计算神经科学 · 果蝇 · 2026 年 9 月Computational neuroscience · Drosophila · September 2026
当一个黑影朝果蝇扑来，它脑中一根叫巨纤维的神经元会发出信号，果蝇随即起跳逃走。这个项目把这段回路按一只真实果蝇的完整接线图，一个神经元一个神经元地搬进电脑。然后先检查它像不像真果蝇，再一步步打乱它的接线，看它的反应会不会变。When a dark shape rushes at a fruit fly, a neuron in its brain called the giant fiber fires and the fly jumps away. This project copied that circuit into a computer, neuron by neuron, from the complete wiring diagram of a real fly. It then checked whether the copy behaves like a real fly, and scrambled its wiring step by step to see whether its response changes.
这是静态的 PDF 版。可以交互的网页版：https://ewanwu06.github.io/fly-escape-switch/This is the static PDF edition. The interactive web page: https://ewanwu06.github.io/fly-escape-switch/
连接组（connectome）是一张大脑的完整接线图：有哪些神经元，每个神经元连向谁，每条连接有多少个突触。2026 年，研究者正式发布了雄性果蝇整个中枢神经系统的连接组 MaleCNS v1.0，数据公开，任何人都能下载。A connectome is a complete wiring diagram of a brain: which neurons exist, which other neurons each one connects to, and how many synapses each connection has. In 2026 researchers released MaleCNS v1.0, the connectome of a male fruit fly's entire central nervous system, as open data that anyone can download.
有了接线图，一个很自然的想法是把它直接放进电脑“跑起来”，看它能不能产生果蝇的行为。网上已经有不少“果蝇大脑玩游戏”之类的演示。但有一个问题很容易被略过：就算模拟出来的网络能做事，这个能力究竟来自这张图里具体的连线方式，还是只来自一些更粗略的统计特征，比如每个神经元大概有多少条连接？With a wiring diagram in hand, a natural idea is to run it on a computer and see whether it produces fly behaviour. There are already many demos of "a fly brain playing games". One question is easy to skip, though. Even if a simulated network can do something, does that ability come from the specific way it is wired, or only from coarser statistics, such as how many connections each neuron has?
区分这两者有一个标准的对照实验，叫保度数重连（degree-preserving rewiring）。一个神经元的连接数叫它的“度”。保度数重连会随机打乱“谁连向谁”，但让每个神经元发出和收到的连接数都保持不变。如果打乱之后行为照旧，就说明这个行为并不依赖具体的接线。A standard control separates the two. It is called degree-preserving rewiring. A neuron's number of connections is its "degree". Degree-preserving rewiring shuffles who connects to whom at random, while every neuron keeps the same number of outgoing and incoming connections. If the behaviour survives the shuffle, it did not depend on the specific wiring.
打个比方An analogy
想象一座城市的公路网。保度数重连就像把所有路口重新接一遍：每个路口进出的路还是那么多条，但每条路通向哪里全变了。如果重接之后交通照样顺畅，那顺畅就不是靠原来那套具体的路线设计。Picture a city's road network. Degree-preserving rewiring reconnects every intersection: each one keeps the same number of roads in and out, but where each road leads changes completely. If traffic still flows as well afterwards, the flow never depended on the original street plan.
这个对照本身并不新：已经有项目在一整个果蝇大脑上做过一次“打乱或不打乱”的二选一对比。这个项目的不同在于三点：打乱的程度是逐步加大的；只针对一个小而清楚的回路；并且先用已发表的实验数据检验这个回路像不像真的。The control itself is not new. One earlier project ran a single all-or-nothing shuffle on a whole fly brain. This project differs in three ways: it scrambles the wiring in graded steps; it targets one small, well-understood circuit; and it first checks that circuit against published experiments.
许多演示把连接组当成一个固定不动的网络，再在外面接一个经过训练的“读出层”，真正学会任务的是这个读出层。这叫储备池计算（reservoir computing）。它能完成任务，不代表是接线本身产生了这个行为。保度数重连正好能区分两者：如果换成一个随机接线的网络也能做到，功劳就不在具体的接线上。Many demos treat the connectome as a fixed network and attach a trained "readout" layer on top; the readout is what learns the task. This is called reservoir computing. A network that can be trained to do something has not shown that its wiring produces the behaviour. Degree-preserving rewiring separates the two: if a randomly rewired network does just as well, the credit does not belong to the specific wiring.
