5D Library › Neurobiology of Ritual, Spatial Boundary Transitions and Computational Biomedicine of Self-Transcendence · 中文

Chapter 6 Incense, Sound and Cost: Multisensory Integration, Breathing Rhythms and Costly Signals

Three claims about multisensory ritual

Many rituals engage several senses at once: incense, chanting, bells, candlelight and fixed movements. The framework makes three claims about this: multisensory inputs enhance each other most when they are weak (“inverse effectiveness”); terpene odorants bypass the thalamus to reach the piriform cortex and amygdala, act together with acoustic harmonics, and influence the brain through respiration-locked theta–gamma phase–amplitude coupling; and the high cost of ritual sustains group cooperation through a signalling game. This chapter checks each. Book 34 discussed maximum-likelihood multisensory integration and causal inference, Book 44 collective synchrony, and Book 35 cooperation and reputation; they are only recalled here.

Inverse effectiveness: neuronal non-linearity or statistical artefact?

Evidence

Single-neuron evidence of multisensory integration (animal experiments). Meredith and Stein (1983) recorded neurons in the cat superior colliculus that receive visual, auditory and somatosensory input: when two stimuli were close in space and time, the response to the combination could far exceed the response to either alone (Meredith and Stein 1983). The review of Stein and Stanford (2008) summarised integration at the single-neuron level as a spatial rule, a temporal rule and inverse effectiveness: the weaker the single stimuli, the larger the relative enhancement from combining them (Stein and Stanford 2008). Multisensory enhancement is usually measured as ME=(CM−max⁡(V,A))/max⁡(V,A)\text{ME} = (CM – \max(V, A))/\max(V, A), with CMCM the response to the combined stimulus.

Inverse effectiveness has a simple source: a neuron’s output is a sigmoid function of its input. In part A of the next program, a model neuron adds visual and auditory drives and passes the sum through a sigmoid (spontaneous rate 2, maximum 50 spikes). Part B illustrates the statistical problem pointed out by Holmes (2009) (Holmes 2009): if neurons are split into “weak” and “strong” by their observed unisensory responses and enhancement is computed from the same data, inverse effectiveness appears even when every neuron is purely additive with exactly the same true enhancement.

Calculation

# Inverse effectiveness: from an output nonlinearity, and as a statistical artefact
# (Chapter 6). Synthetic. Multisensory enhancement (Meredith & Stein):
#   ME = (CM - max(V, A)) / max(V, A) x 100 %,  CM = response to the combined stimulus.
# A: a model neuron sums visual and auditory drives and passes them through a sigmoid
#    (rate = 2 + 48 / (1 + exp(-(u - 5)/1.2)) spikes); V and A drives are equal.
# B: purely additive neurons (true ME = 100 % for every neuron) with Poisson spike counts
#    (5 trials per condition). Sorting neurons by their OBSERVED unisensory response and
#    computing ME from the same data produces "inverse effectiveness" from noise alone.
import math, random

def rate(u):
    return 2 + 48 / (1 + math.exp(-(u - 5) / 1.2))

print("A  drive per sense   V = A (spikes)   CM (spikes)   ME (%)   CM / (V + A)")
for u in (1.5, 2.5, 3.5, 4.5, 5.5, 7.0):
    v = rate(u); cm = rate(2 * u)
    print(f"   {u:10.1f} {v:15.1f} {cm:13.1f} {100 * (cm - v) / v:9.0f}"
          f" {cm / (2 * v):12.2f}")

def poisson(rng, lam):
    lim, k, p = math.exp(-lam), 0, 1.0
    while True:
        p *= rng.random()
        if p <= lim:
            return k
        k += 1

