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

第七章 原典并读:「先起如黍粟」与剂量的边界

注

对勘一律是类比,且本章不含任何可照着做的内容。 本章把古代文本当作算法与命题来分析,不是用古籍支持任何现代做法,也不是把古代的用药经验转译成建议。本章讨论的「剂量探索规则」是在合成人群上的算法研究,不针对任何物质,任何人都不应据此选择任何东西的用量。

三品分类与「轻身延年」

原典

《神农本草经》序录把药分为三品。上品一百二十种为君,「主养命以应天,无毒,多服久服不伤人,欲轻身益气不老延年者本上经」;中品为臣,「主养性以应人」;下品为佐使,「主治病以应地,多毒,不可久服」(Shen Nong 1998)。

白话:上品的药养生命,没有毒,多吃久吃不伤人,想要身体轻健、气力充足、不老延年的人以上品为依据;中品养性;下品治病,毒性多,不可以长期服用。

「轻身延年」正是本书副题所引的措辞。而这里有一个必须写明的问题。

证据

上品包含今天已知有毒的物质。 《神农本草经》上品的第一味是丹砂,即硫化汞。现代毒理学对中药中的汞作过专门研究:Liu 等人指出硫化汞与常见汞化合物(如氯化汞、甲基汞)在毒性上并不等同,口服吸收率低得多 (Liu et al. 2008)。但这一点不足以支持「多服久服不伤人」或「久服通神明」。

所以

上品的「无毒、久服」措辞不能当作安全性的历史证据。它记录的是一个分类体系的判断,而那个判断在至少一个重要的例子上与现代毒理学不符。

这一点对本书的主题是直接相关的:如果有人用「古人说某物轻身延年」来支持长期低量使用某种物质,那么这个论证形式本身在《本经》自己的上品里就有反例。

「先起如黍粟」:一个剂量探索算法

《本经》序录的另一段则完全不同——它是一条规则,而规则可以分析。

原典

《神农本草经》序录:「若用毒药疗病,先起如黍粟,病去即止,不去倍之,不去十之,取去为度。」(Shen Nong 1998)

白话:如果用有毒的药治病,先从像黍米粟米那样小的量开始,病好了就停;不好就加倍,还不好就加到十倍,以病除为标准。

这段话有算法的全部要素:起始剂量(最小可行量)、停止规则(病去即止)、递增规则(倍之,然后十之)、终止判据(取去为度)。而剂量探索算法有可以计算的运作特性。

证据

现代的剂量探索设计。 Storer 系统论述了第一期临床试验的设计与分析,其中包括后来通称「3+3」的一类规则 (Storer 1989)。O’Quigley、Pepe 与 Fisher 提出连续重估法,用一个参数模型把已观察到的毒性反馈到下一个剂量的选择上 (O’Quigley, Pepe, and Fisher 1990)。两者都以毒性为升降的依据,并在人群层面学习。

推导

# Chapter 7. The Shennong Bencao Jing's escalation rule, treated as an
# algorithm. The preface to the Bencao Jing instructs, for toxic drugs: begin
# with an amount "like a millet grain"; if the illness goes, stop; if it does
# not, double it; if it still does not, go tenfold; take the removal of the
# illness as the measure. That is a dose-finding rule, and dose-finding rules
# have operating characteristics. The program compares it with a modern 3+3
# design and a one-parameter continual reassessment method on the same
# synthetic patient population, and reports what each finds and what each
# costs. Everything is in dose-fraction units relative to an unspecified
# reference. Synthetic; fixed seed; standard library only. This is a study of
# the RULE, not advice to use it.
import math
import random