果蝇左右脑各有一根巨纤维（giant fiber，细胞类型名 DNp01）。它是一根又粗又长的下行神经元，从大脑一直延伸到控制腿和翅膀的胸部神经节。它一发放，就会触发果蝇的快速起跳。Each side of a fly's brain has one giant fiber (cell type DNp01). It is a thick, long descending neuron that runs from the brain down to the thoracic ganglia, which control the legs and wings. When it fires, it triggers the fly's fast escape jump.
这段回路的输入来自眼睛。两类视觉神经元把信号送到巨纤维：The circuit's input comes from the eyes. Two kinds of visual neuron send signals to the giant fiber:
这些视觉神经元还会激活一批抑制性神经元，它们的作用像刹车，又连回巨纤维。所以视觉信号有两条路到达巨纤维：一条直接踩油门，一条先经过刹车。These visual neurons also excite a set of inhibitory neurons, which act like brakes and connect back onto the giant fiber. The visual signal therefore reaches the giant fiber by two routes: one presses the accelerator directly, the other goes through the brakes first.
2019 年，Ache 等人在另一只果蝇（一只雌蝇的全脑电镜数据 FAFB）里，数过连到右侧巨纤维的 LC4：55 个。从 MaleCNS 这只雄蝇里数出来的也是 55 个，突触数只多 5.7%（2,580 对 2,442）。性别不同、个体不同、数据集不同，数字对上了。LPLC2 的数字差得多一些（84 个对 108 个），原因还不清楚。In 2019, Ache and colleagues counted the LC4 neurons connected to the right giant fiber in a different fly (FAFB, whole-brain electron microscopy of a female). They found 55. Counting in the male MaleCNS fly also gives 55, with only 5.7% more synapses (2,580 against 2,442). Different sex, different individual, different dataset, same number. The LPLC2 counts differ more (84 against 108), for reasons not yet known.
果蝇实验里常用的“威胁”，是屏幕上一个不断变大的黑色圆盘，模拟一个直冲过来的物体。它在果蝇眼里有多大，用张角 θ 表示：物体两侧边缘对眼睛张开的角度。The usual "threat" in fly experiments is a black disk on a screen that keeps growing, imitating an object on a collision course. How big it looks to the fly is its angular size θ: the angle between the object's two edges as seen from the eye.
θ(t) = 2 · arctan( r / (v · |t|) )
这里 r 是物体半径，v 是速度，t 是距离碰撞还有多久。决定逼近节奏的只有 r/v 这一个比值，单位是毫秒：r/v 越小，逼近越快、越突然。已发表的实验和这个项目都用了 10、20、40、80 ms 四种，圆盘都从 10° 长到 63°。Here r is the object's radius, v its speed and t the time left before collision. Only the ratio r/v, in milliseconds, sets the pace: the smaller r/v, the faster and more sudden the approach. The published experiments and this project both use 10, 20, 40 and 80 ms, with the disk growing from 10° to 63°.
注意曲线的形状：开始时张角几乎不动，最后一刻才猛地变大。LC4 的输入跟着张角变大的速度走，所以在最后时刻最强，而且逼近越快越强。LPLC2 的输入在张角约 42° 时最强，和速度无关。Notice the shape of the curve: the angle barely moves at first and shoots up only at the very end. LC4's input follows how fast the angle grows, so it is strongest at the end, and stronger for faster approaches. LPLC2's input peaks when the angle is about 42°, whatever the speed.
视觉输入怎么编码，是这个模型里唯一的人为假设，来自 Ache 等人 2019 年的实验结论：LC4 的输入与张角变化速度成正比，LPLC2 的输入是以 42° 为中心的钟形曲线。钟形曲线的宽度 σ = 5°，是用第一项基准测试 B1 校准的，之后冻结不再改动。另外，每个视觉神经元的输入都加了最多 ±10 ms 的随机时间偏移，因为真实的神经元不会像阅兵一样完全同步（第 10 章会讲为什么必须加这一步）。从视觉神经元到巨纤维之间怎么整合，全部交给连接组里测出来的连接。How the visual input is encoded is the model's one explicit assumption, taken from the findings of Ache et al. (2019): LC4's input is proportional to how fast the angle grows, and LPLC2's input is a bell curve centred on 42°. The bell curve's width, σ = 5°, was calibrated once against the first benchmark, B1, and then frozen. Each visual neuron's input is also shifted by a random offset of up to ±10 ms, because real neurons do not fire in perfect lockstep (chapter 10 explains why this step was needed). Everything between the visual neurons and the giant fiber is left to the measured connections of the connectome.