rng = random.Random(1983)
cells = []
for _ in range(3000):
    m = 6.0                                                 # identical true means
    obs = [math.fsum(poisson(rng, lam) for _ in range(5)) / 5 for lam in (m, m, 2 * m)]
    cells.append(obs)
cells.sort(key=lambda o: max(o[0], o[1]))
print("B  additive neurons, all with V = A = 6 and CM = 12 spikes (true ME = 100 %)")
for lab, part in (("weakest third", cells[:1000]), ("middle third", cells[1000:2000]),
                  ("strongest third", cells[2000:])):
    me = [100 * (c[2] - max(c[0], c[1])) / max(c[0], c[1]) for c in part]
    print(f"   {lab:16s} by observed max(V, A): mean ME {math.fsum(me) / len(me):6.1f} %")
# A  drive per sense   V = A (spikes)   CM (spikes)   ME (%)   CM / (V + A)
#           1.5             4.5           9.6       116         1.08
#           2.5             7.3          26.0       255         1.78
#           3.5            12.7          42.4       234         1.67
#           4.5            21.1          48.3       129         1.15
#           5.5            30.9          49.7        61         0.80
#           7.0            42.4          50.0        18         0.59
# B  additive neurons, all with V = A = 6 and CM = 12 spikes (true ME = 100 %)
#    weakest third    by observed max(V, A): mean ME  115.1 %
#    middle third     by observed max(V, A): mean ME   82.7 %
#    strongest third  by observed max(V, A): mean ME   57.9 %

Part A: with a unisensory drive of 2.5, the single response is 7.3 spikes, the combined response 26.0, the enhancement 255 %, and the combined response is 1.78 times the sum of the two (superadditive); with a drive of 7.0 the enhancement is only 18 % and the combined response is 0.59 times the sum (subadditive, near saturation). At the weakest end (drive 1.5) enhancement drops back to 116 %, because the response is dominated by spontaneous firing; so “the weaker, the more enhancement” holds in this model only over a range. Part B: for 3,000 purely additive neurons, all with true enhancement 100 %, splitting by observed unisensory response into thirds gives an average enhancement of 115.1 % in the weakest third and 57.9 % in the strongest. This is produced by regression to the mean and by the bias of “taking the maximum”. The left panel of Figure 5 shows the curve of part A.

So inverse effectiveness can come from a real non-linearity and partly from the method of analysis. Applied to ritual, “weak incense and weak sound together have the largest effect” is a testable prediction, but it needs the strength groups fixed in advance and “weak” and “strong” defined from independent data.

Smell: a main route without the thalamus, and a thalamic side route

Olfaction has a unique anatomy: the output of the olfactory bulb reaches the piriform cortex, the cortical nuclei of the amygdala, the lateral entorhinal cortex and other structures without a thalamic relay; every other sense passes through the thalamus first. The framework is right about this. But “bypassing the thalamus” needs a supplement: the piriform cortex also projects through the mediodorsal thalamus to the orbitofrontal cortex, forming a thalamic side route that may take part in olfactory attention and odour discrimination (Courtiol and Wilson 2015).

The framework singles out “terpenes”. Terpenes are a large class of molecules in many plant scents (pine, lemon, lavender); like other odorants they are detected by olfactory receptors, and there is no evidence of a special pathway for them. Human studies of inhaling a particular scent to improve mood are mostly small and hard to blind (participants can smell the scent), and this book does not present them as evidence of efficacy.

Respiration-locked oscillations and phase–amplitude coupling

Evidence

Nasal breathing entrains limbic oscillations (intracranial recording and behavioural experiments). In intracranial recordings from patients with drug-resistant epilepsy, Zelano and colleagues (2016) found that natural nasal breathing entrains slow oscillations in the piriform cortex, synchronised with oscillations in the amygdala and hippocampus; oscillatory power was highest during inspiration and dissipated when breathing was diverted from nose to mouth. Parallel behavioural experiments in healthy people found faster discrimination of fearful faces and better memory retrieval during inspiration (Zelano et al. 2016). This shows that the breathing rhythm can organise limbic activity; it is a correlation observed in a few patients and does not show that any particular “ritual breathing” has a therapeutic effect by this route.

The framework goes on to speak of “respiration-locked theta–gamma phase–amplitude coupling”. Phase–amplitude coupling (PAC) means that the phase of a slow rhythm modulates the amplitude of a fast one. The modulation index of Tort and colleagues (2010) is a common measure (Tort et al. 2010): divide the phase of the slow rhythm into NN bins, compute the mean amplitude of the fast rhythm in each bin, normalise to a distribution PP, and compute its distance from the uniform distribution, MI=(log⁡N−H(P))/log⁡N\text{MI} = (\log N – H(P))/\log N. The next program computes it on synthetic signals, with all filtering done by a fast Fourier transform written in the standard library. There are three signals: first, the phase of a 6-Hz theta rhythm modulates the amplitude of 60-Hz gamma, with the modulation stronger during the “inspiration” half of a 0.25-Hz breathing cycle; second, the same without modulation; third, a signal with no gamma oscillation at all, only a sharp, non-sinusoidal 6-Hz waveform. Significance is assessed against a surrogate distribution from 200 circular shifts of the amplitude series.