# 0.05 .. 12.8, dimensionless
DOSE_LEVELS = [0.05 * (2 ** k) for k in range(9)]

def p_toxic(d, d50=3.0, slope=1.6):
    return 1.0 / (1.0 + (d50 / d) ** slope)

def p_effective(d, d50=0.8, slope=1.8):
    return 1.0 / (1.0 + (d50 / d) ** slope)

def classical(rng, max_steps=12):
    """Begin minimal; stop if effective; else double; if still not, go
    tenfold."""
    d, used, doubled = DOSE_LEVELS[0], 0, 0
    tox = False
    for _ in range(max_steps):
        used += 1
        if rng.random() < p_toxic(d):
            tox = True
            break
        if rng.random() < p_effective(d):
            return d, used, tox
        if doubled < 2:
            d *= 2.0
            doubled += 1
        else:
            d *= 10.0
            doubled = 0
        if d > DOSE_LEVELS[-1]:
            break
    return d, used, tox

def three_plus_three(rng, target=0.33):
    """Classic 3+3: escalate while 0/3, expand on 1/3, stop on >=2."""
    exposures = 0
    for i, d in enumerate(DOSE_LEVELS):
        t = sum(1 for _ in range(3) if rng.random() < p_toxic(d))
        exposures += 3
        if t == 0:
            continue
        if t == 1:
            t2 = sum(1 for _ in range(3) if rng.random() < p_toxic(d))
            exposures += 3
            if t + t2 <= 1:
                continue
        return (
            (DOSE_LEVELS[i - 1] if i > 0 else DOSE_LEVELS[0]),
            exposures,
            True,
        )
    return DOSE_LEVELS[-1], exposures, False

def crm(rng, n=24, target=0.25):
    """One-parameter power model CRM on a skeleton, grid posterior over
    beta."""
    skeleton = [p_toxic(d) for d in DOSE_LEVELS]
    betas = [0.2 * (10 ** (0.05 * k)) for k in range(41)]
    logpost = [0.0] * len(betas)
    idx, exposures, tox_total = 0, 0, 0
    for _ in range(n):
        d = DOSE_LEVELS[idx]
        y = 1 if rng.random() < p_toxic(d) else 0
        exposures += 1
        tox_total += y
        for bi, b in enumerate(betas):
            p = min(max(skeleton[idx] ** b, 1e-9), 1 - 1e-9)
            logpost[bi] += math.log(p) if y else math.log(1 - p)
        m = max(logpost)
        w = [math.exp(v - m) for v in logpost]
        s = math.fsum(w)
        w = [v / s for v in w]
        best, bestgap = 0, 9e9
        for j in range(len(DOSE_LEVELS)):
            est = math.fsum(
                wi * (skeleton[j] ** b) for wi, b in zip(w, betas)
            )
            if abs(est - target) < bestgap:
                best, bestgap = j, abs(est - target)
        idx = min(best, idx + 1)                        # no skipping upward
    return DOSE_LEVELS[idx], exposures, tox_total > 0

if __name__ == "__main__":
    print(
        "Dose levels (dimensionless fractions of an unspecified reference):"
    )
    print("   " + " ".join("{:.2f}".format(d) for d in DOSE_LEVELS))
    print(
        "\nSynthetic population: P(toxic) has its midpoint at 3.00 and"
        " P(effective)"
    )
    print(
        "at 0.80, so there is a real therapeutic window. Target toxicity"
        " 25 %."
    )
    print("   dose   P(effective)   P(toxic)")
    for d in DOSE_LEVELS[:7]:
        print(
            "   {:5.2f} {:14.3f} {:10.3f}".format(
                d, p_effective(d), p_toxic(d)
            )
        )

    REPS = 3000
    print("\n{} simulated dose-findings per rule.\n".format(REPS))
    print("   rule                    median dose found  mean exposures"
          "  any toxicity")
    for name, fn in (("Bencao Jing escalation", classical),
                     ("3+3", three_plus_three),
                     ("CRM, 24 patients", crm)):
        rng = random.Random(777)
        doses, exps, toxes = [], [], 0
        for _ in range(REPS):
            d, e, t = fn(rng)
            doses.append(d)
            exps.append(e)
            toxes += t
        doses.sort()
        print(
            "   {:24s} {:16.2f} {:15.1f} {:13.1%}".format(
                name, doses[REPS // 2], math.fsum(exps) / REPS, toxes / REPS
            )
        )