每个神经元都用漏积分-发放模型（LIF）模拟。可以把它想成一个会漏水的水桶：Every neuron is simulated with the leaky integrate-and-fire (LIF) model. Think of it as a leaky bucket:
模型里的数值都是真实单位：静息电位 −52 mV，阈值 −45 mV，所以这个水桶只有 7 mV 深；漏水的时间常数是 20 ms；每个突触传来 0.275 mV。一个 LC4 连到巨纤维平均约有 50 个突触，所以单个 LC4 发放一次还远远不够，要好几个几乎同时发放才行。这些参数直接照搬一个已发表的果蝇全脑模型（Shiu 等，2024，《自然》），没有为这个项目调过。The numbers are in real units: rest at −52 mV and threshold at −45 mV, so the bucket is only 7 mV deep; the leak's time constant is 20 ms; each synapse contributes 0.275 mV. One LC4 neuron connects to a giant fiber through about 50 synapses on average, so a single LC4 spike is far from enough: several have to arrive almost together. These parameters are copied from a published whole-brain fly model (Shiu et al., 2024, Nature) and were not tuned for this project.
关键在于：模型里没有任何“逃跑程序”。视觉信号怎样汇总到巨纤维，完全由连接组里测出来的连接和突触数决定。The key point: there is no "escape program" anywhere in the model. How the visual signals add up at the giant fiber is set entirely by the connections and synapse counts measured in the connectome.
在跑任何模拟之前，研究设计就已写好并提交到版本库（git），每次提交都带着时间戳：要检验哪几项、用什么指标、怎样算通过，以及结果出来后每一种情况分别怎么解读。这叫预注册。Before any simulation ran, the study design was written down and committed to version control (git), where every commit carries a timestamp: which results would be tested, with which measures, what counts as a pass, and how each possible outcome would be read. This is called pre-registration.
这是为了防止一个常见的问题：看到结果之后再回头调整标准，让结果“看起来”更好。预注册之后，任何改动都必须留下带日期的公开记录，并注明发生在看到结果之前还是之后。It guards against a common problem: adjusting the standards after seeing the results, so the results look better. After pre-registration, every change has to leave a dated public record that says whether it happened before or after the results were seen.
时间为 git 提交记录中的时间。Times are as recorded in the git commits.
从已发表的电生理实验里挑出四项结果，作为这个虚拟回路的“考试”，编号 B1–B4。通过标准在正式运行这些测试之前就写进了代码并提交。Four results from published electrophysiology experiments serve as the virtual circuit's "exams", numbered B1 to B4. The pass criteria were written into code and committed before the tests were formally run.
σ 正是用这一项校准的，所以它不是完全独立的检验。把输入增益调高到 1.2 倍或更高，最快速度下的结果会跌出范围。σ was calibrated on this test, so it is not fully independent. Raise the input gain to 1.2 times or more and the fastest speed falls out of the band.
代码只检查“反应的峰值时间有没有移动至少 2 ms”。巨纤维没有反应时，这一条被平凡地满足了。登记的标准还要求移动方向正确，代码没有检查方向，而实际方向是反的。The code only checks whether the response's peak time moves by at least 2 ms. With no response at all, that is met trivially. The registered criterion also requires the right direction; the code never checks it, and the shift runs the wrong way.
代码里的指标把两根巨纤维平均起来，只看到第一次发放为止，所以它量到的其实是两根巨纤维有多同步。它显示的“亚线性”（指数 0.81）就是这么来的。The coded measure averages the two giant fibers and stops at the first spike, so what it actually reads is how closely the two fire together. That is where its "sublinear" index of 0.81 comes from.
四项里只有 B1 真正复现了。对 B2–B4 来说，是模型本身不像真果蝇，这和接线有没有被打乱无关。Only B1 was truly reproduced. For B2 to B4, the model itself does not behave like a real fly, before any wiring is scrambled.
