If you Google brain models, you'll find a lot of glossy talk about "decoding the mind" and other phrases that sound like they were focus-grouped by a startup and a TED stage. What you will not find often enough is the question that actually hurts: if your model claims it understands a brain circuit, does it still make sense after scientists hit that circuit with light and deliberately mess with it? That is the fight in a new eLife paper published on January 16, 2026 (Sourmpis et al., 2026; PMCID: PMC12810953).

The Problem With a Pretty Map

Neuroscientists increasingly build recurrent neural networks, or RNNs, from huge mouse electrophysiology datasets. The pitch is simple: if the network mimics recorded activity, maybe it captured something real about cortical traffic. Nice idea. Slight problem. A model can look smart right up until reality throws a chair at it.

That reality check is optogenetics. In plain English, researchers give specific neurons light-sensitive proteins, then use light to activate or suppress those cells with absurd precision. It is one of the field's sharpest tools for asking causal questions instead of collecting fancy correlations (Lee et al., 2022; PMCID: PMC10543742; Adesnik and Abdeladim, 2021; PMCID: PMC9793863).

Sourmpis and colleagues asked a brutal question: train an RNN on normal cortical recordings, then see whether it can predict what happens during unseen optogenetic perturbations. Not just "can it vibe with the data?" but "can it keep its story straight when the supply lines get bombed?"

Black-Box Models, Meet Artillery

The answer for generic RNNs was not flattering. These standard models often fit unperturbed activity well, but generalized poorly when unseen optogenetic perturbations arrived. They could pass for experts in peacetime and then lose the plot once the first flare went up. Classic consultant behavior.

The team's alternative was a biologically informed RNN. Instead of treating every unit like a generic mathematical widget, they added some cortical common sense: excitatory and inhibitory neurons were structured differently, the sign of connections followed Dale's law, inhibitory projections stayed local, and the network used spiking-style dynamics. Those are guesses about how the cortical chain of command is actually organized.

That mattered. The biologically informed models did better at predicting unseen perturbations in both simulated data and real mouse data. The especially useful ingredients were not "more biology" in the abstract. Two constraints did most of the work: Dale's law, which keeps neurons from being both excitatory and inhibitory depending on their mood, and local-only inhibition, which fits the idea that many inhibitory interneurons act like neighborhood enforcers rather than long-range empire builders. Spiking helped less dramatically, which is a refreshing bit of scientific honesty. Not every biologically inspired idea gets a medal.

Why This Is More Than Model Nerd Drama

This matters because neuroscience has a recurring problem: a model predicts data well enough to impress a seminar room, but nobody knows whether it captured a mechanism or just learned a clever imitation. Recent work in cortical circuits has shown how targeted perturbations can expose recurrent circuit logic, especially in visual cortex (Oldenburg et al., 2024). The broader field is also moving toward models with more biological detail attached (Bazinet et al., 2023).

So this paper's real contribution is not "biology beats AI," because that bumper sticker would be nonsense. It is narrower and more useful: if you want a circuit model that survives causal testing, some biological constraints appear to improve out-of-distribution generalization. Translation: the model is less likely to panic when the experiment stops being polite.

There is also a spicy forward-looking angle. The authors show that gradients from the fitted model could, in principle, help design tiny targeted perturbations, what they call micro-perturbations, to bias behavior or dissect circuit mechanisms with finer control. That does not mean we are five minutes away from remote-controlling a mouse like a tiny furry drone. It does mean models might help experimentalists choose smarter perturbations instead of blasting whole populations and hoping for wisdom.

The Catch, Because There Is Always a Catch

No, this does not mean we have "solved" cortex. Please put the victory trumpet back in the case. These models were tested in a specific mouse setting, perturbation data were limited, and even the winning assumptions are still simplified.

Still, this is the kind of progress worth watching. Recent expert roadmaps frame optogenetics not just as a flashy lab trick, but as a way to build causal knowledge that could later inform therapies and better interventions (Lüscher et al., 2025). Before any of that gets near a clinic, though, we need models that do not collapse the moment the lights come on. This paper makes a decent case that a little biological discipline helps.

References

Sourmpis C, Petersen CCH, Gerstner W, Bellec G. Biologically informed cortical models predict optogenetic perturbations. eLife. 2026;14:RP106827. DOI: https://doi.org/10.7554/eLife.106827. PMCID: https://pmc.ncbi.nlm.nih.gov/articles/PMC12810953/

Lee JH, Liu Q, Dadgar-Kiani E. Solving brain circuit function and dysfunction with computational modeling and optogenetic fMRI. Science. 2022;378(6619):493-499. DOI: https://doi.org/10.1126/science.abq3868. PMCID: https://pmc.ncbi.nlm.nih.gov/articles/PMC10543742/

Adesnik H, Abdeladim L. Probing neural codes with two-photon holographic optogenetics. Nature Neuroscience. 2021;24:1356-1366. DOI: https://doi.org/10.1038/s41593-021-00902-9. PMCID: https://pmc.ncbi.nlm.nih.gov/articles/PMC9793863/

Oldenburg IA, Hendricks WD, Handy G, et al. The logic of recurrent circuits in the primary visual cortex. Nature Neuroscience. 2024;27:137-147. DOI: https://doi.org/10.1038/s41593-023-01510-5

Bazinet V, Hansen JY, Misic B. Towards a biologically annotated brain connectome. Nature Reviews Neuroscience. 2023;24(12):747-760. DOI: https://doi.org/10.1038/s41583-023-00752-3

Luscher C, Emiliani V, Farahany N, et al. Roadmap for direct and indirect translation of optogenetics into discoveries and therapies for humans. Nature Neuroscience. 2025;28:2415-2431. DOI: https://doi.org/10.1038/s41593-025-02097-9

Disclaimer: The image accompanying this article is for illustrative purposes only and does not depict actual experimental results, data, or biological mechanisms.