressed to improve clarity and consistency before the manuscript can be considered for publication.

(1) The formatting of the bibliographic entries in the References section is inconsistent. For example, the journal name appears as "Nature Methods" in some entries (e.g., Ref. 10) and as "Nat. Methods" in others (e.g., Ref. 22). I recommend that the authors adopt a single, uniform citation style throughout the reference list.

Reply: We thank the reviewer for the suggestion. “Nature Methods” is adopted to unify style.

Reply: Our mistake. "single-molecule tracking (SMT) has now been uniformly adopted throughout.

Reply: The suggestion is well taken and the revised manuscript clarifies this point.

Reply: We thank the reviewer for the suggestion. We consider that discussing non-Markovian vs Markovian processes from quantum mechanics deviates from the central topic of the review, that is, ML applied to single-molecule tracking. Having said this, the distinction between Markovian and non-Markovian processes from a statistical point of view is relevant. We now indicate that many ML approaches are not restricted to the Markov assumption. In contrast to Hidden Markov Models, which rely solely on the current state to determine future transitions, most deep learning methods can incorporate both short- and long-term temporal correlations. By leveraging information from the entire trajectory, these approaches can capture complex dependencies, enabling more accurate inference.

Reply: We thank the reviewer for the comment.

ML methods are first validated with simulations. Then, these methods are applied to experimental data and obtained results are compared with those obtained in previous work. Experimental validation is developed such that the difference between previous and new results is quantified. Another approach is to recreate certain particle behaviors (e.g., free diffusion) and use them as a baseline for validation. However, other diffusion modes may be complicated to recreate intentionally (e.g., confinement). We added a paragraph discussing this in the conclusions.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors addressed all my concerns and suggestions. I hence recommend publication as it is.