下表是一次事后检查（不属于预注册内容）中每根巨纤维在第一次发放前的最大去极化（相对静息电位），三次运行共用同一套随机输入。每根巨纤维只对同侧的眼有反应，双眼时的数值与只看同侧眼时完全相同。The table shows each giant fiber's maximum depolarisation (relative to rest) before its first spike, from a post hoc check that is not part of the pre-registration; the three runs share one input draw. Each giant fiber responds only to its own eye, and its value with both eyes equals its value with its own eye alone.
现在进入核心问题。打乱接线用的是双边交换：随机挑两条连接 A→B 和 C→D，把它们改成 A→D 和 C→B。这样一来，每个神经元发出的连接数和收到的连接数都没变，每条连接的突触数和正负号也跟着发出它的神经元走，变的只是“谁连向谁”。Now for the main question. The wiring is scrambled by double-edge swaps: pick two connections A→B and C→D at random, and turn them into A→D and C→B. Every neuron keeps its number of outgoing and incoming connections, and each connection keeps its synapse count and sign, which travel with the neuron that sends it. Only who connects to whom changes.
打乱的程度用剂量 f 表示：交换次数 = f × 连接总数 ÷ 2。f = 1 时，平均每条连接会被动到一次。但因为每次都是随机挑选，有的连接被挑中好几次，有的一次也没被挑中，所以最后仍有约 41% 的原始连接保留。How much the wiring is scrambled is set by the dose f: number of swaps = f × number of connections ÷ 2. At f = 1 each connection is touched once on average. Because the picks are random, though, some connections are picked several times and others never, so about 41% of the original connections survive.
实验做法：在 6 个剂量下（f = 0、0.05、0.10、0.25、0.50、1.00）各生成 100 个打乱的回路，全部用同一套输入、同一种逼近速度（r/v = 40 ms）模拟。然后看真实回路的反应，在这 100 个打乱回路中间是不是显得特别。The experiment: at each of 6 doses (f = 0, 0.05, 0.10, 0.25, 0.50, 1.00), 100 rewired circuits were generated and all were simulated with the same input at one approach speed (r/v = 40 ms). The question is whether the real circuit's response stands out among those 100.
p 值是“如果真实回路只是这 100 个打乱回路中普通的一员，看到至少这么大差别的概率”。经过 Benjamini–Hochberg 多重比较校正后，没有一项显著；即使不校正，也没有一项低于 0.05。f = 0 时“打乱”的回路和真实回路完全一样，所以 p = 1。四个指标并不独立（发放时的张角和峰值时间几乎同步变化），所以这 24 项实际上远少于 24 个独立检验。A p-value is the probability of seeing a difference at least this large if the real circuit were just an ordinary member of its 100 rewired neighbours. After Benjamini–Hochberg correction for multiple comparisons none is significant, and none falls below 0.05 even uncorrected. At f = 0 the "rewired" circuits are identical to the real one, so p = 1. The four measures are not independent (angle at first spike and peak time move almost in lockstep), so the 24 tests amount to far fewer than 24 independent ones.
下面是模拟的原始数据。选一个逼近速度和打乱剂量，看两根巨纤维的平均膜电位怎样随时间变化；把鼠标放在图上读数，或点播放，小屏幕会同步显示果蝇此刻看到的画面。每个剂量这里只展示一个打乱回路作为例子，它和上一节矩阵里显示的是同一个回路；统计结果以上一节为准。这里用的随机输入和上一节不是同一套，所以真实回路的数字会和上一节差零点几度。These are the simulation's raw traces. Pick an approach speed and a rewiring dose to see how the two giant fibers' mean membrane potential changes over time. Hover to read values, or press play, and the small screen shows what the fly sees at that moment. Each dose shows one rewired circuit as an example, the same one shown in the matrix above; the statistics are in the previous chapter. The random input here is a different draw from the previous chapter's, so the real circuit's numbers differ from those by a few hundredths to tenths of a degree.