Calculation

# Phase-amplitude coupling and Tort's modulation index on synthetic signals (Chapter 6).
# Sampling 250 Hz, 16,384 samples (65.5 s). Band-pass filtering and the analytic signal
# are done with a radix-2 FFT (masking), so only the standard library is needed.
#   MI = (log N - H(P)) / log N, P = mean gamma amplitude in N = 18 theta-phase bins,
#   normalised (Tort et al. 2010). Surrogates: the amplitude series circularly shifted by
#   a random lag of at least 2 s, 200 times; z = (MI - mean) / sd of the surrogates.
# Signals (theta frequency jitters around 6 Hz): (1) theta whose phase modulates 60-Hz
#   gamma, the coupling stronger during "inspiration" of a 0.25-Hz breathing cycle;
#   (2) the same without modulation;
#   (3) a sharp, non-sinusoidal 6-Hz wave (harmonics to 15 x 6 Hz) and NO gamma at all.
import cmath, math, random

FS, N = 250.0, 16384

def fft(x, inv=False):
    a = list(x); n = len(a); j = 0
    for i in range(1, n):
        bit = n >> 1
        while j & bit:
            j ^= bit; bit >>= 1
        j |= bit
        if i < j:
            a[i], a[j] = a[j], a[i]
    size = 2
    while size <= n:
        ang = (2 if inv else -2) * math.pi / size
        wl = [complex(math.cos(ang * k), math.sin(ang * k)) for k in range(size // 2)]
        for st in range(0, n, size):
            for k in range(size // 2):
                u = a[st + k]; v = a[st + k + size // 2] * wl[k]
                a[st + k] = u + v; a[st + k + size // 2] = u - v
        size <<= 1
    return [v / n for v in a] if inv else a

def analytic(spec, lo, hi):
    out = [0j] * N
    for k in range(1, N // 2):
        if lo <= k * FS / N <= hi:
            out[k] = 2 * spec[k]
    return fft(out, inv=True)

def mi(phase, amp, bins=18):
    s, c = [0.0] * bins, [0] * bins
    for p, a in zip(phase, amp):
        b = min(int((p + math.pi) / (2 * math.pi) * bins), bins - 1); s[b] += a; c[b] += 1
    m = [s[i] / c[i] for i in range(bins)]; tot = math.fsum(m); P = [v / tot for v in m]
    return (math.log(bins) + math.fsum(p * math.log(p) for p in P)) / math.log(bins), P

def signal(kind, rng):
    x, th, y = [], 0.0, 0.0
    for i in range(N):
        t = i / FS; br = 2 * math.pi * 0.25 * t
        y += -y / (0.5 * FS) + math.sqrt(2 / (0.5 * FS)) * rng.gauss(0, 1)   # OU jitter
        th += 2 * math.pi * (6 + 0.6 * y) / FS
        if kind == "sharp":
            v = math.fsum(math.sin(k * th) / k for k in range(1, 16))
        else:
            kap = 0.0 if kind == "none" else 0.8 * (1 + math.cos(br)) / 2
            g = math.cos(2 * math.pi * 60 * t)
            v = math.cos(th) + 0.25 * (1 + kap * math.cos(th)) * g
        x.append(v + rng.gauss(0, 0.5))
    return x

