In one mouse retina, dendrite depth predicted the polarity and part of the time course of 47 held-out ganglion cell types; bipolar-type overlap and bipolar kinetics added little
We asked how much of the light response of a retinal ganglion cell (GC) type we can predict from the depth of its dendrites in the inner plexiform layer. We used the public e2198 data set: electron microscopy (EM) and calcium imaging of one mouse retina, 315 GCs of 47 types, and responses to a moving bright bar (mean of 8 directions). In a leave-one-type-out test, the dendrite overlap with the EM axon terminals of 13 bipolar cell (BC) types predicted the response shape of held-out types with mean r = 0.81 (95% CI 0.76-0.85). One common shape gave 0.70, the other cells of the same type gave 0.80, and the trial noise ceiling was 0.94. After removal of the common shape, the BC-type model gave centred r = 0.59 (0.46-0.71). This was better than the On/Off arbor fraction (0.51; difference +0.08, CI +0.02 to +0.14), but not better than four classical sublaminae (0.58; +0.02, CI -0.02 to +0.06). Structure predicted polarity across types (Spearman 0.74), but On and Off sustainedness only weakly (0.35 and 0.42). Arbor size gave no clear gain. An a priori model combined the same overlaps with BC glutamate responses that Franke et al. (2017) measured in other retinas. It fitted only three global parameters (calcium decay, time shift, rectification) and gave r = 0.78. But one mean kernel per polarity gave the same value. The published cluster-to-type assignment did better than random reassignment (p = 0.009, 1,000 permutations), but not better than reassignment inside the On and Off groups (p = 0.39). Across GCs, the overlaps with the BC types that Franke et al. call possibly permuted correlated at 0.96-1.00 (0.64-0.96 for the BC5 group). All 864 assignments inside these groups scored within 0.008 of each other. Thus GC depth data cannot resolve that ambiguity. Kernels inferred backwards from the GC responses had the correct On or Off sign for all 13 BC types. Limitations: one retina, one stimulus, calcium and not spikes, depth overlap and not synapses, and no amacrine inhibition in the model. The package has the code, derived tables and figures; raw data are downloaded from their sources.
retina · retinal ganglion cell · bipolar cell · inner plexiform layer · stratification · electron microscopy · calcium imaging · structure-function · e2198 · Eyewire · connectomics · temporal kinetics · sustained transient · identifiability · mouse
References
13- usesdoi:10.1016/j.cell.2018.04.040
Source of the e2198 GC types, stratification profiles and temporal responses (Eyewire Museum).
- usesdoi:10.1038/nature09818
Original e2198 calcium imaging and EM data set of the same retina patch.
- usesdoi:10.1038/nature13240
e2198 bipolar cell reconstructions and type definitions used for the BC profiles.
- usesdoi:10.1016/j.celrep.2016.02.001
e2198 bipolar cell reconstructions and type definitions used for the BC profiles.
- usesdoi:10.1038/nature21394
Bipolar cell glutamate responses for the a priori model; we test their cluster-to-type assignment.
- usesdoi:10.5061/dryad.rs2qp
Processed bipolar glutamate data file (FrankeEtAl_BCs_2017_v1.mat), via its Zenodo mirror.
- doi:10.1038/nature12346
Dense e2198 IPL reconstruction and bipolar cell type catalogue.
- doi:10.1038/35069068
Proposal that marginal and central IPL strata carry sustained and transient signals.
- doi:10.1038/nature16468
Mouse GC functional types and the response quality index.
- In one mouse retina, starburst contact direction predicted preferred motion direction within a median 12° for 26 On-Off and 23° for 7 transient On DS ganglion cells; arbor direction did not sci:sha256:d70a5a6322e8ec3db5b80f3d7871b4207590d3517332401bc14f8d9baeda0dc2
Same e2198 data; reports that directions came in a fixed order and that baselines carry the previous response, which motivated our response-frames-only check.
- Local contact correction reduced motion prediction error in 27 cells from one mouse retina sci:sha256:be8277b484073d3a865b5e829d4dd0c2c24aaeedc71969d419baf8647372c367
Same e2198 data; structure-based prediction of direction preference.
- With published labels and axon depth control, adding ribbon volume raised local plateau test error by 4.3% in eight mouse ON bipolar types sci:sha256:0307beb60fc74ab63d008205acc12e92d58b46eed0cde260c866902de3ab8432
Also uses the Franke 2017 clusters and their possible label exchanges, with a different structural measure (bipolar ribbon volume) and a bipolar-level outcome.
- With cell types excluded from training, directional contact fractions gave an uncertain 18% motion tuning error reduction in 302 ganglion cells from one mouse retina sci:sha256:68bca6ee92aa35b528db8802c021db39fc418600ab6fd314d7ac087c407f0a08
Same e2198 data; held-out-type prediction of motion tuning from starburst contacts.
Cited by
5 publicationsPublic catalog entries on Sciscape.
- A 128 ms window shift changed a depth model's error decrease from -6.4% to 20.1% in 178 cells from a second mouse retina
This earlier study tested depth prediction of excluded cell types within e2198. We test transfer to another retina and response window sensitivity.
- In one mouse retina, bipolar kinetics read through dendrite depth predicted response transience of On ganglion cells (Spearman 0.58; polarity only −0.12), not of Off cells or within types
[Legacy relation: reproduces] Same e2198 cells and a priori BC-kernel model: our independent code reproduced its whole-trace result (mean cell r 0.778 vs 0.78; polarity-mean kernels equal) and the small effect of the 864 assignments. We add cell-level On/Off-specific kinetics, within-type and label-free tests.
- In 147 ganglion cells from one mouse retina, depth profile similarity gave no clear added gain for movie response similarity in a ridge test with chirp and quality controls
This prior SCI study tested dendrite depth and moving-bar responses in the e2198 retina. It motivates a separate test with repeated natural movies in Eyewire II.
- challengesIn one mouse retina, bipolar release kinetics by layer depth ranked the sustainedness of 20 On ganglion cell types (ρ 0.77) but not of 13 Off types (ρ 0.05, below depth-scrambled maps)
It concluded that BC kinetics add little over one kernel per polarity. We agree for whole traces, but for type-level On sustainedness depth-resolved kinetics beat one kernel per polarity (difference +0.98, CI +0.38 to +1.40).
- In a fixed test of 5,542 pairs from one mouse retina, dendrite depth overlap gave an uncertain 0.21% error decrease for residual calcium correlation
Prior SCI test of depth and mean visual response in the same retina; the current target is residual pair correlation.