5D Library › 双过程睡眠架构、慢波胶质淋巴清除与体温节律相位控制 · English

第二章 睡眠开关:为什么衰老先让夜晚破碎,再让它变短

是开关,不是调光器

过程 S 与过程 C 说明睡眠何时应当开始;它们没有说明大脑如何完成这个切换。Saper 等人提出,切换由两组相互抑制的神经元完成:下丘脑腹外侧视前核(VLPO)中的睡眠活跃细胞,以及脑干与下丘脑中促觉醒的单胺能细胞群 (Saper et al. 2001, 2010)。相互抑制构成一个触发器(flip-flop):哪一边活跃,就压制另一边,于是系统停在两种状态之一,在两者之间切换得很快,中间停留的时间很短。下丘脑外侧的食欲素(下丘脑分泌素)神经元兴奋觉醒一侧,被认为能稳定它——就像一根按在开关上的手指。

这一设想对开关变弱时会发生什么作出了尖锐的预测。它主要改变的不应是我们睡多少,而应是我们在每种状态中停留得多稳。

推导

模型。 令 xx 为睡眠活跃群体、yy 为觉醒活跃群体(都缩放到 0 到 1), dxdt=−x+f(c+D−wy),\frac{dx}{dt} = -x + f(c + D – w\,y), dydt=−y+f(c−D−wx+Oy),\frac{dy}{dt} = -y + f(c – D – w\,x + O\,y), 其中 f(u)=1/(1+e−u/0.1)f(u) = 1/(1+e^{-u/0.1}),c=0.5c = 0.5 为紧张性兴奋,DD 为趋向睡眠的净驱动(来自过程 S 与 C),ww 为相互抑制的强度,OO 为食欲素在觉醒一侧活跃时对它的加强。

开关有多少余量

第三段程序对每一种相互抑制强度,问在多大的驱动范围内开关是双稳态的——能保持任一状态,使驱动的小变化不会把它翻转。

推导

# Chapter 2. The sleep switch as two mutually inhibiting populations (a
# minimal "flip- flop" in the sense of Saper and colleagues). x = sleep-active
# VLPO population, y = wake- active monoaminergic population, both scaled
# 0..1:
#   dx/dt = -x + f(c + D - w y)
#   dy/dt = -y + f(c - D - w x + O y)
# c = 0.5 is tonic excitation, D is the net drive toward sleep (homeostatic
# minus circadian), w is the strength of the mutual inhibition, O is orexin's
# reinforcement of the wake side while it is active, and f(u) = 1 / (1 +
# exp(-u / 0.1)). For each D on a grid (step 0.005) we let the network settle
# from a wake start and from a sleep start. Where the two settle to different
# states the switch is bistable: it holds whichever state it is in. The range
# of D over which that is true is the switch's margin against changes in
# drive.
import math

def f(u):
    return 1.0 / (1.0 + math.exp(-u / 0.1))

def settle(D, w, O, x, y, steps=4000, dt=0.02):
    for _ in range(steps):
        dx = -x + f(0.5 + D - w * y)
        dy = -y + f(0.5 - D - w * x + O * y)
        x += dt * dx
        y += dt * dy
    return x, y

def bistable_range(w, O):
    lo = hi = None
    for i in range(-400, 401):
        D = 0.005 * i
        xa, ya = settle(D, w, O, 0.0, 1.0)
        xb, yb = settle(D, w, O, 1.0, 0.0)
        if abs(xa - xb) > 0.5:
            lo = D if lo is None else lo
            hi = D
    return lo, hi

if __name__ == "__main__":
    print("Range of sleep drive D over which the switch is bistable")
    print(
        "   inhibition w   orexin O   bistable from D   to D      margin"
        " (width)"
    )
    for O in (0.0, 0.3):
        for w in (1.5, 1.0, 0.7, 0.5, 0.3):
            lo, hi = bistable_range(w, O)
            if lo is None:
                print(
                    "   {:8.1f}       {:5.1f}       none: one state only     "
                    "      0.000".format(w, O)
                )
            else:
                print(
                    "   {:8.1f}       {:5.1f}       {:+8.3f}        {:+8.3f} "
                    "   {:7.3f}".format(w, O, lo, hi, hi - lo + 0.005)
                )
    print(
        "\nReading: strong mutual inhibition gives a wide margin; weakening"
        " it (as with"
    )
    print(
        "loss of VLPO neurons) narrows the margin until the switch no longer"
        " holds a"
    )
    print(
        "state at all. Orexin moves the margin toward higher sleep drive: the"
        " wake"
    )
    print(
        "state needs more pressure to give way. Without it, wake gives way"
        " sooner."
    )
# Range of sleep drive D over which the switch is bistable
#    inhibition w   orexin O   bistable from D   to D      margin (width)
#         1.5         0.0         -0.325          +0.325      0.655
#         1.0         0.0         -0.180          +0.180      0.365
#         0.7         0.0         -0.055          +0.055      0.115
#         0.5         0.0       none: one state only           0.000
#         0.3         0.0       none: one state only           0.000
#         1.5         0.3         -0.320          +0.505      0.830
#         1.0         0.3         -0.160          +0.300      0.465
#         0.7         0.3         +0.010          +0.155      0.150
#         0.5         0.3       none: one state only           0.000
#         0.3         0.3       none: one state only           0.000
#
# Reading: strong mutual inhibition gives a wide margin; weakening it (as with
# loss of VLPO neurons) narrows the margin until the switch no longer holds a
# state at all. Orexin moves the margin toward higher sleep drive: the wake
# state needs more pressure to give way. Without it, wake gives way sooner.