在跑实验之前，研究设计就写好了三种可能的结果以及各自的解读。结果落在下表的哪一行：Before the experiment ran, the design listed three possible outcomes and how each would be read. Here is where the results fell:
| 情况Outcome | 预先写好的解读Reading fixed in advance | 落在这里Lands here |
|---|---|---|
| 真实回路通过，打乱后不通过Real circuit passes, rewired ones do not | 具体接线是必要的（结构特异性）The specific wiring is needed (structure specificity) | 无none |
| 打乱后的回路同样通过Rewired circuits pass too | 保持每个神经元的连接数就够了Keeping each neuron's number of connections is enough | B1 |
| 真实回路本身就不通过The real circuit itself fails | 阴性结果：“连接组 + LIF”不足以解释这些反应A negative result: connectome + LIF cannot explain these responses | B2 B3 B4 |
研究很少一次就做对。这个项目把每一处更正都公开记录下来，并注明发生在看到结果之前还是之后。下面是几个例子。Research is rarely right the first time. This project records every correction in public and notes whether it came before or after the results were seen. A few examples:
重新运行之前Before the re-run
第一次运行时，311 个视觉神经元收到几乎一样的输入，像阅兵一样完全同步地发放。巨纤维因此从静息状态一步跨过阈值，好几个指标在所有条件下都恒为 0，根本无法检验。修正方案（给每个视觉神经元加上随机时间偏移）先写进设计并提交，然后才重新运行。In the first run, the 311 visual neurons received almost identical input and fired in perfect lockstep, like a parade. The giant fiber jumped from rest past threshold in a single step, and several measures read zero in every condition, so nothing could be tested. The fix, a random time offset for each visual neuron, was written into the design and committed first; only then was the test re-run.
任何测试之前Before any test ran
LC4 的输入原本在每种逼近速度内部各自归一化，把每种速度的峰值都压成 1。这恰好抹掉了 B2 要检验的东西：快速逼近时 LC4 信号本应比慢速强 8 倍。改成统一的参照之后，才开始任何测试。LC4's input was first normalised separately within each approach speed, which pinned every speed's peak to 1. That erased exactly what B2 tests: for the fastest approach the LC4 signal should be 8 times stronger than for the slowest. It was switched to one shared reference before any test ran.
全部结果之后After all results
终审发现，“最大去极化”这个指标把两根巨纤维平均起来，并在任意一根发放时截止。一旦有一根先发放，它量到的就是两根巨纤维有多同步，而不是反应有多大。B4 的“通过”就是这么来的。已登记的判定没有重新打分，报告如实说明了它们的真实含义，并补做了一项标明为“事后”的逐根检查。The final review found that the "maximum depolarisation" measure averages the two giant fibers and stops as soon as either one fires. Once one fires first, it reads how closely the two fire together, not how large the response is. That is where B4's "pass" came from. The registered verdicts were not re-scored; the report states what they actually mean and adds a per-fiber check labelled post hoc.
全部结果之后After all results
用来验证提取结果的“55 个 LC4”，最初被标成来自 hemibrain 数据集。实际上它来自 Ache 等人 2019 年在 FAFB 数据集上的重建，而那篇论文比 hemibrain 发布得还早。基于这个错误标签提出的一个解释也一并撤回。数字本身没有变。The "55 LC4 neurons" used to validate the extraction were first labelled as hemibrain counts. They actually come from Ache et al.'s 2019 reconstruction in the FAFB dataset, a paper that predates the hemibrain's release. An explanation built on the wrong label was withdrawn with it. The numbers themselves did not change.
全部结果之后披露Disclosed after all results
输入增益是把视觉信号换算成输入电流的系数。在它固定之前，为了排查“整个回路毫无反应”，曾做过一次增益 × σ 的扫描，瞥见了部分 B1 的结果，随后停止，结果也没有保存。这件事在全部结果出来之后公开披露。复核还发现，按当初写下的规则，增益应该是 7.0（或 6.0），而不是实际使用的 8.0。为此补做了一项事先写好方案的增益敏感性分析：B1 在 0.6–1.0 倍增益下通过，在 1.2 和 1.4 倍时不通过。The input gain is the factor that turns the visual signal into input current. Before it was fixed, a gain × σ scan was run while working out why the whole circuit was silent. It glimpsed some B1 outcomes, was stopped, and its results were not kept. This was disclosed after all results were in. A re-check also found that the rule written down for the gain gives 7.0 (or 6.0), not the 8.0 actually used. A gain sensitivity analysis, specified before it ran, was added: B1 passes at 0.6–1.0 times the gain and fails at 1.2 and 1.4 times.