BGPT: Paper Review: The mouse superior colliculus promotes competing actions independently of sensory inputs

Fuel Your Discoveries

Quick Explanation Copied

Core claim

In mice, the superior colliculus (SC)—especially SCm—shapes choice competition by adding stimulus-independent action pressures for eligible actions (Left/Right) relative to inaction (NoGo), rather than primarily providing multisensory integration for perception.

Key causal test: unilateral SCm inactivation shifts contralateral choices toward NoGo without changing sensory sensitivity; bilateral inactivation restores Left/Right balance while increasing NoGo.

Evidence: behavioral 3-option logistic classifier, SC-layer recordings with sparse sensory vs action kernel separation, and optogenetic inactivation.

Long Explanation

Paper Review (Evidence-first, skeptical): SC action competition independent of sensory inputs

Manuscript: The mouse superior colliculus promotes competing actions independently of sensory inputs

What’s new (hypothesis)

SCm provides stimulus-independent pressures favoring eligible actions over NoGo.

Main causal readout

Unilateral SCm shifts choices toward NoGo without changing sensory sensitivity terms.

Computational backbone

A 3-option logistic classifier and a modified DDM explain Left/Right/NoGo simultaneously.

1) Behavioral observables the model must explain

NoGo outcomes were reported as 9.6 ± 1.2% across mice (N=18 mice; 491 sessions), even after excluding sequences associated with disengagement and correcting for delayed response windows.

2) Unilateral SCm inactivation re-times choices

Reported effects: unilateral SCm inactivation sped ipsilateral choices by 68 ± 37 ms and slowed contralateral choices by 99 ± 52 ms; these timing effects canceled under bilateral inactivation.

3) SC neurons: mostly unimodal sensory encoding + separate task kernels

Passive-kernel fitting with a variance-explained threshold (>0.5%) is reported to yield ~20% visual-only and ~10% auditory-only neurons, with only ~3% showing both visual and auditory kernels.

Visualize the causal-computational core

The paper proposes a three-option winner-take-all-like competition among Left, Right, and NoGo, implemented by a 3-option logistic classifier (softmax) that adds (i) sensory decision variables and (ii) two stimulus-independent action-pressure intercept terms.

The network view is schematic (faithful to the paper’s stated logic), not a mechanistic circuit diagram with asserted connectivity.

Methods-to-conclusion traceability (what is strong vs what remains uncertain)

A) Behavioral modeling strength

The key methodological choice is to treat behavior as three-choice (Left/Right/NoGo) rather than collapsing to only Left vs Right. The authors argue that a previous two-class model would absorb NoGo-related structure as a single bias, whereas their three-class model can separately estimate sensory weights and two action-pressure intercepts for Left and Right relative to NoGo.

B) Neural encoding claims: use kernels + thresholds (but thresholding is a choice)

The authors fit linear models with sensory kernels (visual low/mid/high contrast kernels; auditory base + spatial kernels) and separate engagement and action kernels. They report layer-dependent segregation (visual-dominant in superficial layers, auditory-dominant in deeper/motor layers) and very low audiovisual co-tuning among SC neurons, arguing against SC being a canonical multisensory integrator in this task context.

C) Causality hinge: optogenetic SCm inactivation

The central causal statement is: unilateral SCm inactivation alters contralateral choice preference by shifting it toward NoGo, while leaving sensory sensitivity terms (contralateral visual and auditory weights) unchanged in the three-option model.

D) Modeling RT & NoGo: modified DDM with lazy gain

The “lazy” gain mechanism is used to explain why NoGo trials are frequent and how choice timing changes under unilateral vs bilateral SCm manipulations. This is an interpretive model class—useful for quantifying—but it is not the same thing as identifying the neural computation.

Critical evaluation (skeptical blindspots & falsification pressure)

1) Thresholding & definition of “sensory vs action kernels”

The claim that SC neurons are largely unimodal relies on a variance-explained threshold and on the particular kernel basis (linear kernels with prescribed temporal supports). If audiovisual effects are nonlinear, subthreshold, or distributed across time windows not captured by the kernel design, the “multisensory segregation” conclusion could be underestimated. This doesn’t refute the core causal effect, but it weakens the strength of the neural encoding-to-computation mapping.

2) Optogenetic inactivation spread vs specific population inhibition

Inactivation is performed with eNpHR3.0 and is delivered via fibers; the authors estimate affected tissue areas using an eNpHR3.0 EPD50 threshold with Monte Carlo light spread. Even with estimates, the spatial-temporal uncertainty means SC inactivation might also suppress neighboring or passing axons, potentially altering network states rather than a clean “SCm local action-pressure” effect. The paper partly addresses this by recording during inactivation in separate validation sessions, but this still leaves some uncertainty about exact circuitry.

3) Generality: training-dependent audiovisual representation

The authors explicitly note that low audiovisual co-tuning could be task- or training-specific (e.g., trained mice, fixed lateralized stimuli, and wheel-based orienting report). Many prior SC studies emphasize multisensory integration; this paper’s conclusion is consistent with its task design, but does not settle whether SC is universally non-integrative across contexts.

4) Circuit claim: “SC pressures to prefrontal or downstream integrator” is still open

The paper provides inactivation comparisons (SCm vs VIS vs MOs) to infer a functional division of labor. But it does not directly identify the anatomical output targets responsible for the action-pressure term. Given that SCm has multiple ascending and descending pathways, a mechanistic circuit-level dissection remains a known gap for this specific mechanism.

Connections to broader literature (for orientation, not to overclaim)

The SC’s involvement in action selection/competition is broadly consistent with SC-focused reviews describing multisensory mapping and roles in guiding movements and decisions.
Multisensory integration in SC neurons has been reported historically (e.g., auditory-visual interactions in SC), and this paper’s emphasis on stimulus-independent action pressures is therefore best viewed as a task- and mechanism-specific decomposition rather than a blanket denial of SC multisensory physiology.

Conclusions & what would change my mind

Most supported conclusion (by internal consistency across behavior + SCm inactivation + model parameter shifts): SCm causally biases which eligible actions win against NoGo via stimulus-independent action-pressure changes, while sensory sensitivity in the three-option model is largely preserved under unilateral SCm inactivation.

Confidence note: I’m fairly confident in the computational readout and the causal directionality implied by parameter changes, but less confident in the mechanistic mapping from “kernel separations” to “no SC multisensory integration,” due to thresholding/model assumptions.

What would disprove or materially revise the paper’s central interpretation? If future experiments show that SCm inactivation robustly changes sensory sensitivity (not just action pressures) across task manipulations that hold choice strategy constant, or if a competing model without separate action-pressure parameters fits NoGo prevalence and inactivation effects equally well, the “stimulus-independent action-pressure” framing would weaken.

Bespoke next actions on BGPT

Author reviews (quick access)

Feedback:

Updated: July 06, 2026