5D Library › 亚知觉色胺药代动力学、5-HT2A 受体脱敏与神经可塑性窗口的生物数学模型 · English
第三章 药代—药效耦合:知觉阈不是一个数
效应室
血浆浓度与效应在时间上不同步:效应的峰通常晚于浓度的峰,而且在同一浓度下,上升段与下降段的效应不同。处理这件事的标准办法是效应室模型。
证据
效应室(方法学原始文献)。 Sheiner 等人 1979 年提出把药代与药效同时建模,引入一个与血浆之间以一阶速率 交换的假想「效应室」,从而解释效应与浓度之间的滞后 (Sheiner et al. 1979)。Holford 与 Sheiner 随后系统阐述了剂量—效应关系的临床应用 (Holford and Sheiner 1981)。这是药理学中最标准的工具之一。
推导
# Chapter 3. The effect compartment, and why the time course of an effect is
# not the time course of a concentration. A plasma concentration Cp drives a
# notional effect-site concentration Ce through a first-order link:
# dCe/dt = ke0 (Cp - Ce)
# and the effect follows the Hill relation of p01 applied to Ce. The
# consequence is
# hysteresis: plotting effect against plasma concentration traces a loop, not
# a curve, and the peak effect lags the peak concentration. The lag is set by
# ke0, which is estimated only by fitting effect data and cannot be obtained
# from concentrations alone. Dimensionless throughout; Cp here is the
# active-species curve of p02.
import math
DT, T_END = 0.002, 16.0
KEL = math.log(2.0)
KA_LOSS = 0.05
def plasma(D=1.0, Km=0.3, Vmax=1.0):
A, P, out = D, 0.0, []
for i in range(int(T_END / DT)):
conv = Vmax * A / (Km + A)
A = max(0.0, A + DT * (-conv - KA_LOSS * A))
P = max(0.0, P + DT * (conv - KEL * P))
out.append(P)
return out
def effect_site(cp, ke0):
Ce, out = 0.0, []
for v in cp:
Ce += DT * ke0 * (v - Ce)
out.append(Ce)
return out
def hill(c, ec50=0.35, h=1.3):
if c <= 0:
return 0.0
r = (c / ec50) ** h
return r / (1.0 + r)
def summarise(series):
k = max(range(len(series)), key=lambda i: series[i])
return series[k], (k + 1) * DT
if __name__ == "__main__":
cp = plasma()
cmax, tmax = summarise(cp)
print(
"Active-species (plasma) curve: Cmax = {:.4f} at t = {:.2f}".format(
cmax, tmax
)
)
print("Time unit = elimination half-life of the active species.\n")
print(" ke0 Ce,max t(Ce,max) lag Emax t(Emax) lag")
for ke0 in (10.0, 3.0, 1.0, 0.3, 0.1):
ce = effect_site(cp, ke0)
e = [hill(v) for v in ce]
cem, cet = summarise(ce)
em, et = summarise(e)
print(" {:5.1f} {:8.4f} {:10.2f} {:7.2f} {:8.4f} {:9.2f} {:6.2f}"
.format(ke0, cem, cet, cet - tmax, em, et, et - tmax))
print(
"\nAs ke0 falls, the effect-site peak is lower and later: at ke0 = 10"
" the lag"
)
print(
"is almost nothing, at ke0 = 0.1 it is more than a half-life. The"
" plasma"
)
print(
"curve is IDENTICAL in every row. So the shape of the effect over"
" time is"
)
print(
"not a property of the pharmacokinetics; it is a property of ke0,"
" which has"
)
print(
"to be fitted to effect measurements and cannot be read off"
" concentrations."
