The openretina Project: Collaborative Retina Modelling Across Datasets and Species

Reviewer #1 (Public review):

Summary:

This "Tools and Resources" submission describes a platform for the modeling of stimulus-response relationships in the retina. It includes a repository for experimental data sets with standardized programmatic access, and a suite of software for constructing stimulus-response models and evaluating them.

Strengths:

(1) The paper is well written.

(2) The platform could serve an integrative function by connecting different research programs and offering a common baseline for evaluating stimulus-response models.

(3) The finding that there is "substantial explainable variance remains uncaptured by current models" is a useful insight to motivate further work and measure progress.

Weaknesses:

(1) The modeling supported by the package focuses on predictive accuracy at the cost of less interpretability.

(2) The article needs to make a stronger argument that this style of modeling is fruitful, especially when applied to the retina.

Main comments:

(1) Abstract machine learning vs mechanistic models. The "Core + Readout" architecture advocated here seems to be divorced from all the neurobiological detail that is already known in the retina. It mostly aims at prediction, not interpretation. Such a black-box modeling framework is useful in brain regions where we know very little about connectivity, or mechanisms, or even about the primary function being performed there, like in the mammalian cortex. In those cases, any model that can deliver a prediction is a step forward, even if it does not connect to biological mechanisms. But that's decidedly not the situation in the retina, where so many mechanistic details are known: from consensus cell types, to synaptic detail, to single-neuron biophysics, to circuit motifs. How can one connect this ML modeling approach with the extensive mechanistic knowledge available in retinal neuroscience? And can the combination somehow lead to a better understanding? The authors seem to recognize this tension (e.g. line 215ff and 370ff) but don't give it much weight. A stronger case needs to be made here for how this kind of modeling will advance the field.

(2) The "gradient field" approach. Figure 4c illustrates a case of this dissonance. The gradient field of the response increases with contrast in multiple directions. This is obvious a priori (see line 274) from the more mechanistic model we already have of this On-Off cell. These are the W3 cells described in www.pnas.org/cgi/doi/10.1073/pnas.1211547109. The circuit-based model from that paper, with rectifying on and off subunits from bipolar cells, gives a much more compact explanation for what the neuron does. Because each of the subunits has a spatio-temporal receptive field, this model can predict the entire dynamics to arbitrary stimuli, rather than just 2 dimensions of static stimuli as in the present analysis. So what is the value added here? Again, a stronger case needs to be made that these "Core + Readout" modeling activities enhance understanding.

(3) The "most exciting input" approach (Line 193ff):

- Presumably, some power constraint must be put on the stimulus? Otherwise, increasing the contrast will make it more exciting. What are these constraints?

- Presumably, this optimal stimulus is computed from the model based on non-optimal stimuli? What are the assumptions going into that?

- The most exciting stimulus is not necessarily the most useful characterization. Near its maximal firing rate, the neuron doesn't discriminate stimuli much, because the slope there is zero (line 237). Instead (or in addition), one would like to know along which stimulus axis the neuron is most sensitive. See e.g. discussion in Dayan & Abbott 2000, Figure 3.11.

Reviewer #2 (Public review):

Summary

openretina is a Python package for training and applying convolutional neural network-based models of retinal ganglion cell responses. The package integrates dataloading, model training, and evaluation in a unified framework built on PyTorch Lightning and Hydra, and ships with pre-trained model checkpoints and publicly available datasets (whitenoise, natural scenes) spanning multiple species (marmoset, mouse, axolotl, salamander) and recording modalities (multielectrode array recordings or 2-p calcium imaging). Beyond predictive modelling, openretina includes a suite of in silico analysis tools for probing learned representations, including maximally exciting input synthesis, discriminatory stimulus optimisation, and model weight visualisation. The broader openretina initiative aims to establish a community-driven platform for computational retina research, lowering barriers to entry and facilitating cross-dataset model benchmarking. This is a valuable contribution given the longstanding fragmentation of datasets, codebases, and analysis practices across retina laboratories.

Strengths:

The tool has several strengths. By providing a framework built on deep learning infrastructure, the package substantially lowers the barrier to entry for researchers without extensive machine learning backgrounds. The inclusion of pre-trained model checkpoints across multiple species and recording modalities will allow users to apply state-of-the-art models. The in silico toolkit - and in particular the MEI synthesis pipeline - has already demonstrated its scientific potential, with prior work using optimised stimuli to discover a previously uncharacterised RGC type confirmed experimentally, illustrating what becomes possible when these tools are made broadly accessible. The current modular Core + Readout architecture is a well-suited architecture for modeling retina responses. The HDF5-based data standard provides a sensible common format for contributing new datasets. Overall, the initiative is well-motivated, the engineering is competent, and the vision of a collaborative, community-driven platform for retina modelling is one that the retina community would benefit from.

