PAPER: INTRODUCTIONHigh-density electrode arrays now allow for simultaneous extracellular recording from hundreds to thousands of neurons across interconnected brain regions (Jun et al., 2017;Steinmetz et al., 2021;Ye et al., 2023b;Trautmann et al., 2023). While significant progress has been made in developing algorithms for tracking neural activity (Hennig et al., 2019;Buccino et al., 2020;Magland et al., 2020;Boussard et al., 2023;Pachitariu et al., 2024), identifying cell types and brain regions solely from electrophysiological features remains an open problem.Traditional approaches for electrophysiological cell-type classification utilize simple features of the extracellular action potential (EAP) such as its width or peak-to-trough amplitude (Mountcastle et al., 1969;Matthews & Lee, 1991;Nowak et al., 2003;Barthó et al., 2004;Vigneswaran et al., 2011) or features of neural activity, such as the inter-spike interval distribution (Latuske et al., 2015;Jouty et al., 2018). These simple features are interpretable and easy to visualize but lack discriminative power and robustness across different datasets (Weir et al., 2015;Gouwens et al., 2019). Current automated featurization methods for EAPs (Lee et al., 2021;Vishnubhotla et al., 2024) and neural activity (Schneider et al., 2023a) improve upon manual features but are limited to a single modality.There has been a recent push to develop multimodal methods that can integrate information from both recorded EAPs and spiking activity. PhysMAP (Lee et al., 2024) is a UMAP-based (McInnes et al., 2018a) approach that can predict cell-types using multiple physiological modalities through a weighted nearest neighbor graph. Another recently introduced method utilizes variational autoencoders (VAEs) to embed each physiological modality separately and then combines these embeddings before classification (Beau et al., 2025). Although both methods show promising results, PhysMAP is hard to fine-tune for downstream tasks as it is nondifferentiable, and the VAE-based method captures features that are important for reconstruction, not discrimination, impairing downstream performance (Guo et al., 2017). Neither approach has been used to classify brain regions.In this work, we introduce a multimodal contrastive learning method for neurophysiological data, Neuronal Embeddings via MultimOdal Contrastive Learning (NEMO), which utilizes large amounts of unlabeled paired data for pre-training and can be fine-tuned for different downstream tasks including cell-type and brain region classification. We utilize a recently developed contrastive learning framework (Radford et al., 2021) to jointly embed individual neurons' activity autocorrelations and average extracellular waveforms. The key assumption of our method is that jointly embedding different modalities into a shared latent space will capture shared information while discarding modalityspecific noise (Huang et al., 2024). We evaluate NEMO on cell-type classification using optotagged Neuropixels Ultra (NP Ultra) data from the mouse visual cortex (Ye et al., 2023b) and optotagged Neuropixels 1 data from the mouse cerebellum (Beau et al., 2025). We evaluate NEMO on brain region classification using the International Brain Laboratory (IBL) Brain-wide Map dataset (IBL et al., 2023). Across all datasets and tasks, NEMO outperforms current unsupervised (PhysMAP and VAEs) and supervised methods, with particularly strong performance in label-limited regimes. These results demonstrate that NEMO is a significant advance towards accurate cell-type and brain region classification from electrophysiological recordings.
================================================================================

REVIEW
--------------------------------------------------------------------------------
**Summary Of The Paper**

The paper introduces *NEMO* (Neuronal Embeddings via Multimodal Contrastive Learning), a pretraining framework designed for neurophysiological data that integrates extracellular action potentials (EAPs) and spiking activity autocorrelograms (ACGs) using a contrastive learning objective inspired by CLIP. The core idea is to jointly embed these two modalities into a shared latent space, thereby capturing commonalities while filtering out modality-specific noise. NEMO is evaluated on three datasets—NP Ultra (visual cortex), C4 (cerebellum), and IBL Brain-wide Map—for both cell-type and brain region classification tasks. The results indicate that NEMO outperforms existing methods, particularly in label-limited scenarios. Key components of the methodology include preprocessing steps to generate ACG images and EAP templates, data augmentation strategies tailored to each modality, and a contrastive loss function optimized for multimodal alignment.

---

**Strengths**

1. **Comprehensive Evaluation Across Tasks and Datasets:** The paper thoroughly evaluates NEMO on three diverse datasets (NP Ultra, C4, IBL) and applies it to both cell-type and brain region classification tasks. This breadth strengthens the claim of NEMO’s versatility and effectiveness across different neurophysiological challenges.

2. **Novel Application of Contrastive Learning to Neuroscience Data:** Leveraging a CLIP-like contrastive objective for neurophysiological data is a notable innovation. The paper provides a compelling argument for integrating ACGs and EAPs, which are traditionally treated separately.

3. **Robust Performance in Label-Limited Regimes:** The results highlight that NEMO performs exceptionally well even with limited labeled data—a crucial requirement in neuroscience where obtaining ground truth is challenging. This is supported by the label-ratio sweep experiments, where NEMO outperforms competitors with just 50% of available labels.

4. **Insightful Ablations and Analysis:** The paper includes ablation studies that explore the benefits of joint versus independent learning, the role of each modality, and the impact of different encoder architectures. These analyses offer meaningful insights into the mechanism behind NEMO’s performance.

