Signal Amplification Versus Fidelity In Cascades

Amplification Is Not the Same as Fidelity

A signaling cascade can produce a large change in its final output when an input changes. That does not, by itself, mean the cascade preserves information about the input well. The same layers that amplify a signal can also amplify fluctuations, and the effect depends on where those fluctuations enter the cascade. To compare cascades, we therefore need to ask two distinct questions: how sensitive is the output to a change in input, and how large is the output variation relative to the separation between inputs of interest?

The calculations here are local: they describe small changes around a specified operating point. This makes them useful for comparing cascade behavior near a steady state, but they do not describe large input changes automatically. Earlier material on gain and attenuation in linear signaling cascades introduced the basic idea that layer responses combine; here we use local response calculations to connect that combination to noise and discrimination.

The distinction between the two gains matters. A positive-valued response need not have a positive slope: a decreasing response can have negative gain. For positive, nondecreasing layers, gains are nonnegative; for positive, strictly increasing layers with positive slopes, gains are positive. The sign indicates whether a small input increase raises or lowers the output. The magnitude indicates the size of the local response.

Combining Gains Across Layers

For the full cascade, define \(F=f_m\circ f_{m-1}\circ\cdots\circ f_1\), so \(y_m=F(y_0)\). Applying the chain rule gives the absolute local gain of the cascade as the product of its layer gains. For positive operating-point values, the corresponding logarithmic gains also multiply.

The second equality follows because \(d\ln y_m/d\ln y_0=(y_0/y_m)(dy_m/dy_0)\), and each \(E_i=(y_{i-1}/y_i)g_i\). Multiplying the layer expressions cancels the intermediate factors \(y_1,\ldots,y_{m-1}\), leaving \((y_0/y_m)\prod_i g_i\). Thus, a local relative input change \(\delta y_0/y_0\) produces, to first order, a relative output change of approximately \(E_{\mathrm{cascade}}\delta y_0/y_0\). A negative product means the final output moves in the opposite direction from the input.

The product also shows why the location of a strong layer matters. Two cascades can have the same total gain but different intermediate responses. Those differences matter when fluctuations enter at different layers: a fluctuation introduced early is passed through more of the cascade than one introduced near the output.

How Small Fluctuations Propagate

To make noise propagation precise, work with logarithmic deviations around positive operating-point values. Let \(\xi_i\) be a small random fluctuation added to \(\ln y_i\) immediately after layer \(i\), and let \(\xi_0\) be a small fluctuation in the input log value. Assume the response is smooth enough to linearize locally. To first order, a layer maps its incoming log deviation to \(E_i\) times that deviation, then adds its own fluctuation:

Repeated substitution shows that input fluctuations are multiplied by the full cascade gain, while a fluctuation after layer \(i\) is multiplied only by the gains of the downstream layers. Write \(P_{i+1:m}=\prod_{j=i+1}^{m}E_j\), with an empty product equal to 1. Then:

If the fluctuations are mutually independent, have mean zero, and have variances \(v_i=\operatorname{Var}(\xi_i)\), this linear approximation gives:

The variances are multiplied by squared gains, so the sign of a gain does not change its contribution to variance. If fluctuations are correlated, covariance terms must also be included; the independent-noise formula is then not sufficient. These equations describe small, local fluctuations, not arbitrary large stochastic excursions.

Discrimination Depends on Separation Relative to Noise

A simple local measure of discrimination compares the output separation caused by two nearby inputs with the output variability. For positive inputs \(y_0^{(A)}\) and \(y_0^{(B)}\), define their small log separation as \(\Delta_0=\ln y_0^{(B)}-\ln y_0^{(A)}\). Locally, their mean output log separation is approximately \(\Delta_m=E_{\mathrm{cascade}}\Delta_0\). If output log noise has standard deviation \(s_m\) in each condition and is comparable between them, a useful separation-to-noise measure is \(d_{\mathrm{local}}=|\Delta_m|/s_m\). It is a local comparison, not a universal probability of correct classification.

If there is no downstream noise and only input variation, both the input separation and input fluctuations are multiplied by the same cascade gain. Increasing gain alone then does not improve their ratio. By contrast, if important noise is added after an early layer, changing layer strengths can alter how much that noise reaches the output while preserving the product of gains. Amplification can therefore raise output magnitude without improving fidelity, or can improve local separation relative to specific downstream noise, depending on the noise sources and where they enter.

Assumptions, Diagnostics, and Limitations

These local calculations rely on a smooth response near the chosen operating point, small perturbations, positive values for logarithmic gains, and a specified noise model. The independent-variance calculation additionally requires independent, mean-zero fluctuations. Check these conditions before interpreting the result. For example, positive-valued outputs are required to take their logarithms, but positivity alone does not imply a positive slope. A decreasing positive layer has a negative local gain and can reverse the direction of a signal.

A useful diagnostic is to compare the predicted local change with direct evaluation of the response functions at nearby inputs. If the finite-change response differs substantially from the derivative-based estimate, the perturbation may be too large or the operating point may be near a steep or saturating region. Local gain can vary with the operating point, so report where it was evaluated. If measured fluctuations are large, skewed, state-dependent, or correlated across layers, a first-order independent-noise approximation may misrepresent output variability; simulation or a fuller stochastic model may be needed.

Local gain is not the only way to compare cascades. Finite-difference calculations evaluate outputs at the actual input values of interest and do not rely on a small-change approximation, though they do not by themselves describe noise. Stochastic simulation can represent nonlinear responses and complex noise, but requires a defined model and enough replicates to estimate variability reliably. Neither alternative removes the need to specify the operating conditions and noise sources. A local gain calculation is most useful for identifying which layers control sensitivity and which noise sources are passed forward.