在一段 DNA 上换掉一个字母，有时候什么都不会发生。不是影响很小，是字面意义上的没有：造出来的蛋白质一模一样，细胞根本不知道这儿动过。
这种改动也不是边角料。基因组里有大片区域，你怎么改它都不吭声。
于是有个问题躲不掉：这些没人管的改动，最后是怎么定下来的？
先说为什么会有「换了等于没换」这种事。DNA 上的字母三个一组被读出来，一组指定一种氨基酸——氨基酸就是拼成蛋白质的零件。四种字母三个一组共 64 种组合，而要指定的氨基酸只有 20 种，外加一个「到此为止」的信号。编码比要编的东西多，多出来的就成了同义词。好几种组合指向同一个氨基酸，差别常常只在第三个字母上。第三个字母换掉，蛋白质分毫不差。
「看不见」在这里有个很硬的定义：一个改动如果对携带它的个体能留下多少后代毫无影响，那它在选择面前就是透明的。不是选择宽容，是压根没有可挑的差别——选择挑的从来不是「好基因」，是「留下后代多的那些个体」；改动如果不落在这个数上，它就不在场。
还有更彻底的。基因组里造蛋白质的部分只占很小一块，剩下大片区域躺着各种残骸，最干净的一类叫假基因：某个基因半路坏了，再也造不出能用的东西，尸体却留在原地继续变，没有谁在乎。
证据摆在任何两个亲缘物种的基因组对比里。把同一个基因在两边排齐了数差异：「换了字母但氨基酸没变」的位置，差异明显比「氨基酸也跟着变」的位置多，不挑基因、不挑物种。而假基因积累差异最快，几乎等于突变本身发生的速度——那儿真的一个把关的都没有。
EN Start with why "changed but unchanged" is possible at all. DNA letters are read three at a time, and each triplet specifies one amino acid — amino acids being the parts that proteins are built from. Four letters taken three at a time gives 64 combinations, while there are only 20 amino acids to specify plus a "stop here" signal. There are more codes than things to encode, and the surplus turns into synonyms: several combinations point at the same amino acid, usually differing only in the third letter. Swap the third letter and the protein is identical. "Invisible" has a hard definition here: if a change makes no difference at all to how many offspring its carrier leaves, it is transparent to selection. Selection is not being lenient; there is simply no difference to pick between. Selection never picks "good genes" — it picks the individuals who leave more offspring, and a change that does not land on that number is not in the room. There is something more thorough than third letters, too. Only a small slice of the genome codes for protein; the rest holds wreckage of various kinds, the cleanest sort being pseudogenes — a gene that broke partway and can no longer make anything usable, whose corpse stays where it was, changing letter by letter, with nobody caring. The evidence sits in any comparison between two related genomes. Line the same gene up on both sides and count: positions where a letter changed but the amino acid did not carry visibly more differences than positions where the amino acid changed too. The gap shows up everywhere, in every gene and every species. And pseudogenes accumulate differences fastest of all, at nearly the rate at which mutations themselves occur — because there really is nothing filtering them.
