Bayesian reaction optimization with fixed MaleCNS features: performance, wiring controls, and reward–aversion diagnostics
- 1Research Center for Macromolecules and Biomaterials, National Institute for Materials Science (NIMS, 1-2-1 Sengen, Tsukuba, Ibaraki, Japan, 305-0047
Abstract
Connectome-derived features provide one possible representation for reaction optimization, but their utility need not depend on measured neuronal pairing. We examine a Bayesian predictor using fixed mushroom-body features derived from MaleCNS connectivity, with observed prediction errors separated into engineered reward and aversion signals. An offline evaluator returns only selected recorded outcomes. We distinguish this predictor from earlier models that modify KC–MBON efficacy incrementally or reconstruct it from observed response ranks. The rank comparison changes both the learning head and acquisition rule. In the exploratory predictive-model evaluation, the mean best response over post-initial observations exceeded rank reconstruction by 0.486 percentage points on C–N coupling and 3.346 on Suzuki–Miyaura coupling. Differences from development-tuned Gaussian-process optimization were smaller, and their cluster-bootstrap intervals included zero. Native-versusrewired intervals also included zero. C–N responses were analytical yields; Suzuki responses were UV-area percentages. High/low response diagnostics separate input-dependent KC activity from outcome-dependent prediction updates: circuit activity remains fixed under an identical-cue outcome intervention. Reused datasets and input ensembles, and one rewired graph restrict inference. These results support evaluating anatomical representation, learning rule, search performance, and prediction accuracy separately; they establish neither statistical equivalence nor a unique advantage of native connectivity.
Related
- Enforcing Thermodynamic Consistency in Parameter Optimization and Bayesian Inference of Biochemical Reaction Networks
- Simulation-free approximate Bayesian computation for stochastic reaction networks
- Simulation-free approximate Bayesian computation for stochastic reaction networks
- Characterizing fatigue-related neural signatures via multi-band EEG features and a three-stage feature optimization strategy for driver fatigue detection
- The ethical and regulatory framework of medical artificial intelligence (AI) system’s behavioral performance for image-based diagnostics: An ethical evaluation of the International and European Union Scope
BibTeX
@misc{nagata2026bayesian,
title = {{Bayesian reaction optimization with fixed MaleCNS features: performance, wiring controls, and reward–aversion diagnostics}},
author = {Yuuya Nagata},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008959/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008959/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Bayesian reaction optimization with fixed MaleCNS features: performance, wiring controls, and reward–aversion diagnostics
AU - Yuuya Nagata
PY - 2026
DA - 2026/09/16
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008959/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008959/v1
AB - Connectome-derived features provide one possible representation for reaction optimization, but their utility need not depend on measured neuronal pairing. We examine a Bayesian predictor using fixed mushroom-body features derived from MaleCNS connectivity, with observed prediction errors separated into engineered reward and aversion signals. An offline evaluator returns only selected recorded outcomes. We distinguish this predictor from earlier models that modify KC–MBON efficacy incrementally or reconstruct it from observed response ranks. The rank comparison changes both the learning head and acquisition rule. In the exploratory predictive-model evaluation, the mean best response over post-initial observations exceeded rank reconstruction by 0.486 percentage points on C–N coupling and 3.346 on Suzuki–Miyaura coupling. Differences from development-tuned Gaussian-process optimization were smaller, and their cluster-bootstrap intervals included zero. Native-versusrewired intervals also included zero. C–N responses were analytical yields; Suzuki responses were UV-area percentages. High/low response diagnostics separate input-dependent KC activity from outcome-dependent prediction updates: circuit activity remains fixed under an identical-cue outcome intervention. Reused datasets and input ensembles, and one rewired graph restrict inference. These results support evaluating anatomical representation, learning rule, search performance, and prediction accuracy separately; they establish neither statistical equivalence nor a unique advantage of native connectivity.
ER -