rng = random.Random(2010)
for kind, lab in (("coupled", "theta-gamma coupling"), ("none", "no coupling"),
                  ("sharp", "sharp 6-Hz wave, no gamma")):
    spec = fft(signal(kind, rng))
    ph = [cmath.phase(z) for z in analytic(spec, 4, 8)]
    am = [abs(z) for z in analytic(spec, 45, 75)]
    m, P = mi(ph, am)
    sur = []
    for _ in range(200):
        lag = rng.randrange(int(2 * FS), N - int(2 * FS))
        sur.append(mi(ph, am[lag:] + am[:lag])[0])
    mu = math.fsum(sur) / len(sur)
    sd = math.sqrt(math.fsum((v - mu) ** 2 for v in sur) / (len(sur) - 1))
    print(f"{lab:27s} MI {m:.5f}  surrogates {mu:.5f} +- {sd:.5f}"
          f"  z {(m - mu) / sd:6.1f}")
    if kind == "coupled":
        br = [math.cos(2 * math.pi * 0.25 * i / FS) for i in range(N)]
        for half, sel in (("inspiration half", [b > 0 for b in br]),
                          ("expiration half", [b <= 0 for b in br])):
            mh = mi([p for p, s in zip(ph, sel) if s],
                    [a for a, s in zip(am, sel) if s])[0]
            print(f"   {half:24s} MI {mh:.5f}")
# theta-gamma coupling        MI 0.00187  surrogates 0.00012 +- 0.00008  z   21.8
#    inspiration half         MI 0.00506
#    expiration half          MI 0.00030
# no coupling                 MI 0.00006  surrogates 0.00009 +- 0.00005  z   -0.6
# sharp 6-Hz wave, no gamma   MI 0.00827  surrogates 0.00013 +- 0.00008  z  103.4

The modulated signal has MI 0.00187 against surrogates of 0.00012 ± 0.00008, z=21.8z = 21.8; MI in the inspiration half (0.00506) is far above that in the expiration half (0.00030). The unmodulated signal gives z=−0.6z = -0.6. But the third signal, with no gamma oscillation at all, has MI 0.00827 and z=103.4z = 103.4, more “significant” than the real coupling. The reason is that the sharp 6-Hz waveform contains harmonics at 48, 54, 60, 66 and 72 Hz that fall inside the 45–75 Hz “gamma band”, and their amplitude naturally rises and falls with the 6-Hz phase. Aru and colleagues (2015) discussed such artefacts in detail: non-sinusoidal waveforms, spikes and the choice of filters can all produce spurious cross-frequency coupling (Aru et al. 2015). The right panel of Figure 5 shows the phase–amplitude distributions of the three signals.

Left: multisensory enhancement against unisensory response in the model neuron (the author’s model). Right: distribution of gamma-band amplitude over theta phase in three synthetic signals: real coupling, no coupling, and a sharp 6-Hz waveform with no gamma oscillation.

So “ritual works through respiration-locked theta–gamma coupling” needs at least three steps of evidence: coupling recorded in people; artefacts of waveform and filtering excluded; and changes in coupling linked to subjective or behavioural outcomes. The literature cited by the framework supports only part of the first step (breathing entrains limbic oscillations).

Costly signals: why ritual is “hard”

Many religious rituals demand fasting, austerities, long labour or bodily pain. Irons (2001) proposed that religion can be seen as a “hard-to-fake signal of commitment”: only people who truly identify with the group are willing to pay these costs, so the costs can screen out free-riders (Irons 2001).

Evidence

Commune longevity and ritual requirements (analysis of historical records). Sosis and Bressler (2003) analysed 83 nineteenth-century American communes and found that communes demanding more costly behaviour (rules on diet, dress, sex and property) survived longer, and that this relation interacted with religiosity (Sosis and Bressler 2003). This is a correlation in historical data; the founding date, size and leadership of the communes may all affect the conclusion.

Evidence

Extreme ritual and prosocial behaviour (field quasi-experiment). At the Thaipusam festival in Mauritius, Xygalatas and colleagues (2013) studied two rituals: a low-ordeal collective prayer and singing, and the high-ordeal kavadi (body piercing and long walks carrying heavy structures). Afterwards participants could donate anonymously to the temple; high-ordeal participants and observers gave more, and the more pain participants reported, the more they gave (Xygalatas et al. 2013). Participants were not randomly assigned to the two rituals, so this is a quasi-experiment.

A screening model shows when “cost” is useful (the author’s model). Among applicants to a group, a share qq are committed, and they value membership of a group of “quality” kk (the share of committed members) at vHkv_H k; the rest are free-riders who value it at only vLkv_L k (vL<vHv_L < v_H). Everyone who joins pays the ritual cost cc.