)
print("\nHysteresis: effect against plasma concentration, ke0 = 0.3,")
print("on the way up and on the way down.")
ce = effect_site(cp, 0.3)
e = [hill(v) for v in ce]
peak_i = max(range(len(cp)), key=lambda i: cp[i])
print(" Cp effect (rising) effect (falling)")
for target in (0.05, 0.10, 0.20, 0.30):
up = next((i for i in range(peak_i) if cp[i] >= target), None)
dn = next(
(i for i in range(len(cp) - 1, peak_i, -1) if cp[i] >= target),
None,
)
if up is not None and dn is not None:
print(" {:6.3f} {:17.4f} {:18.4f}".format(target, e[up], e[dn]))
print(
"\n At the same plasma concentration the effect is larger on the"
" way down"
)
print(
" than on the way up. Any attempt to read a concentration-effect"
" curve"
)
print(
" from paired samples without modelling the lag will therefore"
" produce a"
)
print(" loop and, if the loop is averaged away, a wrong EC50.")
print(
"\nWhat this means for the word 'sub-perceptual'. Suppose the"
" threshold of"
)
print("noticing sits at an effect of 0.25. Time spent above it:")
for ke0 in (10.0, 1.0, 0.3, 0.1):
ce = effect_site(cp, ke0)
e = [hill(v) for v in ce]
above = sum(1 for v in e if v >= 0.25) * DT
print(
" ke0 = {:5.1f}: {:.2f} half-lives above the threshold, peak"
" effect {:.3f}".format(ke0, above, max(e))
)
print(
"\n The result is not a graded one. For ke0 of 0.3 and above the"
" same"
)
print(
" exposure sits above the threshold for close to four"
" half-lives; at"
)
print(
" ke0 = 0.1 it never reaches the threshold at all, because the"
" effect-site"
)
print(
" peak (0.162) falls short of 0.25. One unmeasured parameter"
" therefore"
)
print(
" decides between 'clearly noticeable for most of a day' and 'never"
)
print(
" noticeable', at an identical plasma curve. 'Sub-perceptual' is"
" not a"
)
print(
" statement about an amount; p05 shows it is not even a statement"
" about"
)
print(" one person.")
