Optogenetics

An electrode cannot choose what it stimulates. Current spreads from the tip and excites whatever has the lowest threshold nearby, usually an axon — including axons of passage belonging to cells somewhere else entirely. Histed, Bonin and Reid saw this directly in 2009, imaging calcium with two-photon microscopy rather than recording through the artefact: microstimulation lit up a sparse, scattered set of neurons, some millimetres from the electrode.

Optogenetics inverts the arrangement. Rather than confine the stimulus, confine the susceptibility to it: put a light-gated ion channel in the cells you mean, illuminate the whole region, and only the cells carrying the gene respond. Zemelman and colleagues, who built the first version that worked in neurons, called this an inversion of the logic of existing stimulation methods. The receiver is encoded in DNA.

What was asked for, and what was borrowed

Francis Crick asked for it in print. His 1999 paper, expanded from a Kuffler Lecture given at San Diego that March, names what neurophysiology lacked: a way “to turn the firing of one or more types of neuron on and off in the alert animal in a rapid manner”. The signal he proposed was light, “probably at an infrared wavelength to allow the light to penetrate far enough”. The infrared half is still outstanding.

The 2002 version, from Zemelman, Lee, Ng and Miesenböck, reconstituted Drosophila phototransduction in rat hippocampal neurons — the opsin NinaE, arrestin-2 and the α subunit of the cognate G protein, together “chARGe”. Illuminated, chARGed neurons fired and untransfected neighbours did not. But a cascade is slow: spike latencies ran from a few hundred milliseconds to tens of seconds, and peak rates reached 7.5 Hz.

What fixed that came out of algal biophysics. In 2003 Nagel, Szellas, Hegemann, Bamberg and colleagues showed that channelrhodopsin-2, from Chlamydomonas reinhardtii, is a directly light-gated cation channel rather than a receptor driving a cascade, and could depolarise a cell simply by being illuminated.

Why one channel was enough

In 2005 Boyden, Zhang, Bamberg, Nagel and Deisseroth delivered ChR2 to cultured rat hippocampal neurons by lentivirus and pulsed blue light at them. Photocurrents reached a maximal rise rate of 160 ± 111 pA/ms within 2.3 ± 1.1 ms of light onset and peaked at 496 ± 336 pA — enough to cross spike threshold unaided.

The timing is the result. Steady illumination gave one well-timed spike and then ragged firing, so they pulsed instead: 5, 10 or 15 ms flashes, one spike each, latency 8.0 ± 1.9 ms and trial-to-trial jitter 0.5 ± 0.3 ms. Repeated Poisson-distributed trains agreed spike for spike on over 95% of pulses, and trains ran from 5 to 30 Hz at 8–12 mW/mm², fidelity falling at the top of that range.

Three properties made that a tool rather than a finding. The channel is a single open reading frame of 315 amino acids, so it fits in a virus behind a cell-type-specific promoter. Background retinal served as the chromophore, with nothing added. And light does not enter the recording amplifier, which Lima and colleagues turned into a labelling trick in 2009: a flash identifies which recorded unit belongs to which genetic class.

Fibres, micro-LEDs, and where the artefact went

ChR2’s first light source was a xenon lamp behind a fast shutter on a microscope. Everything that made the method usable in an animal is engineering, which is the direction this entry records. Aravanis and colleagues went in vivo in 2007 with a laser diode pulsed to the millisecond, 20 mW at 473 nm into a 200 µm multimode fibre, about 380 mW/mm² at the tip.

A fibre illuminates a cone below its face, brightest where the light leaves it, so reaching cells at the bottom of a structure means passing the fibre through everything above or turning the power up. Tae-il Kim, with Rogers and Bruchas’s groups, deleted the fibre in 2013: GaN micro-LEDs 50 × 50 µm and 6.45 µm thick, a thousand-odd times smaller than a packaged LED, printed with a microelectrode and a temperature sensor into a 20 µm stack, powered by RF scavenging at 910 MHz — 7 mW/mm² optical against the 1 mW/mm² or so many constructs need, with no tether.

