Biologically informed cortical models predict optogenetic perturbations

Reviewer #1 (Public review):

I congratulate the authors on this beautiful work.

This manuscript introduces a biologically informed RNN (bioRNN) that predicts the effects of optogenetic perturbations in both synthetic and in vivo datasets. By comparing standard sigmoid RNNs (σRNNs) and bioRNNs, the authors make a compelling case that biologically grounded inductive biases improve generalization to perturbed conditions. This work is innovative, technically strong, and grounded in relevant neuroscience, particularly the pressing need for data-constrained models that generalize causally.

I have some suggestions for improvement, which I present in the order of re-reading the paper.

Major

(1) In line 76, the authors make a very powerful statement: 'σRNN simulation achieves higher similarity with unseen recorded trials before perturbation, but lower than the bioRNN on perturbed trials.' I couldn't find a figure showing this. This might be buried somewhere and, in my opinion, deserves some spotlight - maybe a figure or even inclusion in the abstract.

(2) It's mentioned in the introduction (line 84) and elsewhere (e.g., line 259) that spiking has some advantage, but I don't see any figure supporting this claim. In fact, spiking seems not to matter (Figure 2C, E). Please clarify how spiking improves performance, and if it does not, acknowledge that. Relatedly, in line 246, the authors state that 'spiking is a better metric but not significant' when discussing simulations. Either remove this statement and assume spiking is not relevant, or increase the number of simulations.

(3) The authors prefer the metric of predicting hits over MSE, especially when looking at real data (Figure 3). I would bring the supplementary results into the main figures, as both metrics are very nicely complementary. Relatedly, why not add Pearson correlation or R2, and not just focus on MSE Loss?

(4) I really like the 'forward-looking' experiment in closed loop! But I felt that the relevance of micro perturbations is very unclear in the intro and results. This could be better motivated: why should an experimentalist care about this forward-looking experiment? Why exactly do we care about micro perturbation (e.g., in contrast to non-micro perturbation)? Relatedly, I would try to explain this in the intro without resorting to technical jargon like 'gradients'.

Minor

(1) In the intro, the authors refer to 'the field' twice. Personally, I find this term odd. I would opt for something like 'in neuroscience'.

(2) Line 45: When referring to previous work using data-constrained RNN models, Valente et al. is missing (though it is well cited later when discussing regularization through low-rank constraints).

(3) Line 11: Method should be methods (missing an 's').

(4) In line 250, starting with 'So far', is a strange choice of presentation order. After interpreting the results for other biological ingredients, the authors introduce a new one. I would first introduce all ingredients and then interpret. It's telling that the authors jump back to 2B after discussing 2C.

(5) The black dots in Figure 3E are not explained, or at least I couldn't find an explanation.

Reviewer #2 (Public review):

Sourmpis et al. present a study in which the importance of including certain inductive biases in the fitting of recurrent networks is evaluated with respect to the generalization ability of the networks when exposed to untrained perturbations.

The work proceeds in three stages:
(1) a simple illustration of the problem is made. Two reference (ground-truth) networks with qualitatively different connectivity, but similar observable network dynamics, are constructed, and recurrent networks with varying aspects of design similarity to the reference networks are trained to reproduce the reference dynamics. The activity of these trained networks during untrained perturbations is then compared to the activity of the perturbed reference networks. It is shown that, of the design characteristics that were varied, the enforced sign (Dale's law) and locality (spatial extent) of efference were especially important.
(2) The intuition from the constructed example is then extended to networks that have been trained to reproduce certain aspects of multi-region neural activity recorded from mice during a detection task with a working-memory component. A similar pattern is demonstrated, in which enforcing the sign and locality of efference in the fitted networks has an influence on the ability of the trained networks to predict aspects of neural activity during unseen (untrained) perturbations.
(3) The authors then illustrate the relationship between the gradient of the motor readout of trained networks with respect to the net inputs to the network units, and the sensitivity of the motor readout to small perturbations of the input currents to the units, which (in vivo) could be controlled optogenetically. The paper is concluded with a proposed use for trained networks, in which the models could be analyzed to determine the most sensitive directions of the network and, during online monitoring, inform a targeted optogenetic perturbation to bias behavior.

The authors do not overstate their claims, and in general, I find that I agree with their conclusions. A couple of points to be made:

(1) Some aspects of the methods are unclear. For comparisons between recurrent networks trained from randomly initialized weights, I would expect that many initializations were made for each model variant to be compared, and that the performance characteristics are constructed by aggregating over networks trained from multiple random initializations. I could not tell from the methods whether this was done or how many models were aggregated.

1. It is possible that including perturbation trials in the training sets would improve model performance across conditions, including held-out (untrained) perturbations (for instance, to units that had not been perturbed during training). It could be noted that if perturbations are available, their use may alleviate some of the design decisions that are evaluated here.