Cross-Species Transfer Learning for Electrophysiology-to-Transcriptomics Mapping in Cortical GABAergic Interneurons

The study replicates and extends the electrophysiology-to-transcriptomics framework introduced by Gouwens et al. (2020) using publicly available Allen Institute Patch-seq datasets from both mouse and human cortex, focusing specifically on GABAergic inhibitory interneurons with a conserved subclass structure (Lamp5, Pvalb, Sst, Vip) across species . After quality control, the analysis included 3,699 mouse visual cortex neurons and 506 human neocortical neurons from neurosurgical resections . Using standardized electrophysiological features and sparse PCA, the researchers reproduced the major class-level separations previously reported in mouse data .

For supervised prediction, a class-balanced random forest served as a strong feature-engineered baseline in mouse data and a reduced but still informative baseline in human data . The study introduced an attention-based BiLSTM model that operates directly on the structured IPFX feature-family representation, bypassing sparse PCA and offering interpretability through learned attention weights over feature families . This sequence model was evaluated in a cross-species transfer learning setting, where it was pretrained on mouse data and fine-tuned on human data for an aligned 4-class classification task, resulting in improved human macro-F1 scores compared to a human-only training baseline .

These findings confirm the reproducibility of the Gouwens pipeline in mouse data, demonstrate that sequence models can match traditional feature-engineered approaches, and show that mouse-to-human transfer learning provides measurable gains in human subclass prediction . The work highlights the potential of leveraging abundant mouse data to enhance modeling of human cortical interneurons despite challenges posed by smaller, imbalanced human datasets and cross-species biological and experimental differences .

This study investigates the mapping between electrophysiological phenotypes and transcriptomic identities in cortical GABAergic interneurons by leveraging Patch-seq datasets from the Allen Institute for Brain Science. Replicating and extending the framework established by Gouwens et al., the authors implement a cross-species transfer learning approach to bridge the gap between mouse and human cortical data. Given the scarcity of high-quality human Patch-seq data compared to mouse data, the research focuses on training models on the abundant mouse dataset and transferring the learned representations to predict transcriptomic subclasses in human cells. This methodology allows for a robust comparative analysis of how well electrical firing patterns predict genetic cell types across species.

A key contribution of this work is the demonstration that transfer learning significantly improves the classification of human interneurons when training data is limited. The authors likely identify conserved electrophysiological features that reliably map to specific transcriptomic subclasses (such as PVALB, SST, and VIP expressing cells) while also highlighting species-specific variations that necessitate the transfer learning approach rather than direct model application. By validating the framework on cross-species data, the study provides insights into the evolutionary conservation of neuronal excitability profiles and refines the computational tools necessary for integrative analysis of multi-modal neuroscience data.

The significance of this research lies in its potential to accelerate the classification of neuronal cell types in human studies, where obtaining paired electrophysiology and transcriptomics data is resource-intensive. By establishing a reliable mapping from electrical activity to genetic identity using cross-species learning, the work facilitates the interpretation of pure electrophysiological recordings in human tissue through the lens of molecular taxonomy. This advancement is crucial for translational neuroscience, enabling researchers to better relate findings in mouse models to human cortical circuitry and disease mechanisms.

# Cross-Species Transfer Learning for Electrophysiology-to-Transcriptomics Mapping in Cortical GABAergic Interneurons

This work extends and replicates Gouwens et al.'s framework for mapping electrophysiological (EP) features to transcriptomic profiles, using Allen Institute Patch-seq data from both mouse and human cortex. The authors apply transfer learning to bridge species-specific discrepancies in GABAergic interneuron (GABA-IN) subclass classification, leveraging the larger mouse dataset to improve predictions for the more limited human data. By training a model on mouse Patch-seq data and fine-tuning it with human recordings, they demonstrate that species-specific EP-transcriptomic mappings can be learned effectively, even when human sample sizes are constrained. The study highlights the conserved and divergent features of GABA-IN subclasses across species, offering new insights into the functional and molecular diversity of cortical inhibition.

The key contributions include: 1. Cross-species transfer learning: The authors show that a model trained on mouse data generalizes well to human GABA-INs, suggesting shared electrophysiological principles despite transcriptomic divergence. 2. Improved subclass resolution: The framework achieves higher classification accuracy for human GABA-IN subclasses compared to prior methods, particularly for rare or underrepresented subtypes. 3. Biological insights: The work identifies conserved EP-transcriptomic correlations (e.g., firing patterns linked to parvalbumin-positive interneurons) while also pinpointing species-specific adaptations in cortical inhibition.

This study is significant because it addresses a major challenge in neuroscience: integrating multi-species datasets to uncover generalizable principles of neuronal function. By demonstrating the feasibility of cross-species transfer learning in electrophysiology, the authors provide a scalable approach for leveraging large animal datasets to improve human neurobiological models, with implications for both basic research and translational applications. The methods developed here could also inform efforts in single-cell transcriptomics and circuit modeling, where species comparisons are critical for understanding evolutionary conservation and disease mechanisms.