抑制较强时(w=1.5w = 1.5),驱动在 −0.325 到 +0.325 之间开关都能保持任一状态,余量为 0.655。减弱抑制,余量迅速缩小:w=1.0w = 1.0 时为 0.365,w=0.7w = 0.7 时为 0.115;到 w=0.5w = 0.5 时开关已不再是开关——每个驱动水平只有一个状态,系统随驱动变化在「睡着」与「醒着」之间滑动。

食欲素改变的是余量的形状,而不只是大小。O=0.3O = 0.3 时,上端从 +0.325 移到 +0.505:觉醒状态需要多得多的睡眠驱动才会让位,而下端几乎不动(−0.320)。这就是「按在开关上的手指」:食欲素守卫着清醒。

噪声与破碎的夜晚

真实的开关是有噪声的。第四段程序让开关度过一个 8 小时的夜晚,睡眠驱动随过程 S 衰减而下降,并给每个群体加上噪声。

推导

# Chapter 2. The same switch with noise, over one 8-hour night. Sleep drive is
# high at lights-out and falls through the night as Process S decays: D(t) =
# 0.45 x exp(-t/4.2)
# - 0.05 (t in hours). Each population receives independent noise
#   (Ornstein-Uhlenbeck,
# sd 0.12, correlation time 30 s). "Asleep" means x > y. A wake bout of at
# least 15 s counts as an awakening. Weaker mutual inhibition w stands for
# loss of VLPO neurons. 200 simulated nights per row, fixed seed; dt = 1 s.
import math
import random

def f(u):
    return 1.0 / (1.0 + math.exp(-u / 0.1))

def night(w, O, rng, sd=0.12, tau_n=30.0, tau=10.0):
    x, y = 1.0, 0.0
    nx = ny = 0.0
    a = math.exp(-1.0 / tau_n)
    b = sd * math.sqrt(1 - a * a)
    asleep_s, awakenings, wake_run, state = 0, 0, 0, True
    for s in range(8 * 3600):
        t = s / 3600.0
        D = 0.45 * math.exp(-t / 4.2) - 0.05
        nx = a * nx + b * rng.gauss(0, 1)
        ny = a * ny + b * rng.gauss(0, 1)
        x += (-x + f(0.5 + D - w * y + nx)) / tau
        y += (-y + f(0.5 - D - w * x + O * y + ny)) / tau
        if x > y:
            asleep_s += 1
            if not state and wake_run >= 15:
                awakenings += 1
            state, wake_run = True, 0
        else:
            state = False
            wake_run += 1
    return asleep_s / (8 * 3600), awakenings

if __name__ == "__main__":
    rng = random.Random(57004)
    print("   mean of 200 nights")
    print("   inhibition w   orexin O   time asleep   awakenings per night")
    for O in (0.3, 0.0):
        for w in (1.5, 1.0, 0.7, 0.5):
            res = [night(w, O, rng) for _ in range(200)]
            fa = sum(r[0] for r in res) / len(res)
            aw = sum(r[1] for r in res) / len(res)
            print(
                "   {:8.1f}       {:5.1f}       {:7.3f}          {:7.2f}"
                .format(w, O, fa, aw)
            )
    print("\nReading (orexin O = 0 rows): weakening the"
          " mutual inhibition from 1.0 to")
    print("0.5 costs less than a tenth of the time asleep but multiplies the")
    print("awakenings - the night fragments long before"
          " it shortens. With orexin left")
    print("active at night (O = 0.3 rows) the wake side"
          " is held too firmly and sleep")
    print("is lost outright.")
#    mean of 200 nights
#    inhibition w   orexin O   time asleep   awakenings per night
#         1.5         0.3         1.000             0.00
#         1.0         0.3         0.933             0.01
#         0.7         0.3         0.640            13.29
#         0.5         0.3         0.546            59.19
#         1.5         0.0         1.000             0.00
#         1.0         0.0         0.989             0.12
#         0.7         0.0         0.938            10.61
#         0.5         0.0         0.908            35.40
#
# Reading (orexin O = 0 rows): weakening the mutual inhibition from 1.0 to
# 0.5 costs less than a tenth of the time asleep but multiplies the
# awakenings - the night fragments long before it shortens. With orexin left
# active at night (O = 0.3 rows) the wake side is held too firmly and sleep
# is lost outright.