    print(
        "\n   The three columns are not like for like, and that matters. The"
    )
    print(
        "   classical rule is a rule for ONE person, so its exposure count"
        " is one"
    )
    print(
        "   person's escalation and its toxicity column is that person's"
        " risk. The"
    )
    print(
        "   3+3 and the CRM are trial-level designs run across many people,"
        " so their"
    )
    print(
        "   exposure counts are totals and 'any toxicity' is near-certain"
        " simply"
    )
    print(
        "   because twenty-odd people were exposed. The comparison is"
        " informative"
    )
    print("   about the SHAPE of each rule, not about which is safer.")

    print(
        "\nThe classical rule reaches a dose near the effective midpoint with"
        " very"
    )
    print(
        "few exposures, because it escalates against the RESPONSE rather than"
    )
    print(
        "against toxicity, and it stops the moment the illness goes. That"
        " is its"
    )
    print(
        "strength: it is individualised and it is frugal. Its weakness is"
        " also"
    )
    print(
        "visible: it has no mechanism for learning across people, so each"
        " person"
    )
    print(
        "repeats the whole search, and the tenfold step after two"
        " doublings is"
    )
    print("large enough to cross a steep toxicity curve in one move.")

    print("\nWhat the tenfold step costs, measured as P(toxic) at each step:")
    d = DOSE_LEVELS[0]
    seq = [d]
    for k in range(4):
        seq.append(seq[-1] * 2)
    seq.append(seq[-1] * 10)
    print("   step    dose    P(toxic)   increment in P(toxic)")
    prev = 0.0
    for i, dd in enumerate(seq):
        pt = p_toxic(dd)
        print(
            "   {:5d} {:7.2f} {:10.4f} {:22.4f}".format(i, dd, pt, pt - prev)
        )
        prev = pt
    print(
        "\n   The doublings move P(toxic) by small amounts; the single"
        " tenfold step"
    )
    print(
        "   moves it by much more than all the doublings combined. Read as an"
    )
    print(
        "   algorithm, the rule is therefore well designed in its early"
        " phase and"
    )
    print(
        "   badly designed in its last step - which is exactly the kind of"
        " judgement"
    )
    print("   a formal comparison can offer and a reverent reading cannot.")
    print(
        "\n   This is an analysis of a RULE on synthetic data. It is not a"
        " claim"
    )
    print(
        "   about any substance, and nothing here should be used to choose an"
        " amount"
    )
    print("   of anything.")
# Dose levels (dimensionless fractions of an unspecified reference):
#    0.05 0.10 0.20 0.40 0.80 1.60 3.20 6.40 12.80
#
# Synthetic population: P(toxic) has its midpoint at 3.00 and P(effective)
# at 0.80, so there is a real therapeutic window. Target toxicity 25 %.
#    dose   P(effective)   P(toxic)
#     0.05          0.007      0.001
#     0.10          0.023      0.004
#     0.20          0.076      0.013
#     0.40          0.223      0.038
#     0.80          0.500      0.108
#     1.60          0.777      0.268
#     3.20          0.924      0.526
#
# 3000 simulated dose-findings per rule.
#
#    rule                    median dose found  mean exposures  any toxicity
#    Bencao Jing escalation               2.00             3.9         38.5%
#    3+3                                  0.80            22.2        100.0%
#    CRM, 24 patients                     1.60            24.0        100.0%
#
#    The three columns are not like for like, and that matters. The
#    classical rule is a rule for ONE person, so its exposure count is one
#    person's escalation and its toxicity column is that person's risk. The
#    3+3 and the CRM are trial-level designs run across many people, so their
#    exposure counts are totals and 'any toxicity' is near-certain simply
#    because twenty-odd people were exposed. The comparison is informative
#    about the SHAPE of each rule, not about which is safer.
#
# The classical rule reaches a dose near the effective midpoint with very
# few exposures, because it escalates against the RESPONSE rather than
# against toxicity, and it stops the moment the illness goes. That is its
# strength: it is individualised and it is frugal. Its weakness is also
# visible: it has no mechanism for learning across people, so each person
# repeats the whole search, and the tenfold step after two doublings is
# large enough to cross a steep toxicity curve in one move.
#
# What the tenfold step costs, measured as P(toxic) at each step:
#    step    dose    P(toxic)   increment in P(toxic)
#        0    0.05     0.0014                 0.0014
#        1    0.10     0.0043                 0.0029
#        2    0.20     0.0130                 0.0086
#        3    0.40     0.0383                 0.0253
#        4    0.80     0.1077                 0.0694
#        5    8.00     0.8277                 0.7200
#
#    The doublings move P(toxic) by small amounts; the single tenfold step
#    moves it by much more than all the doublings combined. Read as an
#    algorithm, the rule is therefore well designed in its early phase and
#    badly designed in its last step - which is exactly the kind of judgement
#    a formal comparison can offer and a reverent reading cannot.
#
#    This is an analysis of a RULE on synthetic data. It is not a claim
#    about any substance, and nothing here should be used to choose an amount
#    of anything.