# Active-species (plasma) curve: Cmax = 0.4972 at t = 1.34
# Time unit = elimination half-life of the active species.
#
# ke0 Ce,max t(Ce,max) lag Emax t(Emax) lag
# 10.0 0.4947 1.44 0.10 0.6106 1.44 0.10
# 3.0 0.4704 1.68 0.34 0.5949 1.68 0.34
# 1.0 0.3740 2.18 0.84 0.5215 2.18 0.84
# 0.3 0.2120 3.06 1.72 0.3426 3.06 1.72
# 0.1 0.0989 4.17 2.83 0.1621 4.17 2.83
#
# As ke0 falls, the effect-site peak is lower and later: at ke0 = 10 the lag
# is almost nothing, at ke0 = 0.1 it is more than a half-life. The plasma
# curve is IDENTICAL in every row. So the shape of the effect over time is
# not a property of the pharmacokinetics; it is a property of ke0, which has
# to be fitted to effect measurements and cannot be read off concentrations.
#
# Hysteresis: effect against plasma concentration, ke0 = 0.3,
# on the way up and on the way down.
# Cp effect (rising) effect (falling)
# 0.050 0.0002 0.2684
# 0.100 0.0013 0.3166
# 0.200 0.0087 0.3424
# 0.300 0.0277 0.3329
#
# At the same plasma concentration the effect is larger on the way down
# than on the way up. Any attempt to read a concentration-effect curve
# from paired samples without modelling the lag will therefore produce a
# loop and, if the loop is averaged away, a wrong EC50.
#
# What this means for the word 'sub-perceptual'. Suppose the threshold of
# noticing sits at an effect of 0.25. Time spent above it:
# ke0 = 10.0: 3.35 half-lives above the threshold, peak effect 0.611
# ke0 = 1.0: 3.88 half-lives above the threshold, peak effect 0.522
# ke0 = 0.3: 3.83 half-lives above the threshold, peak effect 0.343
# ke0 = 0.1: 0.00 half-lives above the threshold, peak effect 0.162
#
# The result is not a graded one. For ke0 of 0.3 and above the same
# exposure sits above the threshold for close to four half-lives; at
# ke0 = 0.1 it never reaches the threshold at all, because the effect-site
# peak (0.162) falls short of 0.25. One unmeasured parameter therefore
# decides between 'clearly noticeable for most of a day' and 'never
# noticeable', at an identical plasma curve. 'Sub-perceptual' is not a
# statement about an amount; p05 shows it is not even a statement about
# one person.第一张表的全部内容是:血浆曲线在每一行里完全相同,而效应的峰高与峰时随 大幅变化。 时滞后几乎为零; 时滞后 2.83 个半衰期,而峰值只有前者的约四分之一。
所以
效应的时程不是药代动力学的性质,而是 的性质。
而 只能由效应数据拟合,不能从浓度数据读出。这对本书的主题有一个尖锐的后果,见程序的最后一段:假定知觉阈位于效应 0.25,则同一条血浆曲线在 时有将近四个半衰期在阈上,而在 时从未越过阈(峰值效应 0.162)。一个未测量的参数就决定了「明显可察觉大半天」与「完全察觉不到」之间的差别。
滞后回线
程序的第二段给出同一模型的另一面:在同一血浆浓度下,下降段的效应远大于上升段(在 处为 0.3166 对 0.0013)。效应对浓度的图不是一条曲线而是一个回线。
这有一个实际的后果:任何试图从配对的血浆—效应样本直接读出浓度—效应曲线的做法,如果不建模滞后,就会得到一个回线;而如果把回线平均掉,得到的 是错的。第 34 本讨论过心理物理函数的拟合,这里的问题是同一类的,只是多了一个时间维度。
「亚知觉」是一个人群分布
第一章说明受体不提供阈值,阈值只能由知觉给出。本节说明知觉给出的也不是一个数。
在一个剂量比例与一个人之间,至少隔着两层个体差异:同样的量在不同人身上产生不同的暴露(药代动力学的个体差异),而同样的暴露在不同人身上跨过不同的察觉阈(药效学的个体差异)。
推导
# Chapter 3. "Sub-perceptual" is a distribution across people, not a dose. Two
# sources of variability sit between a dose fraction and whether anyone
# notices anything:
# (1) pharmacokinetic: exposure per unit dose varies between people
# (log-normal, CV_pk) (2) pharmacodynamic: the concentration at which a
# person notices varies (log-normal)
# The program computes, for a range of dose fractions, the proportion of a
# population who would notice an effect, and then asks the design question
# that matters for Chapter 8: at what dose fraction is the proportion noticing
# low enough for a trial to stay blind? Dimensionless; synthetic population;
# fixed seed; standard library only.
import math
import random
from statistics import NormalDist
N = NormalDist()
CV_PK = 0.45 # between-person variability in exposure per unit dose
CV_THR = 0.60 # between-person variability in the noticing threshold
POP = 40000
def lognormal_sigma(cv):
return math.sqrt(math.log(1.0 + cv * cv))
def proportion_noticing(dose_fraction, rng):
s_pk = lognormal_sigma(CV_PK)
s_th = lognormal_sigma(CV_THR)
hit = 0
for _ in range(POP):
exposure = dose_fraction * math.exp(rng.gauss(-s_pk ** 2 / 2, s_pk))
thresh = math.exp(rng.gauss(-s_th ** 2 / 2, s_th))
hit += exposure >= thresh
return hit / POP
def analytic(dose_fraction):
"""Both log-normal with unit median: P(exposure >= threshold) is a normal
cdf."""
s = math.sqrt(lognormal_sigma(CV_PK) ** 2 + lognormal_sigma(CV_THR) ** 2)
return N.cdf(math.log(dose_fraction) / s)
if __name__ == "__main__":
rng = random.Random(5204)
print("Between-person CV: exposure {:.0%}, noticing threshold {:.0%}."
.format(CV_PK, CV_THR))
print(
"Dose fraction 1.0 is defined as the population MEDIAN noticing"
" exposure.\n"
)
print(" dose fraction proportion noticing (simulated) (analytic)")
for d in (0.05, 0.10, 0.15, 0.20, 0.30, 0.50, 1.00):
print(" {:13.2f} {:32.4f} {:12.4f}"
.format(d, proportion_noticing(d, rng), analytic(d)))
print(
"\nThere is no dose fraction at which the proportion is zero. At one"
" tenth of"
)
print(
"the median noticing exposure, {:.1%} of people still notice"
" something; at".format(analytic(0.10))
)
print(
"one fifth, {:.1%} do. A blind that depends on nobody noticing"
" cannot be".format(analytic(0.20))
)
print(
"achieved by lowering the amount; it can only be made imperfect to a"
" known"
)
print(
"degree, which is why Chapter 8 MEASURES the blind rather than"
" assuming it."
