Research map / Probabilistic and scientific modeling / Dynamical systems
Most discriminative stimuli for functional cell type clustering
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an EM-style clustering of neurons that alternates between optimizing most discriminative stimuli (MDS) on digital twin models and reassigning neurons, recovering functional cell types in mouse/marmoset retina and macaque V4.
Problem
Unbiased identification of functional cell types from neural responses without domain knowledge or hand-crafted fingerprinting stimuli, especially in visual cortex (V4) where types are unknown, and fast cell typing in new experiments.
Approach
Jointly cluster neurons and optimize a per-cluster stimulus (MDS) on a digital twin that maximally drives one cluster while suppressing all others via a softmax objective, alternating M-steps (stimulus optimization) and E-steps (neuron reassignment), with split-based model selection for the number of clusters.
Key results
- On mouse RGCs, MDS clustering found seven clusters that map to OFF, ON-OFF, fast ON, slow ON and color-opponent groups; the reduced chirp classifier dropped from 79% to 49% accuracy on test data.
- Using only MDS, nine repeats identified the correct cluster better than the chirp-based baseline (73.5% accuracy) while saving 20% (27 s) of experimental time.
- Marmoset MDS clustering yielded four clusters (ON/OFF split into fast and slow), with median ARI 0.96 across nine comparison runs; macaque V4 yielded 12 well-separated clusters with ARI 0.53 across seeds.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use nine MDS repetitions instead of chirp responses | Mouse RGC cell-type assignment · Mouse RGC | time saved: 20% (27 sec) | Helped |
| Use chirp responses only; omit direction selectivity and soma size (ablation) | Mouse RGC type classifier · Mouse RGC type classification, test | accuracy: 79% to 49% | Hurt |
| Initialize with 30 instead of five clusters (ablation) | Mouse RGC MDS clustering · Mouse RGC robustness | ARI: 0.87 | No clear effect |
| Change the digital-twin architecture by adding a third layer and changing convolutional kernel sizes (ablation) | Mouse RGC MDS clustering · Mouse RGC | cluster similarity | No clear effect |
| Vary the objective temperature (ablation) | Mouse RGC MDS clustering · Mouse RGC | objective value and clustering performance: between 0.8 and 12.8; five to seven | Mixed |
| Optimize two MDS per cluster instead of one (ablation) | Mouse RGC MDS clustering · Mouse RGC | cluster specificity for ON-OFF cells | Helped |
| Initialize with a different random seed and ten instead of five clusters (ablation) | Marmoset RGC MDS clustering · Marmoset RGC robustness | number of final clusters | No clear effect |
| Initialize with a different seed and ten instead of five clusters (ablation) | Macaque V4 MDS clustering · Macaque V4 robustness | number of final clusters | No clear effect |
Surprising findings
- MDS clustering did not capture direction selectivity, because the digital twin is limited in modeling it; none of the MDS showed significant lateral motion.
- Fewer clusters (7) than the 17 reference types were found, with several fine types merged because they are not separable without direction selectivity or soma size.
Limitations
- Assumes discrete cell types; cannot represent continuous variation within or between types.
- Digital twin cannot capture direction selectivity, so direction-selective types are not separated.
- Ground-truth labels are imperfect (classifier-based RGC labels and MEI-based V4 labels), and results are not verified in vivo.
Experimental setup
Datasets: mouse RGC calcium imaging dataset (naturalistic UV/green movies, 2,448 RGCs, 17 types) marmoset RGC MEA recordings (167 cells) macaque V4 spiking responses to natural images (1,224 neurons, 1,030 used)
Benchmarks: mouse RGC functional types (Baden/Franke-style 17 types) marmoset RGC STA-based clusters macaque V4 MEI-based cluster labels
Metrics: objective J_c adjusted rand index classification accuracy confusion matrices identification accuracy vs presentation time
Baselines: chirp-based classifier (Baden/Franke-style RGC type classifier) STA-based clustering MEI-based clustering of V4 neurons (Ponce et al.) Klindt et al. digital twin parameter clustering
Builds on: maximally exciting inputs (MEI) digital twin CNN core-readout models Ecker et al. retina functional clustering
Closest papers by meaning
- Rotation-invariant clustering of neuronal responses in primary visual cortex (ICLR 2020 · similarity 0.91)
- Learning to cluster neuronal function (NeurIPS 2025 · similarity 0.90)
- Reproducibility of predictive networks for mouse visual cortex (NeurIPS 2024 · similarity 0.88)
- A data and task-constrained mechanistic model of the mouse outer retina shows robustness to contrast variations (NeurIPS 2025 · similarity 0.87)
- Information Geometry of the Retinal Representation Manifold (NeurIPS 2023 · similarity 0.87)
- Disentangling the Roles of Distinct Cell Classes with Cell-Type Dynamical Systems (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.