Calculation

# A screening model of costly ritual (Chapter 6). The author's model, after the costly-
# signalling idea of Irons and Sosis. A share q of applicants are committed and value
# membership in a group of quality k (share of committed members) at v_H k; the rest are
# free-riders who value it at v_L k (v_L < v_H). Everyone who joins pays the cost c.
# Equilibrium group quality: k = q if c <= q v_L (everyone joins); free-riders join with
# a probability that leaves them indifferent, k = c / v_L, if q v_L < c < v_L; k = 1 if
# v_L <= c <= v_H; nobody joins if c > v_H. Net value to a committed member: v_H k - c.
VH, VL = 10.0, 4.0

def quality(c, q):
    if c > VH:
        return None
    if c <= q * VL:
        return q
    return min(1.0, c / VL)

def free_rider_join(c, q):
    k = quality(c, q)
    if k is None or k >= 1.0:
        return 0.0
    if k == q:
        return 1.0
    return q * (1 - k) / (k * (1 - q))        # join probability giving quality k

print(f"v_H = {VH}, v_L = {VL}; ritual helps only if q < 1 - v_L/v_H"
      f" = {1 - VL / VH:.2f}")
print("ritual cost c   q = 0.3: quality   net   free-riders join"
      "   q = 0.7: quality   net")
for c in (0.0, 1.0, 2.0, 3.0, 4.0, 6.0, 8.0, 11.0):
    row = []
    for q in (0.3, 0.7):
        k = quality(c, q)
        row.append(("-", "-", "-") if k is None else
                   (f"{k:.2f}", f"{VH * k - c:5.2f}", f"{free_rider_join(c, q):.2f}"))
    a, b = row
    print(f"   {c:5.1f} {a[0]:>17s} {a[1]:>6s} {a[2]:>14s} {b[0]:>18s} {b[1]:>6s}")
for q in (0.3, 0.5, 0.6, 0.7):
    vals = [(VH * quality(c / 100, q) - c / 100, c / 100) for c in range(0, 1001)]
    top = max(v for v, _ in vals)
    costs = [c for v, c in vals if abs(v - top) < 1e-9]
    where = " and ".join(f"{c:.2f}" for c in costs)
    tie = " (tie)" if len(costs) > 1 else ""
    print(f"q = {q:.1f}: best cost {where}{tie} with net value {top:.2f};"
          f" without ritual {VH * q:.2f}")
# v_H = 10.0, v_L = 4.0; ritual helps only if q < 1 - v_L/v_H = 0.60
# ritual cost c   q = 0.3: quality   net   free-riders join   q = 0.7: quality   net
#      0.0              0.30   3.00           1.00               0.70   7.00
#      1.0              0.30   2.00           1.00               0.70   6.00
#      2.0              0.50   3.00           0.43               0.70   5.00
#      3.0              0.75   4.50           0.14               0.75   4.50
#      4.0              1.00   6.00           0.00               1.00   6.00
#      6.0              1.00   4.00           0.00               1.00   4.00
#      8.0              1.00   2.00           0.00               1.00   2.00
#     11.0                 -      -              -                  -      -
# q = 0.3: best cost 4.00 with net value 6.00; without ritual 3.00
# q = 0.5: best cost 4.00 with net value 6.00; without ritual 5.00
# q = 0.6: best cost 0.00 and 4.00 (tie) with net value 6.00; without ritual 6.00
# q = 0.7: best cost 0.00 with net value 7.00; without ritual 7.00

With q=0.3q = 0.3, a group with no ritual (c=0c = 0) has quality only 0.30, and a committed member’s net value is 3.00; when the cost rises to vL=4v_L = 4, free-riders stop joining altogether, quality is 1.00, and net value is 6.00, the optimum; higher costs lower net value (2.00 at c=8c = 8), and above vH=10v_H = 10 nobody joins. When the cost lies between qvLq v_L and vLv_L, free-riders join with a probability that leaves them exactly indifferent (0.43 at c=2c = 2). With q=0.7q = 0.7, the best cost is 0: there are few free-riders to begin with and screening does not pay. In this model, ritual cost is useful only when q<1−vL/vH=0.60q < 1 – v_L/v_H = 0.60 (at q=0.6q = 0.6 the two are equal, and the program lists costs 0 and 4 as a tie), and the best cost is the smallest that just screens out free-riders, vLv_L; any cost beyond that is pure waste.

For the framework, the model implies that high cost is meaningful only when a group faces a free-rider problem, and that “harder” does not mean “better”. For a personal practice (with no free-rider problem) the screening model does not apply; there, the cost of a ritual is simply a cost.