想象一个几十户人家的村子。几代人之后去看，大部分姓氏都不见了，剩下那几个占了大半个村。没有哪个姓更优秀，只是每家生几个、是儿是女全带着运气，而一个姓只要有一代没人接上就永远回不来。1870 年代 Galton 和 Watson 算的正是这件事。
基因的频率也这么走。每一代都是从上一代里抽一批出来当父母，抽样就有偏差，某个改动的比例会一年年随机地晃，晃到 0（消失）或者 100%（全群都带上了）才停。这就是遗传漂变。村子越小，每代偏差越大，晃得越快；大城市里姓氏几百年也灭不掉几个。
接下来是这套想法里最漂亮的一步。一个中性突变刚冒出来时全群只有一份，它最终铺满整个群体的机会很小，群体越大机会越小。但反过来，群体越大，每代新冒出来的突变总数也越多。两边正好抵消：中性改动被定下来的速度，就等于中性突变本身出现的速度，跟这个物种有多少个体完全无关。
1968 年木村资生把这条写出来，顺手解掉了当时一个很难受的疙瘩。那几年刚测出蛋白质变化的速度，快得离谱；如果每次变化都得靠选择推上去，按 Haldane 早先算的账，群体每代为此损失的个体多到根本撑不住。中性的改动不欠这笔账——没有谁为它死，它自己就铺开了。第二年 King 和 Jukes 用「非达尔文式演化」这个挑衅的标题把同一件事又说一遍。
「速度是常数」还有个用处大得多的后果：差异可以当尺子量时间。Zuckerkandl 和 Pauling 在 1960 年代初就注意到，两个物种的同一种蛋白差多少个零件，大致跟它们分家多久成正比。这就是分子钟，「寒武纪大爆发」那期里化石记录和分子钟吵的那一架，吵的就是这把尺子准不准。诚实地说它是把软尺：不同谱系走得快慢不一样，一代人隔多久、修复机器好不好使都掺在里面，实用时必须拿化石校准。说它是钟不如说是沙漏——总体稳，单次读数糙。
EN Picture a village of a few dozen households. Look a few generations later and most of the surnames are gone, with the surviving few covering half the village. No surname was better; it is just that how many children each family has, and whether they are sons or daughters, all carry luck — and one generation with nobody to carry it and a name is gone for good. This is exactly what Galton and Watson worked out in the 1870s, worried about why aristocratic names kept dying out. Gene frequencies walk the same way. Every generation a batch is drawn from the last to be the parents, drawing has sampling error, and the proportion carrying some change wanders at random year after year until it hits 0 (gone) or 100% (everyone carries it). That is genetic drift. The smaller the village, the larger each generation's sampling error, and the faster the wander; in a big city hardly any surname dies out in centuries. Then comes the prettiest step in the whole idea. A neutral mutation starts as a single copy, so its chance of eventually filling the whole population is small — and the larger the population, the smaller that chance. But equally, the larger the population, the more new mutations appear each generation. The two cancel exactly: the rate at which neutral changes get fixed equals the rate at which neutral mutations arise, with no dependence whatsoever on how many individuals the species has. Motoo Kimura wrote this down in 1968, and in passing it dissolved a very uncomfortable knot. Rates of protein change had just been measured and were absurdly fast; if every change had to be pushed through by selection, then on Haldane's earlier accounting the losses a population must absorb each generation to pay for it were far more than it could bear. Neutral changes owe no such bill — nothing dies for them, and they spread on their own. The following year King and Jukes said the same thing again under the deliberately provocative title "Non-Darwinian Evolution." That constancy has a far more useful consequence: differences can be used as a ruler for time. Zuckerkandl and Pauling noticed in the early 1960s that how many parts the same protein differs by between two species is roughly proportional to how long ago they split. That is the molecular clock. The argument between the fossil record and the molecular clock in the piece on the Cambrian explosion is an argument about how accurate this ruler is. Honestly, it is a soft ruler. Different lineages run at different speeds — generation length, how well the repair machinery works, body size all feed in — so in practice it has to be calibrated against fossils, and the dates it reports come with wide ranges. Less a clock than an hourglass: steady overall, coarse on any single reading.