Mounting the emitter on a recording shank, though, brings back the artefact light was supposed to avoid. Kanghwan Kim and colleagues at Michigan measured 2,478 ± 1,734 µV peak-to-peak in 2020 on a single-metal-layer probe, against a detection threshold near 100 µV. Part of it is capacitive coupling from the drive traces into the recording interconnects:

$$i_c = C_m \frac{dV}{dt}$$

where $C_m$ is the mutual capacitance between the two and $dV/dt$ the slew rate of the LED drive. A grounded shield layer attacks $C_m$, worth 5.2× alone; a 1 ms rise time with the off-state parked just under turn-on attacks $dV/dt$, cutting the swing from 3.5 V to 0.7 V. The rest is photovoltaic — light frees carriers in the silicon substrate itself — which heavy boron doping suppresses by shortening their lifetime. Together they reach 46.5 ± 11.3 µV, below threshold.

A loop that has to keep adjusting

Light is also easy to switch, which makes closed loops cheap. Krook-Magnuson, Armstrong, Oijala and Soltesz detected spontaneous hippocampal seizures in epileptic mice in real time and illuminated on a random half of the detected events, the rest serving as internal controls; silencing principal cells stopped 57 ± 12% of seizures within a second.

The clamp version is the more interesting one here. Newman and colleagues’ “optoclamp” (2015) is a proportional–integral controller on population firing rate, ChR2 pushing up and a halorhodopsin pushing down, holding cultured networks at a commanded rate from a minute to a day — tracking error 0.14 ± 0.09 Hz per unit, settling in 7.8 ± 6.0 s, stable for proportional gains below about one. The load-bearing control is that replaying a successful control signal open-loop failed to hold the rate at all. Opsins desensitise and networks adapt, so the loop is not a convenience but the mechanism — the argument the voltage clamp makes about membrane conductances, moved up to firing rate, which is where the name came from.

Perturbing a system in a known way to learn what it does is system identification, and it is what the method is for. Gradinaru and colleagues used it in 2009 to re-ask what deep brain stimulation does: driving the elements around the subthalamic nucleus one class at a time, the therapeutic effect in parkinsonian rodents followed stimulation of afferent axons — the axon rather than the cell body, where modelling of the clinical electrode had also landed.

What it costs

Genetic access is the binding constraint, and it mostly confines the method to laboratory animals. Expression varies with promoter, viral serotype, dose and cell type, so how much opsin a given cell carries is not something you set — which is why an opsin-negative animal receiving identical light is the standard control, and why Allen, Singer and Boyden’s rules for interpretable experiments are largely about controlling for it. Human use remains thin: Sahel and colleagues reported partial recovery of visual function in one blind patient in 2021, using ChrimsonR in retinal ganglion cells and light-stimulating goggles.

Light scatters, so tissue rather than the aperture sets the illuminated volume, and absorbed light becomes heat — a thermal budget like any implant’s. Stujenske, Spellman and Gordon measured and modelled it: 10 mW of continuous 532 nm light through a 62 µm fibre plateaus about 2.2 °C above baseline within a few hundred micrometres, and holds a degree or more across roughly a cubic millimetre. In mice with no opsin at all, 5 and 10 mW raised firing rates by 31 ± 16% and 43 ± 17%; Owen, Liu and Kreitzer found the opposite sign in 2019, 0.2–2 °C of heating suppressing spiking and biasing turning behaviour. That the confound has no consistent direction is the uncomfortable part. The fixes are familiar power-budget ones: heating falls in proportion to duty cycle, and longer wavelengths are absorbed less.

The deepest limit is that the stimulus is not physiological. Flooding a region drives every expressing cell at once, which is not how a population normally behaves: sustained depolarisation can block spiking outright, hyperpolarising opsins invite rebound spikes and load cells with chloride until the reversal potential for inhibition moves, and driving excitatory cells recruits the inhibition that silences their neighbours. What such an experiment shows is what a synchronous volley into that cell class can cause — a real causal fact, and not quite the one the cell class performs on its own.