Computing Motion in the Retina

A ganglion cell in a rabbit’s retina fires hard when a spot moves one way across its field and hardly at all when it moves the other way. Where in the retina is that direction computed, and by which cells?

The retina lesson built centre-surround fields from bipolar and amacrine circuitry, and sensory coding described lateral inhibition between neighbours. Neither gave a cell a sense of direction. This lesson starts where the retina does that, one synapse before the optic nerve. Motion in the cortex belongs to visual cortex. The cell census this circuit belongs to is in retinal cell types.

Two algorithms for the same job: suppress the null or enhance the preferred

A direction-selective ganglion cell (DSGC) is defined by what it does. It fires strongly to motion in its preferred direction and little or not at all to the opposite, called the null direction. Barlow and Hill reported such cells in the rabbit retina in Science in 1963, and they noted that replacing a light spot with a black spot kept the preferred direction. The 1965 paper by Barlow and Levick is the mechanism paper, and it is often credited with the discovery, which it was not.

Barlow and Levick recorded spikes from ON-OFF direction-selective units in the rabbit, in vivo, from unopened eyes of decerebrate or lightly anaesthetized animals. They stimulated two neighbouring places in one receptive field in sequence. From the results they inferred that small subunits of the field suppress responses to sequences running in the null direction. Their guesses about which cells do this were tentative, and you should read the paper as a statement of what the mechanism must do, not as an identification of starburst cells.

There are two simple ways to build a direction detector from two inputs. One suppresses: a delayed inhibitory signal from one side vetoes the excitation when motion runs the null way. The other enhances: a delayed signal from one side is multiplied with the signal from the other, so the response grows when motion runs the preferred way. The Barlow-Levick scheme is the first, and the Hassenstein-Reichardt scheme is the second. A two-flash test tells them apart. Measure each flash alone, add the two responses, and subtract the sum from the paired-flash response. A positive remainder in one order is enhancement, and a negative remainder in the other order is suppression.

Mechanism

Null-direction suppression (Barlow-Levick scheme) Preferred-direction enhancement (Hassenstein-Reichardt scheme) Both (enhancement and suppression together, as reported in fly T4 and T5 terminals)

Flash order

Preferred order: A then B Null order: B then A

Strength of the delayed signal, k 0.50 (relative units; 0 = none, 1 = strongest)

Sum of the two single flashes

1 + 1 = 2.00

Paired response (this order)

2.00

Remainder = paired − sum (this order)

0.00

Remainder, preferred order (A then B)

0.00

Remainder, null order (B then A)

0.00

Reading of the two orders

no remainder

Each flash alone counts 1, so the sum is 2.

In this sketch: single flashes add linearly, one fixed delay, a delayed veto or boost that shifts the paired response by k (the unaffected order is set to exactly the sum), relative units, no real interval or speed. Barlow and Levick’s preferred order gave about the sum or a little more. Which side carries the delayed signal is a labelling choice of this sketch; the lesson says only “one side”. The bars are computed from the rules, not recordings. The sketch does not say which cells make a negative remainder (Barlow and Levick state what the mechanism must do, not which cells), and the fly and mouse circuits reach a similar algorithm by convergent computation, not homology.

Rule to keep: measure each flash alone, add them, and subtract that sum from the paired response. A positive remainder in one order is enhancement, a negative remainder in the other order is suppression.

A dendrite that is more selective than its own cell body (2002)

The starburst amacrine cell (SAC) is an interneuron whose dendrites radiate from the soma like the spokes of a wheel and which overlap the dendrites of DSGCs. Euler, Detwiler and Denk imaged ON SACs in isolated rabbit retina, using two-photon calcium imaging, and compared the dendrites with the voltage at the soma. Calcium signals in a dendrite were larger for motion outward from the soma, called centrifugal motion, than for motion inward. The somatic voltage did not show the same directional preference under the conditions tested.