)
print(
"\nThe design question: what proportion noticing can a trial"
" tolerate? If a"
)
print(
"fraction f of participants correctly identify their allocation and"
" their"
)
print(
"expectancy contributes an effect of size g (in SD units) while the"
" drug"
)
print(
"itself contributes nothing, the observed between-arm difference is"
" about"
)
print("f * g. For an observed 0.20 SD difference to be pure expectancy:")
print(" proportion noticing f required expectancy effect g (SD)")
for f in (0.05, 0.10, 0.20, 0.40, 0.60):
print(" {:21.2f} {:38.2f}".format(f, 0.20 / f))
print(
"\n At f = 0.40 an expectancy effect of 0.50 SD would suffice,"
" which is"
)
print(
" well within the range reported for open-label expectancy"
" manipulations."
)
print(
" At f = 0.05 it would need 4.0 SD, which is not plausible. So the"
" blind"
)
print(
" integrity measurement is not a formality: it decides whether an"
" observed"
)
print(
" difference of this size can be explained without any drug effect"
" at all."
)
print(
"\nInverting: the dose fraction at which a given proportion notices."
)
s = math.sqrt(lognormal_sigma(CV_PK) ** 2 + lognormal_sigma(CV_THR) ** 2)
print(" target proportion dose fraction")
for p in (0.02, 0.05, 0.10, 0.25, 0.50):
print(" {:17.2f} {:14.4f}".format(p, math.exp(N.inv_cdf(p) * s)))
d02 = math.exp(N.inv_cdf(0.02) * s)
print(
"\n Reaching 2 % noticing requires about {:.0%} of the median"
" noticing".format(d02)
)
print(
" exposure - a fourfold reduction, not the hundredfold that the"
" word"
)
print(
" 'micro' suggests. Note also how flat this is: between 2 % and"
" 25 %"
)
print(
" noticing there is only a factor of {:.1f} in exposure, so a"
" schedule".format(math.exp(N.inv_cdf(0.25) * s) / d02)
)
print(
" intended to stay below the threshold has very little margin."
" Whether"
)
print(
" anything else survives at that exposure is exactly the open"
" question,"
)
print(
" and Chapter 4 shows why the receptor argument does not settle it."
)