Weaknesses:

(1) The in silico tools provided are valuable, but users should interpret their outputs in light of the performance of the underlying models. The predictive performances of current models and datasets in the package are far from performance ceilings.

(2) The authors appropriately note that optimised stimuli reveal what a neuron responds to but not how the computation is implemented. I would encourage readers to keep this distinction in mind when using the weight visualization tools as well - convolutional filters in a shared, unconstrained core do not map onto retinal circuit elements, and should be treated as model descriptors rather than circuit proxies. For example, RGCs of the same type may appear to sample inputs from two different filters, which should have been a single filter. Or a single RGC may be sampling from two filters, which under more constrained conditions could be approximated with a single filter. These are degeneracies in the CNN modeling framework that should be kept in mind when drawing circuit-level interpretations.

(3) The datasets currently distributed with the package vary in recording quality, and users should be aware that model performance may not only reflect architectural limitations but may also be limited by noise and data artifacts, including spike sorting errors.

(4) As the platform grows and community-contributed datasets are added, explicit data quality standards will be essential. I encourage the authors to develop dataset standards to ensure that their resource provides access to highly curated datasets, which I believe is an important step in having high-fidelity models whose functional interpretations can be trusted.

Reviewer #3 (Public review):

Summary:

This manuscript presents openretina, a Python-based platform designed to facilitate collaborative retinal modeling across datasets, laboratories, species, and recording modalities. The package provides standardized model architectures, evaluation metrics, and analysis tools, while also integrating several publicly available retinal datasets. The authors further demonstrate the platform through examples of in silico analyses and model benchmarking.

Strengths:

(1) Emphasis on standardization and reproducibility. Retinal modeling has become increasingly dependent on deep learning approaches, yet datasets and evaluation procedures remain fragmented across laboratories. By providing a unified framework, the authors lower barriers to entry and create opportunities for more systematic comparisons of models and datasets.

(2) The manuscript is clearly written, and the examples effectively illustrate the range of analyses supported by the platform.

(3) The benchmarking results are useful, particularly because they reveal substantial remaining gaps between current model performance and explainable variance ceilings.

Weaknesses:

Not a weakness per se, but rather a limitation, is that the manuscript focuses on software infrastructure rather than new biological or computational insights. While this is appropriate for a resource paper, some of the scientific examples, such as the gradient-field analysis of ON-OFF cells, function more as demonstrations than as rigorous validations of novel hypotheses. It might be useful to add a few sentences discussing potential scientific projects that can be immediately facilitated by the openretina (the current text in the Discussion focuses more on advancements in the technical/social aspects of science that will be supported by openretina).

Overall, this is a valuable and timely resource that is likely to benefit the retinal and computational neuroscience communities.

Author response:

We thank the editors and reviewers for their thoughtful assessment of our manuscript, and for recognizing openretina as a valuable and timely resource for the retinal modelling community.

We are especially glad that the reviewers appreciated the motivation of the project, the focus on standardization and reproducibility, and the potential of the platform to support systematic benchmarking and community-driven model development.

We also understand the concerns raised. In the revision of the manuscript, we will strengthen the conceptual discussion of how predictive models, including the current “Core + Readout” models, can contribute to retinal neuroscience alongside more mechanistic and circuit-based approaches. This is a central matter for us, and one that some of us have recently addressed in a broader review on current trends in retina modelling (see https://doi.org/10.1016/j.visres.2026.108854). We will draw on this perspective to better articulate when predictive models are useful, where their limitations lie, and how openretina can provide infrastructure for comparing functional, normative and mechanistic models within a shared framework.

We will also clarify the scope and limitations of the in-silico analysis methods provided within openretina. This will include a more explicit discussion of how MEIs, gradient-field analyses, and model-weight visualisations should be interpreted.

Furthermore, we will add more information that will help the reader better judge different aspects of dataset quality, including, for example, spike-sorting or calcium-processing information and explainable-variance distributions. We note, however, that there are many subtle details about experimental workflows that are difficult to capture in compact indicators. In addition, we will make it clearer that the manuscript represents a snapshot of a living resource: The website, dataset cards, documentation, and repository will be the primary source of this information, especially as new datasets are contributed.

Finally, we will of course address the technical clarifications raised by the reviewers, with the aim of making the manuscript more accessible overall.

We are grateful for the reviewers’ constructive comments and believe that addressing these points will make our presentation of openretina clearer and more useful to the community.