前言

Google 和合作實驗室花了十年重建一個完整的雄性果蠅中樞神經系統連接體，並將資料集發佈為 MaleCNS v1.0。這項成就以前所未有的細節繪製了每一個神經元和突觸，讓研究人員能追蹤從感覺器官到運動輸出的資訊流。資料發布幾天內，學術界外的工程師已將開放數據改作調皮且具挑釁性的實驗——最著名的是嘗試驅動果蠅模型玩經典電子遊戲 Doom，之後甚至用於交易加密貨幣。本文總結這項科學里程碑，並說明開放科學如何促成軟體工程師迅速且有創意的再利用。目標是以清晰、中立的語氣呈現背景、方法，以及倫理與技術含意。

重點摘要

MaleCNS v1.0 連接體是公開發佈的雄性果蠅神經系統完整接線圖。 在幾天內，工程師將連接體轉換成互動專案：一個嘗試訓練果蠅模型在將遊戲畫面對應為感覺輸入的情況下導航 Doom；另一個則用相同方法在比特幣市場做出交易決策。這些實驗突顯開放資料如何加速意想不到的跨領域創新，同時提出關於能力、詮釋與防範措施的疑問。

主體

The MaleCNS v1.0 release represents a landmark in connectomics. Completed by a consortium including Google Research, HHMI Janelia, the MRC Laboratory of Molecular Biology, and the University of Cambridge, the dataset resulted from slicing a male fruit fly into millions of ultrathin sections, imaging each slice with electron microscopy, reconstructing three-dimensional neuron geometries using machine vision, and manually proofreading the outputs. The final product documents roughly 166,691 neurons and about 125 million synaptic contacts across more than 11,700 cell types, and — importantly — integrates both the brain and the ventral nerve cord (the insect analog of a spinal cord) into a single wiring diagram. This integration enables researchers to follow sensory signals from the eyes down to leg and wing motor circuits, offering a more complete substrate for studying behavior and sex-specific neural differences.

The scale and fidelity of the connectome are notable for multiple reasons. First, the dataset surpasses previous fruit fly connectomes in neuron count and in including the ventral nerve cord, enabling new comparisons between male and female circuits that underpin social behaviors such as courtship or aggression. Second, the open distribution via tools like Neuroglancer means that both neuroscientists and the broader community can inspect, download, and reuse the data. This openness accelerates reproducibility and invites creative applications beyond the original research goals.

Within days of the publication, a software engineer affiliated with Coinbase, Alex Wormuth, announced an experiment converting MaleCNS v1.0 into a real-time control system for the 1993 shooter Doom. The project, named DOOMFLY, mapped each video game frame into thousands of brightness and color signals designed to mimic the stimulation patterns that fly photoreceptors and early visual neurons would receive. These sensory-like inputs were injected into a simplified neural dynamics model running on the preserved circuit graph; a fixed readout mapped activity of selected motor-related neurons to game controls such as turn, move forward, and fire.

Training relied on a minimal reinforcement signal: when the in-game avatar took damage, two PPL101 dopaminergic neurons — interpreted as aversive reinforcement channels — received a synthetic stimulus intended to bias synaptic weights along roughly 4,184 plastic connections. All other connections were held static. The README for DOOMFLY explicitly frames the work as an ongoing real-time experiment and cautions that, at the time of writing, the fly agent had not demonstrated reliable survival behaviors. The project’s candidate versions failed initial visual, conditioning, and survival tests, and prolonged training did not produce a stable learning curve. In short, the simulated fruit fly did not yet learn to consistently survive in Doom.

Beyond gaming, Wormuth and others repurposed the connectome-driven pipeline for different tasks. Wormuth later adapted the system to interpret financial candlestick charts and make trading decisions on a Coinbase account in an open project called Stonkfly. The same raw connectome inspired additional playful experiments: engineers connected the wiring to Super Mario 64 (producing repeated jump-and-bump behaviors), Beat Saber controllers, and even Minecraft creatures. These projects demonstrate how an open, richly detailed biological dataset can be a sandbox for engineers and hobbyists to explore sensorimotor mappings, reinforcement paradigms, and emergent behavior.

These creative uses provoke questions across technical, scientific, and ethical dimensions. Technically, converting a biological connectome into a functioning controller requires many modeling assumptions: how to translate pixel-level stimuli into biologically plausible sensory inputs, how to simulate neural dynamics efficiently, which synaptic subsets to permit plasticity in, and how to define reward signals that map to meaningful biological analogs. Each design choice affects outcomes and interpretability. For example, restricting plasticity to a small set of synapses simplifies training but does not reflect the full plastic repertoire of a living nervous system.

Scientifically, such experiments can be informative when interpreted carefully. They offer a way to test whether wiring structure alone, combined with simple learning rules, suffices for certain sensorimotor tasks. Failed learning is instructive: it signals limits of structural data alone or highlights missing components such as neuromodulatory dynamics, developmental history, or realistic proprioceptive feedback. However, success in a constrained virtual task should not be overgeneralized to claims about cognition or biological intelligence.

Ethically and socially, repurposing biological connectomes raises concerns about data stewardship, dual use, and public perception. Open data promotes transparency and accelerates discovery, but it also enables unconventional and sometimes sensational applications that can be misinterpreted by the public. Researchers and platforms releasing datasets may consider providing clearer guidance about recommended uses, documentation of modeling assumptions, and narratives that contextualize what the data can and cannot demonstrate.

In conclusion, the MaleCNS v1.0 connectome is a major scientific resource. Its rapid uptake by engineers into projects like DOOMFLY and Stonkfly underscores the cultural shift toward open, interdisciplinary experimentation. These efforts illustrate both the promise of open neuroscience — enabling diverse explorations of structure-to-function relationships — and the need for careful interpretation, transparent methodology, and thoughtful discussion about the broader implications of reusing biological datasets in unconventional ways.

關鍵見解表