5. **Practical Implementation Details:** The paper provides detailed descriptions of preprocessing pipelines, data augmentation techniques, and model configurations, enabling reproducibility despite some missing hyperparameter justifications.

---

**Weaknesses**

1. **Lack of Theoretical Justification for Joint Embedding Assumption (Severe):** The central assumption—that joint embedding of ACGs and EAPs reduces modality-specific noise—is stated without formal proof or empirical validation. No quantitative measure is provided to show that NEMO effectively suppresses noise compared to a baseline that processes modalities separately (e.g., PCA on concatenated features).

2. **Ambiguity in Encoder Architectures and Hyperparameters (Major):** The choice of shallow networks (2-layer CNN for ACGs, 2-layer MLP for EAPs) lacks justification. Given the complexity of ACG images (with 10 deciles), deeper architectures may be necessary to capture higher-order patterns. Furthermore, critical hyperparameters such as the temperature $ \tau $ in Equation 1 are fixed without explanation or sensitivity analysis, undermining the robustness of the results.

3. **Overstated Novelty Relative to Prior Work (Moderate):** The paper claims to be the first to apply contrastive learning to brain region classification, yet earlier works such as Azabou et al. (2021) and Urzay et al. (2023) have already explored contrastive learning on spiking activity. The distinction between these works and NEMO is unclear, and the novelty of applying CLIP-style objectives to brain region classification is not sufficiently established.

4. **Limited Statistical Rigor in Reporting Results (Moderate):** While the paper reports mean values and standard deviations, it omits statistical significance tests (e.g., t-tests, ANOVA) to determine whether observed improvements are reliable. This limits the interpretability of the results, especially in cases where margins are narrow (e.g., in the C4 dataset).

5. **Ambiguous Baseline Comparison for VAE (Minor):** The VAE baseline uses a 500D pre-latent representation, which is not part of the original design in Beau et al. (2025). This raises concerns about fairness in the comparison, as the 500D representation may not reflect the intended functionality of the VAE model.

---

**Questions For The Authors**

1. **Empirical Validation of Modality Alignment:** How is the semantic alignment between ACGs and EAPs for the same neuron validated? Please provide quantitative measures (e.g., correlation coefficients, mutual information) to support the claim that these modalities are meaningfully related in the shared latent space.

2. **Justification for Shallow Encoders:** Why are only 2 layers used for the ACG encoder and EAP encoder? Have deeper architectures been tested, and if so, how do they compare in terms of performance?

3. **Impact of Temperature Parameter $ \tau $:** Were sensitivity analyses conducted for the temperature parameter $ \tau $? If yes, please report the optimal value and justify its selection.

4. **Statistical Significance of Results:** Can the authors provide p-values or hypothesis tests confirming that the improvements in NEMO over baseline methods are statistically significant?

5. **Clarification on VAE Baseline:** Why is the 500D pre-latent representation used for the VAE baseline instead of the original 10D latent space? Does this affect the fairness of the comparison?

---

**Limitations Not Addressed By The Authors**

- **No Generalization Beyond Rodents:** The paper only evaluates NEMO on rodent data. The feasibility of extending NEMO to non-human primate or human data—which may exhibit different electrophysiological signatures—is not discussed.

- **Missing Reproducibility and Ethics Sections:** The submission format guidelines mention missing reproductions and ethics sections, which would be essential for assessing the practical deployment and ethical implications of NEMO.

- **Unaddressed Biological Constraints in Multimodal Fusion:** The paper does not discuss whether the fusion of ACGs and EAPs introduces any biologically implausible assumptions or misrepresentations of the underlying neural phenomena.

- **Scarcity of Ground Truth Labels in Multi-Region Classification:** While the IBL dataset contains region labels, the granularity and consistency of these labels across different labs remain unspecified. This affects the reliability of the brain region classification results.

---

**Soundness:** 3 (Good)

While the paper presents a solid implementation and shows promising results, the lack of theoretical grounding for the joint embedding assumption and ambiguity in hyperparameter choices reduce the overall soundness of the method.

**Contribution:** 3 (Good)

NEMO makes a reasonable contribution to the field by introducing a novel application of contrastive learning to neurophysiological data. However, the novelty is somewhat overstated, and the broader implications for neuroscience are not fully elaborated.

**Confidence:** 4 (Very High)

Despite certain shortcomings, the methodology is clearly articulated, and the empirical results are convincing. The paper offers a viable tool for improving classification in electrophysiology, albeit with caveats regarding theoretical foundations and reproducibility.

**Rating:** 7 (Accept)

The paper deserves acceptance due to its innovative approach and strong empirical results. However, it requires revisions to address theoretical underpinning, clarify hyperparameter choices, and strengthen statistical reporting.

**Brief Justification For Rating:**  
The paper presents a novel and practically useful method for neurophysiological classification using contrastive learning. Despite some weaknesses in theoretical justification and statistical rigor, the results are compelling and the methodology is well-articulated. With appropriate revisions, the paper would make a valuable addition to the literature on computational neuroscience and machine learning for electrophysiology.

================================================================================