注

三列不是可比的量。 《本经》的规则是一个人的规则,所以它的暴露次数是一个人的递增次数、毒性列是那个人的风险;3+3 与连续重估法是试验层面的设计,跨许多人实施,所以暴露次数是总数,而「出现毒性」接近必然,只因为有二十余人被暴露。这个比较提供的是三条规则形状上的信息,不是哪一条更安全。

结果有两面。

《本经》规则的长处。 它以反应(病去)而不是毒性为升降依据,并且在反应出现时立刻停止。这使它个体化而且节省——在这个合成人群上,它用约 3.9 次暴露就到达接近有效中点的剂量,而 3+3 与连续重估法分别用了 22.2 与 24.0 次。「取去为度」是一条明确的停止规则,这在剂量探索设计里是必需的,而许多现代的自我实验恰恰没有。

它的短处。 它没有跨人学习的机制,所以每个人都要把整个搜索重做一遍;而最后的「十之」一步,按程序的第二张表,使中毒概率的增量(0.7200)超过前面所有倍增之和(0.0014 + 0.0029 + 0.0086 + 0.0253 + 0.0694 = 0.1076)。

读作算法,《本经》的规则在倍增阶段设计良好,在最后一步设计不良。

这正是形式化比较能给出而尊崇式的阅读不能给出的判断。值得补一句:《本经》说的是「不去十之」,也就是在倍增无效之后才用十倍,所以这一步的处境是「已经两次无效」——在那个处境下接受更大的风险是有道理的。但风险的量级是上表的样子,而那是原文无从知道的。

帕拉塞尔苏斯与双相曲线

原典

帕拉塞尔苏斯(1493–1541)的剂量命题通常被概括为「剂量决定毒性」。他的原话大意是:一切物质都是毒,没有什么不是毒;只有剂量使一物不成为毒。Borzelleca 对帕拉塞尔苏斯在现代毒理学中的位置有专门的评述 (Borzelleca 2000)。

《本经》的「取去为度」与帕拉塞尔苏斯的命题在结构上同侧:两者都把判断的依据从物移到量。而两者都没有说量应该是多少——那需要一条曲线。

把「有益」与「有害」写成两条饱和曲线之差,就得到双相(倒 U 形)的剂量—反应。这正是毒物兴奋效应的形式,Calabrese 与 Baldwin 主张它在毒理学中比单调模型更普遍 (Calabrese and Baldwin 2003);第 30 本整本书讨论这一主题,这里只取一个计算。