先把最常见的误读挡掉。中性理论从来没说演化里没有选择。眼睛、翅膀、细菌几年里长出来的抗药性，没有一样是随机攒出来的。它说的只是：在 DNA 这张账单上，被记下来的行里大多数不影响死活。适应仍然只有选择解释得了。混起来就成了「达尔文被推翻了」这种标题。
真正有意思的是中间地带。1973 年太田朋子指出，一个改动不必非黑即白——它可以只有一点点坏。而「一点点坏」算不算坏，取决于它周围有多少个体。如果这点坏处小到跟每代抽样的随机偏差差不多，在小群体里它就跟中性的没区别，照样被漂变推上去；换到大群体，抽样噪声小，同样这点坏处就露出来了，会被清掉。这叫近中性理论，它给出一个能查的预测：群体小的物种，蛋白质里会攒下更多轻微有害的改动。去查，方向对得上——拿「改了氨基酸的差异」和「没改氨基酸的差异」的比值在物种间比，果蝇这种个体数以百万计的比值低，人和猿这种历史上群体一直很小的比值高。
另一头，选择留下的指纹也清楚得很。人身上最干净的例子是消化牛奶：绝大多数哺乳动物断奶后就把分解乳糖的酶关掉，一部分人群里有个把这开关继续开着的变异，几千年里从罕见变成了常见。更能说明问题的是它周围——那一大段 DNA 被整段带了上去，段内多样性被压得特别平。这是快速选择留下的形状，慢悠悠的随机漂变做不出来。
于是问题变成量的问题：被固定下来的氨基酸改动里，多大比例是被选择推上去的？有一类方法把「种内还剩多少变异」和「种间已经差了多少」对着看：2002 年 Smith 和 Eyre-Walker 用它估果蝇，得出接近一半；同样思路用在人和猿身上估出来低得多，有的版本接近零。这类估计对方法极其敏感，不同团队差出一倍以上很常见，这块还在争。更麻烦的是选择会牵连邻居：它作用在一个位置上，会把旁边一整段 DNA 一起拖着走，于是本来纯中性的邻居看上去也不像纯随机的了。基线本身被选择污染了，这是今天这场争论真正落脚的地方。
所以中性理论今天的位置不是「跟选择对立的那一派」，它是零模型：先算出「假如一点选择都没有，数据该长成什么样」，再看真实数据在哪儿偏离。偏离的地方才是选择的证据。没有这条基线，「这里一定有原因」是一句永远推不翻的话。
EN First, block the commonest misreading. Neutral theory never said there is no selection in evolution. Eyes, wings, the drug resistance bacteria grow in a few years — none of that was assembled at random. What it says is only this: on the ledger that is DNA, most of the entries recorded do not affect survival. Adaptation is still explained by selection alone; selection just occupies far fewer lines of the ledger than intuition suggests. The two statements are about different things, and running them together produces headlines about Darwin being overturned. The interesting part is the middle ground. In 1973 Tomoko Ohta pointed out that a change need not be black or white — it can be only slightly bad. And whether "slightly bad" counts as bad at all depends on how many individuals surround it. If the harm is small enough to be comparable to each generation's random sampling error, then in a small population it is indistinguishable from neutral and drift can push it through anyway; move it to a large population, where sampling noise is small, and that same small harm becomes visible and gets cleared out. This is the nearly neutral theory, and it makes a testable prediction: species with small populations should accumulate more slightly harmful changes in their proteins. Go and check, and the direction matches. Compare, across species, the ratio of "differences that changed an amino acid" to "differences that did not": in fruit flies, with individuals in the millions, the ratio is low; in humans and apes, whose populations have historically been small, it is high. At the other end, the fingerprints of selection are perfectly clear too. The cleanest human example is digesting milk. Most mammals switch off the enzyme that breaks down lactose after weaning, so drinking milk as an adult upsets the gut; in some populations a variant that keeps the switch on went from rare to common within a few thousand years. What clinches it is the neighbourhood: the whole long stretch of DNA around that variant was carried up with it, leaving diversity within the stretch unusually flat. That is the shape a fast episode of selection leaves, and unhurried random drift cannot produce that shape. So the question becomes quantitative: of the amino acid changes fixed in a species, what fraction were pushed through by selection? One family of methods sets "how much variation is still segregating within the species" against "how much the species have already diverged." In 2002 Smith and Eyre-Walker applied this to fruit flies and concluded that close to half of amino acid substitutions were driven by positive selection; applied to humans and apes, the estimates come out far lower, some versions near zero. These estimates are extremely sensitive to method, and differences of twofold or more between groups are common. This part is still being argued. There is a thornier point of contention too: selection acting at one position drags a whole stretch of neighbouring DNA along with it — carried up when the change is good, swept away alongside when a bad one is cleared. So neighbours that were purely neutral stop looking purely random. The baseline itself is contaminated by selection, and that is where today's argument actually lands. Neutral theory's place now is not "the faction opposed to selection." It is the null model: work out what the data should look like if there were no selection at all, then see where real data departs from it. The departures are the evidence of selection. That is its most practical legacy — without that baseline, "there must be a reason here" is a sentence that can never be refuted.