Most discriminative stimuli for functional cell type clustering

TL;DR

The paper tackles the challenge of unbiased functional cell type identification in the retina and cortex, where domain knowledge biases traditional fingerprinting approaches. It introduces an EM-like clustering method that jointly learns Most Discriminative Stimuli (MDS) for functional clusters using deep predictive twins, enabling fast, on-the-fly cell typing across species and recording modalities. The approach yields seven RGC functional clusters in mouse retina, four in marmoset retina, and twelve in macaque V4, with MDS providing interpretable fingerprints that distinguish cell types and achieve speedups over conventional fingerprinting. This framework offers a practical tool for rapid experimental targeting and could accelerate discovery of functional cell types in less-characterized brain areas.

Abstract

Identifying cell types and understanding their functional properties is crucial for unraveling the mechanisms underlying perception and cognition. In the retina, functional types can be identified by carefully selected stimuli, but this requires expert domain knowledge and biases the procedure towards previously known cell types. In the visual cortex, it is still unknown what functional types exist and how to identify them. Thus, for unbiased identification of the functional cell types in retina and visual cortex, new approaches are needed. Here we propose an optimization-based clustering approach using deep predictive models to obtain functional clusters of neurons using Most Discriminative Stimuli (MDS). Our approach alternates between stimulus optimization with cluster reassignment akin to an expectation-maximization algorithm. The algorithm recovers functional clusters in mouse retina, marmoset retina and macaque visual area V4. This demonstrates that our approach can successfully find discriminative stimuli across species, stages of the visual system and recording techniques. The resulting most discriminative stimuli can be used to assign functional cell types fast and on the fly, without the need to train complex predictive models or show a large natural scene dataset, paving the way for experiments that were previously limited by experimental time. Crucially, MDS are interpretable: they visualize the distinctive stimulus patterns that most unambiguously identify a specific type of neuron.

Figures (14)

• Figure 1: Most discriminative stimulus (MDS) clustering based on a digital twin.A. Digital twin model trained to mirror responses of neurons in the visual system. B. MDS clustering (\ref{['alg:clustering']}) iterates between optimizing MDS to drive neurons within one cluster while suppressing all others and reassigning neurons to the cluster associated with the MDS they respond most to.
• Figure 2: Most discriminative stimuli clustering on mouse RGCs. Results on held-out test neurons and empirical comparison to baden_functional_2016 types shown. A. Cluster averaged digital twin response to the optimized MDS. Elements within a column normalized to highest value. B. MDS video snippets per cluster decomposed into time, UV- and green-channel stimulus components by singular value decomposition. For display, spatial components were re-scaled to a symmetric color scale between -1 (blue) and 1 (red). C. Summary of the functional characteristics for each cluster (UV-pref: UV-preference, T-freq. pref.: Temporal frequency preference; Surr.: Surround properties; RF: Receptive Field; $+ \text{/} \circ \text{/} -$: High/medium/low). D. Confusion matrix between MDS clusters and baden_functional_2016 types (predicted by a classifier for the mouse RGC dataset; qiu_efficient_2023). Annotations in same color belong to the same hierarchical functional type. Elements within rows were normalized to sum to 1, annotated numbers display the number of neurons in percent. E. Confusion matrix of the classifier qiu_efficient_2023 used to predict baden_functional_2016 types for our dataset only having access to the same type of information as MDS clustering between baden_functional_2016 type labels and predicted type labels on held out test data. Types confused by the classifier are merged into MDS clusters. Rows normalized (across baden_functional_2016 types) to sum to 1, annotations displaying the number of neurons in percent.
• Figure 3: Accuracy of identifying MDS clusters under simulated observational noise for varying repeats. Presentation time for the full set of seven MDS. Mean and standard deviation across ten simulations shown. MDS outperform the baseline after 1 min 45 s (9 repeats).
• Figure 4: MDS clustering on marmoset RGCs. Cluster averaged digital twin response to MDS normalized by mean predictions across the twin's training stimuli. Elements within a column normalized to highest value. MDS space component displayed on symmetric color scale between $-1$ (black) and 1 (white).
• Figure 5: MDS clustering on macaque V4 on held-out test neurons. Cluster averaged digital twin response to MDS normalized by mean predictions across the twin's training stimuli. Elements within a column normalized to highest value. MDS displayed on symmetric color scale between -1 (white) and 1 (black).