推导

# Chapter 7. "Light the body and extend the years" as a biphasic
# dose-response, and why the location of the optimum is the hardest thing to
# measure. A benefit term saturates while a cost term grows with a steeper
# slope:
#   net(d) = B d^a / (d^a + kB^a) - C d^b / (d^b + kC^b)
# With b > a the net response is an inverted U: zero at the bottom, a maximum
# somewhere, and negative beyond. Volume 30 treated hormesis in general; what
# this program adds is an uncertainty calculation. Given plausible uncertainty
# in the four parameters, how precisely is the optimum located? The answer
# bounds what any recommendation could be. Dimensionless dose fractions;
# synthetic; fixed seed; standard library only.
import math
import random

def net(d, B=1.0, kB=0.8, a=1.6, C=1.0, kC=4.0, b=3.0):
    if d <= 0:
        return 0.0
    ben = B * d ** a / (d ** a + kB ** a)
    cost = C * d ** b / (d ** b + kC ** b)
    return ben - cost

def optimum(**kw):
    grid = [0.02 * (10 ** (0.01 * k)) for k in range(301)]      # 0.02 .. 20
    best = max(grid, key=lambda d: net(d, **kw))
    return best, net(best, **kw)

if __name__ == "__main__":
    print("Benefit saturates (exponent 1.6, midpoint 0.8); cost is steeper")
    print(
        "(exponent 3.0, midpoint 4.0). All in dimensionless dose fractions.\n"
    )
    print("   dose   benefit   cost    net")
    for d in (0.1, 0.3, 0.8, 1.5, 2.5, 4.0, 6.0, 10.0):
        ben = 1.0 * d ** 1.6 / (d ** 1.6 + 0.8 ** 1.6)
        cost = 1.0 * d ** 3.0 / (d ** 3.0 + 4.0 ** 3.0)
        print(
            "   {:6.2f} {:9.3f} {:7.3f} {:7.3f}".format(
                d, ben, cost, ben - cost
            )
        )

    d_opt, v_opt = optimum()
    print("\nOptimum at d = {:.3f}, net = {:.3f}".format(d_opt, v_opt))

    print(
        "\nHow flat is the peak? The range of doses within 5 % and 10 % of"
        " the best:"
    )
    grid = [0.02 * (10 ** (0.005 * k)) for k in range(601)]
    for frac in (0.05, 0.10, 0.25):
        ok = [d for d in grid if net(d) >= v_opt * (1 - frac)]
        print(
            "   within {:3.0%} of best: {:.3f} to {:.3f} (a factor of {:.2f})"
            .format(frac, min(ok), max(ok), max(ok) / min(ok))
        )

    print(
        "\nNow the uncertainty. Each of the four parameters is given a modest"
    )
    print("log-normal uncertainty (30 % CV) and the optimum is recomputed:")
    rng = random.Random(5213)
    opts = []
    for _ in range(2000):
        kw = dict(kB=0.8 * math.exp(rng.gauss(0, 0.29)),
                  a=1.6 * math.exp(rng.gauss(0, 0.29)),
                  kC=4.0 * math.exp(rng.gauss(0, 0.29)),
                  b=3.0 * math.exp(rng.gauss(0, 0.29)))
        opts.append(optimum(**kw)[0])
    opts.sort()
    print("   median optimum          {:.3f}".format(opts[len(opts) // 2]))
    print("   5th to 95th percentile  {:.3f} to {:.3f}  (a factor of {:.1f})"
          .format(opts[len(opts) // 20], opts[-len(opts) // 20],
                  opts[-len(opts) // 20] / opts[len(opts) // 20]))

    print(
        "\n   A 30 % uncertainty in the parameters becomes a factor of"
        " roughly {:.0f}".format(
            opts[-len(opts) // 20] / opts[len(opts) // 20]
        )
    )
    print(
        "   in the location of the optimum. That is the quantitative reason a"
        " dose"
    )
    print(
        "   recommendation cannot be derived from a dose-response shape that"
        " has not"
    )
    print(
        "   been measured: the optimum is the least identifiable feature"
        " of the"
    )
    print(
        "   curve, because it depends on the DIFFERENCE of two uncertain"
        " terms."
    )