没有食欲素的那几行(O=0O = 0,相当于睡眠中食欲素神经元安静的情形)包含本章的主要结果。把相互抑制从 1.0 减弱到 0.7 和 0.5,睡着的时间从 0.989 降到 0.938 和 0.908——不到十分之一——而每夜的觉醒次数从 0.12 升到 10.61 和 35.40。夜晚在变短之前很久就先破碎了。这正是老年人睡眠的模式:总睡眠时间适度下降,而入睡后觉醒与分期转换显著增多 (Ohayon et al. 2004)。

夜里食欲素仍然活跃的那几行(O=0.3O = 0.3)展示了另一面:觉醒一侧被加强时,变弱的开关会直接失去睡眠(只睡夜晚的 0.640 和 0.546)。在这个模型里,食欲素夜里的工作就是关掉。

双稳态余量(开关能保持任一状态的睡眠驱动范围)随相互抑制强度的变化,有无食欲素两种情形。右:无食欲素时,有噪声开关每夜的睡眠时间与觉醒次数。

证据

支持开关图景的证据。 对生前多年佩戴活动记录仪、死后做尸检的老年人的研究发现,VLPO/中间核神经元越少,睡眠越破碎 (Lim et al. 2014)。在人类发作性睡病中,大部分下丘脑分泌素(食欲素)神经元丧失 (Thannickal et al. 2000),而这种疾病的特征是双向的状态不稳定——睡眠侵入清醒、清醒侵入睡眠——而不只是睡得太多。反过来,阻断食欲素受体的药物被用于治疗失眠,在随机试验中相对安慰剂改善了睡眠 (Herring et al. 2016);这里提及只是为了说明药理,不是推荐。

假说 · 尚未验证

作者的假说(尚未验证): 在不同年龄的健康成年人中,若把两者都表示为最年轻组数值的百分比,每小时睡眠中的觉醒次数随年龄上升的幅度,大于总睡眠时间下降的幅度;在多年随访的个体中,觉醒次数的增加先于总睡眠可测的减少。什么结果会证明它错:两者的比例变化大小相近,或者总睡眠先下降。

小结

睡眠与清醒由两个相互抑制的群体维持,相互抑制构成一个有余量的开关:在一定驱动范围内,它保持自己所处的状态。减弱抑制——如衰老中 VLPO 神经元的丧失——会缩小余量,直到开关不再是开关。加上噪声后,这几乎不损失总睡眠,却使觉醒成倍增加:模型中每夜从 0.12 次增到 35.40 次,而睡着的时间下降不到十分之一。食欲素守卫清醒状态;在这个模型里,它夜里需要关掉。

参考文献

Herring, W. Joseph, Kathryn M. Connor, et al. 2016. “Suvorexant in Patients with Insomnia: Results from Two 3-Month Randomized Controlled Clinical Trials.” Biological Psychiatry 79 (2): 136–48.
Lim, Andrew S. P., Brian A. Ellison, et al. 2014. “Sleep Is Related to Neuron Numbers in the Ventrolateral Preoptic/Intermediate Nucleus in Older Adults with and without Alzheimer’s Disease.” Brain 137 (10): 2847–61.
Ohayon, Maurice M., Mary A. Carskadon, Christian Guilleminault, and Michael V. Vitiello. 2004. “Meta-Analysis of Quantitative Sleep Parameters from Childhood to Old Age in Healthy Individuals: Developing Normative Sleep Values across the Human Lifespan.” Sleep 27 (7): 1255–73.
Saper, Clifford B., Thomas C. Chou, and Thomas E. Scammell. 2001. “The Sleep Switch: Hypothalamic Control of Sleep and Wakefulness.” Trends in Neurosciences 24 (12): 726–31.
Saper, Clifford B., Patrick M. Fuller, Nigel P. Pedersen, Jun Lu, and Thomas E. Scammell. 2010. “Sleep State Switching.” Neuron 68 (6): 1023–42.
Thannickal, Thomas C., Robert Y. Moore, et al. 2000. “Reduced Number of Hypocretin Neurons in Human Narcolepsy.” Neuron 27 (3): 469–74.

Fan, Y. (2026). Two-Process Sleep Architecture, Slow-Wave Glymphatic Clearance and Thermoregulatory Phase Control: Adenosine Homeostasis, Flip-Flop Switch Bistability, Core-Distal Heat Flux and Sleep-Stage Markov Chains [双过程睡眠架构、慢波胶质淋巴清除与体温节律相位控制] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090162 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).