# Between-person CV: exposure 45%, noticing threshold 60%.
# Dose fraction 1.0 is defined as the population MEDIAN noticing exposure.
#
# dose fraction proportion noticing (simulated) (analytic)
# 0.05 0.0000 0.0000
# 0.10 0.0008 0.0005
# 0.15 0.0045 0.0034
# 0.20 0.0138 0.0109
# 0.30 0.0505 0.0430
# 0.50 0.1847 0.1615
# 1.00 0.5379 0.5000
#
# There is no dose fraction at which the proportion is zero. At one tenth of
# the median noticing exposure, 0.1% of people still notice something; at
# one fifth, 1.1% do. A blind that depends on nobody noticing cannot be
# achieved by lowering the amount; it can only be made imperfect to a known
# degree, which is why Chapter 8 MEASURES the blind rather than assuming it.
#
# The design question: what proportion noticing can a trial tolerate? If a
# fraction f of participants correctly identify their allocation and their
# expectancy contributes an effect of size g (in SD units) while the drug
# itself contributes nothing, the observed between-arm difference is about
# f * g. For an observed 0.20 SD difference to be pure expectancy:
# proportion noticing f required expectancy effect g (SD)
# 0.05 4.00
# 0.10 2.00
# 0.20 1.00
# 0.40 0.50
# 0.60 0.33
#
# At f = 0.40 an expectancy effect of 0.50 SD would suffice, which is
# well within the range reported for open-label expectancy manipulations.
# At f = 0.05 it would need 4.0 SD, which is not plausible. So the blind
# integrity measurement is not a formality: it decides whether an observed
# difference of this size can be explained without any drug effect at all.
#
# Inverting: the dose fraction at which a given proportion notices.
# target proportion dose fraction
# 0.02 0.2368
# 0.05 0.3155
# 0.10 0.4071
# 0.25 0.6231
# 0.50 1.0000
#
# Reaching 2 % noticing requires about 24% of the median noticing
# exposure - a fourfold reduction, not the hundredfold that the word
# 'micro' suggests. Note also how flat this is: between 2 % and 25 %
# noticing there is only a factor of 2.6 in exposure, so a schedule
# intended to stay below the threshold has very little margin. Whether
# anything else survives at that exposure is exactly the open question,
# and Chapter 4 shows why the receptor argument does not settle it.第一张表的结论是:没有任何剂量比例使察觉比例为零。这不是参数选择的问题,而是两个对数正态分布的尾部必然相交。
第二段把这件事翻译成设计语言。若比例 的参与者正确识别了自己的分组,而他们的预期贡献一个 个标准差的效应,则表观的组间差别约为 。要用纯预期解释一个 0.20 个标准差的观测差别: 时只需 ,这在开放标签的预期操控研究中是常见的量级; 时需要 ,不可信。所以设盲完整性的测量不是形式——它决定一个这个量级的观测差别能否在完全没有药理效应的情况下被解释掉。
第三段给出反向计算,并指出一件设计上要紧的事:从 2% 察觉到 25% 察觉之间只有 2.6 倍的暴露余地。一个想要留在阈下的安排几乎没有余量。
注
对第八章的后果。 设盲不可能靠压低暴露来保证。第八章因此(一)测量设盲完整性并为检出它供能,(二)把「按信念分层」的对照列为预先指定的关键次要分析,(三)在判定规则中把设盲破坏写成限定条件。缺乏参与者设盲会造成偏倚,这有系统综述的证据 (Hróbjartsson et al. 2014);本书的做法是测量它而不是假设它不存在。
假说 · 尚未验证
作者的假说(尚未确立): 在这一类化合物的双盲试验中,猜中分组的比例与活性条件的暴露比例单调相关,且在任何产生可检出客观效应的暴露下,猜中比例都显著高于随机。什么样的结果能证明它是错的: 在事先登记的剂量—反应双盲研究中,找到一个暴露水平,使预先指定的客观结局相对安慰剂的改善的置信区间排除零,而 Bang 设盲指数 (Bang, Ni, and Davis 2004) 的置信区间包含零。
本章小结
效应室模型是药理学的标准工具,而它给出一个对本书主题关键的结论:效应的时程是 的性质而不是药代动力学的性质。在完全相同的血浆曲线下, 从 0.3 降到 0.1 就使「是否越过知觉阈」从「将近四个半衰期在阈上」变为「从未越过」,而 只能由效应数据拟合。同一模型给出滞后回线:在同一血浆浓度下下降段的效应远大于上升段,所以不建模滞后的浓度—效应拟合会给出错误的 。最后,「亚知觉」是一个人群分布而不是一个剂量:在暴露与阈值各有个体差异的人群中,没有任何剂量比例使察觉比例为零,而从 2% 察觉到 25% 察觉之间只有 2.6 倍的余地。要用纯预期解释一个 0.20 个标准差的观测差别,在 40% 猜中率下只需 0.50 个标准差的预期效应,所以设盲完整性必须被测量。
参考文献
Fan, Y. (2026). Biomathematical Models of Sub-Perceptual Tryptamine Pharmacokinetics, 5-HT2A Receptor Desensitisation and Neuroplastic Window Dynamics: From Michaelis–Menten Kinetics of Prodrug Deacetylation and Whole-Brain Connectivity Entropy to the Microdose Tolerance Ratchet, with a Textual Examination of the Dose Boundary in the Classical Materia Medica [亚知觉色胺药代动力学、5-HT2A 受体脱敏与神经可塑性窗口的生物数学模型] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24019347 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).