    print(
        "\nThe asymmetry of being wrong. Net response at multiples of the"
        " optimum:"
    )
    print("   multiple of optimum   dose    net    fraction of best")
    for mult in (0.25, 0.5, 1.0, 2.0, 3.0, 4.0):
        d = d_opt * mult
        print("   {:19.2f} {:7.3f} {:7.3f} {:17.1%}"
              .format(mult, d, net(d), net(d) / v_opt))
    print(
        "\n   A quarter of the optimum keeps {:.0%} of the benefit; four"
        " times it".format(net(d_opt * 0.25) / v_opt)
    )
    print(
        "   gives {:.0%}. The curve is far more forgiving below the optimum"
        " than".format(net(d_opt * 4.0) / v_opt)
    )
    print(
        "   above it, which is the formal content of the classical"
        " instruction to"
    )
    print(
        "   begin with the smallest possible amount and increase only if"
        " needed."
    )
    print(
        "   Note that this conclusion depends on b > a - on the cost curve"
        " being"
    )
    print(
        "   steeper than the benefit curve - which is an assumption of the"
        " model,"
    )
    print("   not a measurement.")
# Benefit saturates (exponent 1.6, midpoint 0.8); cost is steeper
# (exponent 3.0, midpoint 4.0). All in dimensionless dose fractions.
#
#    dose   benefit   cost    net
#      0.10     0.035   0.000   0.035
#      0.30     0.172   0.000   0.172
#      0.80     0.500   0.008   0.492
#      1.50     0.732   0.050   0.682
#      2.50     0.861   0.196   0.665
#      4.00     0.929   0.500   0.429
#      6.00     0.962   0.771   0.190
#     10.00     0.983   0.940   0.043
#
# Optimum at d = 1.867, net = 0.703
#
# How flat is the peak? The range of doses within 5 % and 10 % of the best:
#    within  5% of best: 1.400 to 2.461 (a factor of 1.76)
#    within 10% of best: 1.219 to 2.729 (a factor of 2.24)
#    within 25% of best: 0.883 to 3.396 (a factor of 3.85)
#
# Now the uncertainty. Each of the four parameters is given a modest
# log-normal uncertainty (30 % CV) and the optimum is recomputed:
#    median optimum          1.867
#    5th to 95th percentile  1.233 to 2.825  (a factor of 2.3)
#
#    A 30 % uncertainty in the parameters becomes a factor of roughly 2
#    in the location of the optimum. That is the quantitative reason a dose
#    recommendation cannot be derived from a dose-response shape that has not
#    been measured: the optimum is the least identifiable feature of the
#    curve, because it depends on the DIFFERENCE of two uncertain terms.
#
# The asymmetry of being wrong. Net response at multiples of the optimum:
#    multiple of optimum   dose    net    fraction of best
#                   0.25   0.467   0.295             42.0%
#                   0.50   0.933   0.549             78.1%
#                   1.00   1.867   0.703            100.0%
#                   2.00   3.733   0.473             67.3%
#                   3.00   5.600   0.225             32.0%
#                   4.00   7.466   0.106             15.1%
#
#    A quarter of the optimum keeps 42% of the benefit; four times it
#    gives 15%. The curve is far more forgiving below the optimum than
#    above it, which is the formal content of the classical instruction to
#    begin with the smallest possible amount and increase only if needed.
#    Note that this conclusion depends on b > a - on the cost curve being
#    steeper than the benefit curve - which is an assumption of the model,
#    not a measurement.

三个结果。

第一,峰是平的:在最优点 5% 以内的剂量跨度是一个因子,10% 以内更宽。所以「最优剂量」不是一个点。

第二,最优点的位置是曲线上最难定的特征。给四个参数各 30% 的对数正态不确定度,最优点的 5–95 百分位跨 2.3 倍。原因是最优点取决于两个不确定项之差,而差的相对误差大于各项的相对误差。所以

剂量建议不能从一条未被测量的剂量—反应形状推出来。

第三,错在低侧比错在高侧便宜:最优点的四分之一处保留 42% 的净获益,而四倍处只有 15%。这正是《本经》「先起如黍粟」与「不去倍之」的形式内容——从小量开始、必要时才加——而程序也写明这个结论依赖于成本曲线比获益曲线更陡(b>ab > a),那是模型的假设而不是测量。