This changes what you should picture. A SAC does not have one preferred direction that its soma reports. Each branch has its own, set by where the branch points. Calcium is also not transmitter release, so the result shows a local directional signal and leaves the mechanism of that signal open.

How necessary is the starburst cell? Yoshida and colleagues ablated the SAC population in adult mice with an immunotoxin-based method and then recorded ganglion-cell responses and tested the optokinetic reflex, the eye movement that follows a moving visual field. Both direction selectivity in the ganglion cells and the reflex they tested were abolished. This is necessity in that assay. Because the SACs release both acetylcholine and GABA onto DSGCs (shown in rabbit by Lee and colleagues in 2010, with its role tested by Sethuramanujam and colleagues in 2016), taking the whole population away also removes both transmitters together.

Wiring read directly: imaging followed by electron microscopy (2011)

If centrifugal dendrites inhibit a DSGC, then the dendrites that contact it should point along its null direction, because that is the motion that drives them hardest. Briggman, Helmstaedter and Denk tested this in mouse retina. They first imaged calcium in DSGCs in a retina from a wild-type mouse, using moving bars, to read each cell’s preferred direction. They then cut the same tissue with serial block-face electron microscopy, which images the face of a tissue block after each slice is removed, and reconstructed the cells and their contacts. The microscopy and images lesson covers how that stack is made.

Most reconstructed SAC input onto a DSGC came from dendrites pointing roughly along that cell’s null direction. The orientation of the individual branch predicted whether it connected better than the line between the two cell bodies did. That second result fits the 2002 finding that the branch, not the cell, is the unit.

Two limits apply. Anatomy does not measure the strength of a synapse, so a contact count or area is an indirect sign of connectivity, not a conductance. And a snapshot of an adult retina says nothing about how the pattern arose.

Where the asymmetry comes from: strengthening, not pruning (2011)

An asymmetric circuit could arise by pruning: start with symmetric connections and delete the preferred-side ones. Wei, Hamby, Zhou and Feller tested this in mouse retina with paired recordings from SACs and nasal-preferring ON-OFF DSGCs, stimulating one SAC and measuring the inhibitory current in the ganglion cell. At postnatal days 4 and 7 the inhibition was similar from the two sides. Later, the null-side unitary inhibition increased while the preferred-side inhibition stayed about constant. Preferred-side connections were still detectable in adult retina.

So development strengthened one side and did not delete the other. The authors also found that direction selectivity developed despite their muscimol or gabazine treatment, which they interpret within that treatment window. For the activity patterns that wire the retina early on, see spontaneous activity, and for how cell types find their partners, see partners, layers and self-avoidance. The finding that strength, not presence, differs is one reason why recordings of a single connection matter more here than a count of connections.

Tuned input, tuned output, and a fly that solved it differently

The wiring story is about output from the SAC. The input side is under revision.

The EyeWire project (Kim and colleagues, Nature, 2014) reconstructed 79 OFF SACs in a mouse retinal electron microscopy volume. Contact area suggested that type BC2 bipolar cells contact proximal parts of a SAC dendrite and type BC3a cells contact more distal parts. Their model proposed that this spatial arrangement, with different timing in the two inputs, could favour outward motion. That is a model prediction, resting on contact area as an indirect measure.

Measurements followed. Fransen and Borghuis recorded excitation in ON and OFF SACs in whole-mount retina from mice aged 1.5 to 3 months, in vitro. Distal excitation was faster and more transient than proximal excitation in both types. Srivastava and colleagues imaged glutamate on SAC dendrites directly in mouse retina in vitro and found relatively sustained proximal input. Deconvolution suggested about three times larger steady release proximally, and their constrained model gave its strongest space-time contribution below about 500 micrometres per second.

