When Does the Real Wiring Win? Testing connectome structure-function claims against unmeasured physiology

Yeyuan LiuarXiv 2026 · arXiv

TL;DR

Does the "real" wiring of the brain actually perform better, or are we just tuning our models to make it look that way? This paper introduces a rigorous framework to separate structural advantages from tuning biases, revealing that many "structure-function" claims in neuroscience are fragile and can even reverse when constrained by real-world physiological data.

What Problem Does This Paper Address?

The explosion of high-resolution connectomics (mapping every synapse in a brain) has led to a new kind of argument: "The real wiring diagram performs task X better than a shuffled version, therefore this specific structure is essential for function."

However, a connectome only tells you who connects to whom. It doesn't tell you the "unmeasured physiology"—how excitable a neuron is, how strong a synapse is, or how fast a circuit integrates signals. Modellers typically tune these parameters until the model with the real connectome works, then test "null" (shuffled) models using those same parameters.

The author argues this creates a massive bias, analogous to the "winner's curse" or "under-tuned baselines" in machine learning. If you pick parameters that make the real wiring succeed, you are likely picking parameters that make the null wiring fail, regardless of their inherent potential.

Methods

The paper defines three comparison protocols to isolate these effects:

  1. P1 (Transfer): Tune physiology for the real wiring, then transfer those parameters to nulls. This is current common practice.
  2. P2 (Independent Tuning): Tune each wiring (real and null) independently to find its own "best case."
  3. P3 (Data-Constrained): Tune each wiring independently, but restrict the search to physiological ranges supported by independent experimental data (e.g., calcium imaging).
Figure 1: Schematic of the three protocols. P1 selects for the real wiring's success, P2 compares best-cases, and P3 restricts the comparison to biologically plausible regimes.

The author tests these protocols across three Drosophila (fruit fly) circuit models:

  • Mushroom Body: Associative learning.
  • Head-direction Ring Attractor: Spatial memory and integration.
  • Flyvis: A large-scale trained visual system model.

The study also uses "placebo nulls"—comparing two different instances of the same shuffled family—to measure the inherent bias of the P1 protocol itself.

Datasets

  • FlyWire v783: The synapse-resolution connectome of the adult fruit fly brain.
  • Synthetic Ground Truth: Used to validate the framework where the "true" effect is known.
  • Flyvis Ensemble: A pretrained ensemble of 50 visual system models.
  • Independent Physiological Data: Published calcium imaging data used to set the constraints for the P3 protocol.

Key Insights and Ideas

  • Tuning can reverse structural claims: In the Mushroom Body, the real wiring appears superior under P1 and P2. However, under P3 (constrained by imaging data), the real wiring actually performs worse than a random shuffle.
  • Model Adequacy vs. Wiring Deficit: When a model fails under P3, it often reveals that the model itself is inadequate for biologically realistic parameter regimes, rather than a deficit in the fly's actual wiring.
  • Robust vs. Fragile Claims: Some claims are "bulletproof." For example, the requirement for specific topography in the head-direction circuit survives every protocol. Others, like specific Kenyon cell output wiring, are highly sensitive to unmeasured parameters.
  • The Bias is Quantifiable: Using "placebo nulls," the author shows that P1 consistently reports a spurious advantage for the "real" model even when comparing two random shuffles. This bias can be predicted using a closed-form Gaussian approximation (Proposition 2).

Key Results

  • Mushroom Body Reversal: The real Kenyon-cell output wiring beats a shuffle by +0.09 under P1/P2 but loses to it by -0.09 under P3. The "deciding parameter" is the share of $\alpha'\beta'$ cells among active neurons; the model only "works" when this share is low, but imaging shows it is actually high (~44%).
  • Head-Direction Robustness: The real topology is essential for a persistent "bump" of activity across all protocols, with the real circuit beating rewired versions by massive margins (e.g., +0.72 in memory score).
  • Flyvis Mirror Test: In the visual system, P1 is so biased that it penalizes a mirrored connectome—which has a known, mathematically equivalent solution—more than it penalizes a spatially shuffled one.
Figure 2: Summary of results. Note how the KC->MBON claim (top) reverses sign under P3, while the head-direction claims (bottom) remain robustly positive.

Limitations and Future Work

  • Model Class: The study uses rate-based models; more detailed biophysical or spiking models might behave differently, though the author notes the protocol biases are likely universal.
  • Spatial Constraints: While the author tested "spatially constrained" nulls (swapping synapses only between nearby neurons), these are still proxies for true wiring costs.
  • P3 Availability: The P3 protocol is the most rigorous but requires independent experimental data, which is often unavailable for many brain regions.
  • Future Direction: The author suggests that when a model fails under P3, the next step isn't to claim the wiring is "bad," but to build better models that can actually function in the biologically observed parameter space.

Summary updated

Abstract

Connectome-constrained models are often used to argue that a measured wiring diagram outperforms shuffled or rewired wiring, and hence that structure shapes function. Such comparisons can confound wiring with how the models are tuned: a connectome fixes who connects to whom, not the excitability, gains or plasticity rates the model also needs, and modellers often tune these until the real-wiring model works. We separate three estimands: physiology tuned on the real wiring and transferred to nulls (P1), each wiring tuned on its own (P2), and P2 restricted to physiology consistent with independent data (P3). The biases involved are known in machine learning (under-tuned baselines, inherited weights, winner's curse); we measure what they do to biological conclusions, using synthetic ground truth, placebo nulls and three Drosophila circuit models. Some claims survive every protocol we test: segregated reward and punishment dopamine improves learning, and head-direction topography is required for a persistent bump. Others do not. In the mushroom body, the real Kenyon-cell output wiring beats a cross-subtype identity shuffle by 0.09 under P1 and P2 but loses to it by 0.09 under P3 (95% CIs exclude zero; held-out runs confirm). The verdict is set by the share of alpha'beta' cells among active Kenyon cells, which P1 selects low; at the share implied by imaging (about 44%) the real-wiring model fails to learn, exposing a model inadequacy rather than a deficit of the fly's wiring. In flyvis, P1 penalises a mirrored connectome, which has a constructed solution within 0.07 (median) of the real error, more than a spatially shuffled one in 50 of 50 members; the ranking survives ten-fold longer retraining but depends on the checkpoint at fifty-fold. A credible structural claim needs matched optimisation, data-constrained physiology and a model that still works there.
BibTeX
@misc{liu2026when,
  author = {Yeyuan Liu},
  title = {{When Does the Real Wiring Win? Testing connectome structure-function claims against unmeasured physiology}},
  year = {2026},
  eprint = {2611.14508},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.14508},
}
Figure