左:双相剂量—反应:获益(虚线)与成本(点线)之差给出净反应(实线),最优点标出。右:四个参数各有 30% 对数正态不确定度时,最优点位置的分布;5–95 百分位跨 2.3 倍。

结构对照与界限

文本 它给出的东西 在本书框架中的位置
《本经》三品「轻身延年」 一个分类判断 不作为安全性证据;上品含丹砂是反例
《本经》「先起如黍粟……取去为度」 一个剂量探索算法 倍增阶段良好,「十之」一步使风险增量超过所有倍增之和
帕拉塞尔苏斯的剂量命题 把判断从物移到量 与「取去为度」同侧;都不给出量
双相剂量—反应 量的形式 最优点是最难定的特征(30% 参数不确定度 → 2.3 倍位置不确定度)

三点界限。第一,这些文本都没有讨论本书前六章的任何具体药理机制,所以本章不作那种对勘。第二,《本经》的「毒药」指的是有毒的药,用于治病,不是用于日常增益;把「先起如黍粟」搬到非治疗性的长期使用上,超出了原文的处境。第三,本章的算法分析是在合成人群上做的,它评价的是规则,不是任何物质。

本章小结

《神农本草经》上品的「无毒、多服久服不伤人、轻身益气不老延年」不能当作安全性的历史证据:上品第一味丹砂是硫化汞,而现代毒理学虽指出硫化汞与常见汞化合物毒性不等同,也不支持久服。序录的「若用毒药疗病,先起如黍粟,病去即止,不去倍之,不去十之,取去为度」则是一条完整的剂量探索算法;在合成人群上与 3+3 及连续重估法比较,它以反应而非毒性升降、有明确的停止规则、用约 3.9 次暴露就到达接近有效中点的剂量,但没有跨人学习的机制,而最后的「十之」一步使中毒概率的增量超过前面所有倍增之和。帕拉塞尔苏斯的命题与「取去为度」同侧:把判断从物移到量,而都不给出量。给出量需要一条曲线,而双相曲线的最优点是它最难定的特征——30% 的参数不确定度变成 2.3 倍的位置不确定度——同时曲线在低侧比高侧宽容(四分之一处保留 42% 的净获益,四倍处只有 15%),这是「先起如黍粟」的形式内容,并依赖成本曲线更陡这一假设。

参考文献

Borzelleca, Joseph F. 2000. “Paracelsus: Herald of Modern Toxicology.” Toxicological Sciences 53 (1): 2–4. https://doi.org/10.1093/toxsci/53.1.2.
Calabrese, Edward J., and Linda A. Baldwin. 2003. “Toxicology Rethinks Its Central Belief.” Nature 421: 691–92. https://doi.org/10.1038/421691a.
Liu, Jie, Jing-Zhen Shi, Li-Mei Yu, Robert A. Goyer, and Michael P. Waalkes. 2008. “Mercury in Traditional Medicines: Is Cinnabar Toxicologically Similar to Common Mercurials?” Experimental Biology and Medicine 233 (7): 810–17. https://doi.org/10.3181/0712-MR-336.
O’Quigley, John, Margaret Pepe, and Lloyd Fisher. 1990. “Continual Reassessment Method: A Practical Design for Phase 1 Clinical Trials in Cancer.” Biometrics 46 (1): 33–48. https://doi.org/10.2307/2531628.
Shen Nong. 1998. The Divine Farmer’s Materia Medica: A Translation of the Shen Nong Ben Cao Jing. The Three-Grade Classification and the Escalation Instruction Appear in the Preface. Translated by Shou-zhong Yang. Boulder, Colorado: Blue Poppy Press.
Storer, Barry E. 1989. “Design and Analysis of Phase I Clinical Trials.” Biometrics 45 (3): 925–37. https://doi.org/10.2307/2531693.

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).