Hypothesis · not established · Enhancement from low-intensity multisensory input

The author’s hypothesis (not established): in a pre-registered experiment with scent and sound each set at weak, medium and strong levels fixed in advance (determined by an independent pilot), the combination of weak scent and weak sound gives a larger proportional enhancement of self-rated relaxation (relative to the stronger single stimulus) than the combination of strong and strong. What would show it wrong: a confidence interval for the difference in proportional enhancement across levels that includes zero; or an inverse effectiveness that disappears once weak and strong are defined from independent data (which would show it to be a statistical artefact).

Chapter summary

Inverse effectiveness has a real non-linear source in single neurons, but grouping by observed values can also manufacture it (in the book’s simulation, true enhancement is 100 % throughout, yet the groups show 115.1 % and 57.9 %). The main olfactory pathway has no thalamic relay, but a side route through the mediodorsal thalamus exists; terpenes have no special pathway. The intracranial recordings of Zelano and colleagues show that nasal breathing entrains limbic oscillations; the modulation index of phase–amplitude coupling can be faked by non-sinusoidal waveforms (z=103.4z = 103.4 with no gamma oscillation). Costly signals correlate with cooperation in historical and field studies; the author’s screening model shows that ritual cost helps only when free-riders are common enough, and that the best cost is the smallest that screens.

References

Aru, Juhan, Jaan Aru, Viola Priesemann, et al. 2015. “Untangling Cross-Frequency Coupling in Neuroscience.” Current Opinion in Neurobiology 31: 51–61. https://doi.org/10.1016/j.conb.2014.08.002.
Courtiol, Emmanuelle, and Donald A. Wilson. 2015. “The Olfactory Thalamus: Unanswered Questions about the Role of the Mediodorsal Thalamic Nucleus in Olfaction.” Frontiers in Neural Circuits 9: 49. https://doi.org/10.3389/fncir.2015.00049.
Holmes, Nicholas P. 2009. “The Principle of Inverse Effectiveness in Multisensory Integration: Some Statistical Considerations.” Brain Topography 21 (3–4): 168–76. https://doi.org/10.1007/s10548-009-0097-2.
Irons, William. 2001. “Religion as a Hard-to-Fake Sign of Commitment.” In Evolution and the Capacity for Commitment, edited by Randolph M. Nesse. Russell Sage Foundation.
Meredith, M. Alex, and Barry E. Stein. 1983. “Interactions among Converging Sensory Inputs in the Superior Colliculus.” Science 221 (4608): 389–91. https://doi.org/10.1126/science.6867718.
Sosis, Richard, and Eric R. Bressler. 2003. “Cooperation and Commune Longevity: A Test of the Costly Signaling Theory of Religion.” Cross-Cultural Research 37 (2): 211–39. https://doi.org/10.1177/1069397103037002003.
Stein, Barry E., and Terrence R. Stanford. 2008. “Multisensory Integration: Current Issues from the Perspective of the Single Neuron.” Nature Reviews Neuroscience 9 (4): 255–66. https://doi.org/10.1038/nrn2331.
Tort, Adriano B. L., Robert Komorowski, Howard Eichenbaum, and Nancy Kopell. 2010. “Measuring Phase-Amplitude Coupling between Neuronal Oscillations of Different Frequencies.” Journal of Neurophysiology 104 (2): 1195–210. https://doi.org/10.1152/jn.00106.2010.
Xygalatas, Dimitris, Panagiotis Mitkidis, Ronald Fischer, et al. 2013. “Extreme Rituals Promote Prosociality.” Psychological Science 24 (8): 1602–5. https://doi.org/10.1177/0956797612472910.
Zelano, Christina, Heidi Jiang, Guangyu Zhou, et al. 2016. “Nasal Respiration Entrains Human Limbic Oscillations and Modulates Cognitive Function.” Journal of Neuroscience 36 (49): 12448–67. https://doi.org/10.1523/JNEUROSCI.2586-16.2016.

Fan, Y. (2026). Neurobiology of Ritual, Spatial Boundary Transitions and Computational Biomedicine of Self-Transcendence: From Predictive-Coding Free-Energy Minimisation, Default-Mode Network Decoupling and Vagal–Endocrine Restoration to Classical Chinese Medicine on Shifting the Spirit (Yijing Bianqi) [仪式神经学、空间边界切换与自我转移的计算生物医学] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23839910 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).