Excitation is also not always untuned. Matsumoto and colleagues found that single bipolar cells have output boutons with different preferred directions in mouse retina, and that this depended on SAC cholinergic excitation and on wide-field amacrine inhibition. The DSGCs they recorded received both direction-tuned and untuned glutamatergic input. An upstream cell type does not therefore make a clean feed-forward origin. For the primary-visual-cortex consequences, Hillier and colleagues reported in mice that disrupting retinal direction selectivity abolished the over-representation of posterior-motion-preferring layer 2/3 cells in V1, while other cortical motion computations persisted.

In Drosophila the same job is done by T4 cells (the ON channel) and T5 cells (the OFF channel). In vivo calcium imaging of their axon terminals by Haag, Mishra and Borst found both preferred-direction enhancement and null-direction suppression in all four subtypes. The fly and the mouse reach a similar algorithm with different cells, so this is convergent computation, not homology, and no single idealised detector accounts for either circuit.

On the primate side, a textbook line says that primates lack DSGCs. That is out of date: physiological direction selectivity has been reported in macaque ON-OFF cells (Kim and colleagues, 2022) and ON cells (Wang and colleagues, 2023). The 2023 human cell was identified as a candidate homolog without a human motion recording.

Worked problem: indices and timing

(a) Kim and colleagues (2022) reported, for dendritic calcium in macaque SACs (adult isolated retina, radial gratings), a mean inward-to-outward response ratio of 0.32 (SD 0.18) across 36 regions of interest on the terminal dendrites of one starburst cell. For this exercise only, normalise the outward response to 1 and define D = (Rout - Rin) / (Rout + Rin).

D = (1 - 0.32) / (1 + 0.32) = 0.68 / 1.32 = 0.515, outward-preferring

This transforms a published mean ratio. It is not a published direction-selectivity index, and because the formula is non-linear it is not the mean of individual indices, which would need the paired responses. The 36 counts regions of interest on one cell, not cells or animals.

(b) Fransen and Borghuis found that distal excitation in mouse SACs is faster than proximal excitation. For this exercise only, suppose the proximal-to-distal latency difference is 26 ms per 100 micrometres in an OFF SAC and 8 ms per 100 micrometres in an ON SAC; these are illustrative values, not figures from the paper. A pure delay line makes the signals coincide when the stimulus crosses 100 micrometres in the delay time.

OFF: 100 / 26 = 3.8 micrometres per ms, about 3,800 micrometres per second. ON: 100 / 8 = 12.5 micrometres per ms, about 12,500 micrometres per second.

These speeds sit well above the below-500 micrometres per second regime in which the 2022 constrained model found its strongest space-time contribution. With these invented values a pure delay line would match only far faster stimuli than that regime, which is a property of the made-up numbers and not a finding. Srivastava and colleagues measured similar proximal and distal latencies and found the contribution of sustained and transient kinetics at low speeds. Real latency differences depend on the radial positions and the preparation, so do not restate these values as a universal delay.

Design problem

Inhibition from starburst cells on the null side is stronger, SACs release both acetylcholine and GABA, and removing the SAC population abolished direction selectivity in the assay of Yoshida and colleagues. Does asymmetric GABAergic inhibition cause direction selectivity? Say what experiments would settle it.

One way to do it

Start by separating necessity from sufficiency. Removing the whole population removes two transmitters and every other function of the cells at once, so it shows that the cells are needed, not that GABA is.

Test the transmitter directly. Remove GABA release from SACs while leaving the cells and their cholinergic output in place, and compare the result with acute, reversible silencing of the same cells. If the ratio falls but does not reach one, other mechanisms contribute.

Separate development from the adult requirement. A manipulation begun before about postnatal day 7 could change the wiring, and a manipulation in the adult changes signalling. The two answer different questions.

Add a control for tuned excitation, because bipolar boutons can themselves be direction tuned and could supply selectivity that survives a GABA block.

State the species scope: the physiology is established in rabbit and mouse, and in macaque for direction-selective ganglion cells, while the human cell is a candidate homolog only.