Feedback Regulation of Gene Expression by Protein

Why Protein Feedback Changes Expression Dynamics

A protein can regulate the transcription of its own gene. If more protein suppresses transcription, the loop is negative feedback; if more protein stimulates transcription, the loop is positive feedback. The effect is not limited to changing the amount of mRNA produced. Feedback can shift the steady protein level, alter how quickly the system returns after a perturbation, and, with sufficiently strong positive feedback, allow more than one steady state.

To isolate these effects, use the coupled mRNA–protein turnover model with first-order mRNA and protein loss. Let $M$ be mRNA concentration and $P$ protein concentration. Transcription rate $r(P)$ depends on current protein concentration; translation and loss rates are represented by constants. The model assumes a well-mixed system, constant parameters, nonnegative states, and no explicit delay in transcriptional regulation.

Model: The coupled equations are $$ \frac{dM}{dt}=r(P)-k_mM,\qquad \frac{dP}{dt}=k_tM-k_pP, $$ where $k_m>0$ and $k_p>0$ are first-order loss-rate constants, and $k_t>0$ is the translation rate constant. The function $r(P)$ is a nonnegative transcription rate. Its units are mRNA concentration per time; $r'(P)$ has units of inverse time when $M$ and $P$ are expressed in the same concentration units.

The use of first-order loss here makes the feedback calculation transparent. If protein loss is instead capacity-limited, the equilibrium balance and Jacobian must include the derivative of that loss flux; the simple threshold below applies specifically to first-order protein loss.

Hill Functions for Negative and Positive Feedback

A Hill function describes a saturating response to regulator concentration. The Hill coefficient $h>0$ controls the steepness of the response, and $K>0$ is the protein concentration at which the regulated component reaches half of its maximum. Two common transcription-rate forms are:

r_{\text{neg}}(P)=r_b+\frac{V}{1+(P/K)^h}, \qquad r_{\text{pos}}(P)=r_b+V\frac{P^h}{K^h+P^h}.

Here $r_b\geq 0$ is a basal transcription rate and $V>0$ is the maximum regulated contribution. In the negative-feedback form, transcription falls from $r_b+V$ toward $r_b$ as protein increases. In the positive-feedback form, it rises from $r_b$ toward $r_b+V$. The derivatives make the direction of regulation explicit:

r'_{\text{neg}}(P)=-\frac{VhK^hP^{h-1}}{(K^h+P^h)^2}, \qquad r'_{\text{pos}}(P)=\frac{VhK^hP^{h-1}}{(K^h+P^h)^2}.

For $P>0$, the negative-feedback slope is nonpositive and the positive-feedback slope is nonnegative. The Hill coefficient affects the slope over the transition range, so the feedback direction alone does not determine the size of the dynamical effect.

Steady States and the Local Stability Criterion

At a steady state, both derivatives are zero. The protein equation gives $M^*=(k_p/k_t)P^*$, where a star denotes an equilibrium value. Substituting this into the mRNA equation gives a single condition for the steady protein concentration:

r(P^*)=\gamma P^*, \qquad \gamma=\frac{k_mk_p}{k_t}.

Thus, a steady state is an intersection of the transcription-rate curve $r(P)$ and the straight line $\gamma P$. Once $P^*$ is found, $M^*$ follows from the protein balance. This reduction finds equilibria, but it does not by itself determine whether nearby trajectories approach or move away from them.

For local stability, linearize the two-state system around an equilibrium. The Jacobian matrix is

J= \begin{pmatrix} -k_m & r'(P^*)\\ k_t & -k_p \end{pmatrix}, \qquad \operatorname{tr}(J)=-(k_m+k_p), \qquad \det(J)=k_mk_p-k_tr'(P^*).

Local stability result: Since $k_m+k_p>0$, the trace is negative. If $\det(J)>0$, both eigenvalues have negative real parts and the equilibrium is locally asymptotically stable. If $\det(J)<0$, the equilibrium is a saddle and is unstable. If $\det(J)=0$, linearization is inconclusive; nonlinear terms must be examined.

The determinant condition compares the feedback slope with the turnover terms. Negative feedback has $r'(P^*)\leq 0$, so $\det(J)\geq k_mk_p>0$: every such equilibrium is locally stable under this model. For positive feedback, stability requires $k_tr'(P^*)<k_mk_p$. A sufficiently steep positive-feedback slope can make the determinant negative and destabilize an equilibrium.

Worked Examples

Worked Example: A Negative-Feedback Steady State

Consider $r(P)=2+8/(1+P/2)$, with $k_m=k_p=k_t=1$ in consistent normalized units. Then $\gamma=1$, and the steady-state condition is $r(P)=P$. Find the steady protein concentration and assess local stability.

Rewrite the rate as $r(P)=2+16/(P+2)$. The steady-state equation becomes

P=2+\frac{16}{P+2} \quad\Longrightarrow\quad P(P+2)=2(P+2)+16 \quad\Longrightarrow\quad P^2=20.

The physically relevant solution is $P^*=\sqrt{20}\approx4.472$ concentration units. The other algebraic solution, $-\sqrt{20}$, is not a nonnegative protein concentration. The mRNA steady state is $M^*=P^*\approx4.472$ concentration units because $M^*=(k_p/k_t)P^*$ and both rate constants equal 1.

Differentiate the transcription rate: $r'(P)=-16/(P+2)^2$. At the equilibrium, $r'(P^*)\approx-16/(6.472)^2\approx-0.3820$. Therefore $\det(J)=1-r'(P^*)\approx1.382>0$, and the trace is $-2$. The equilibrium is locally asymptotically stable.

Worked Example: Three Steady States Under Positive Feedback

Use the positive-feedback rate $r(P)=1/13+(99/65)P^2/(13/20+P^2)$, with $k_m=k_p=k_t=1$, so $\gamma=1$. Concentrations and time are in normalized units. Verify three steady states and classify their local stability.

The steady-state condition is $r(P)=P$. Substituting $P=0.1$, $0.5$, and $1$ gives the corresponding transcription rates $0.1$, $0.5$, and $1$, respectively, so all three values are equilibria. Their mRNA concentrations equal the corresponding protein concentrations.

The feedback slope must be differentiated carefully. Since \[ \frac{d}{dP}\left(\frac{P^2}{a+P^2}\right) =\frac{2aP}{(a+P^2)^2}, \] with $a=13/20$, the correct derivative is

r'(P)=\frac{99}{65}\frac{2(13/20)P}{(13/20+P^2)^2} =\frac{99}{50}\frac{P}{(13/20+P^2)^2}.

At $P=0.1$, $r'(P)=0.198/(0.66)^2\approx0.4545$. At $P=0.5$, $r'(P)=0.99/(0.90)^2\approx1.222$. At $P=1$, $r'(P)=1.98/(1.65)^2\approx0.7273$. Because $k_mk_p=k_t=1$, the determinant is $1-r'(P^*)$. It is positive at $0.1$ and $1$, and negative at $0.5$. Thus the first and third equilibria are locally asymptotically stable, while the middle equilibrium is a saddle. The calculation shows how positive feedback can support multiple steady states with different local stability.

Worked Example: Stability Can Include Damped Oscillations

Return to the negative-feedback equilibrium in the first example. Use its slope $r'(P^*)\approx-0.3820$ to determine whether the local return is monotone or oscillatory.

The Jacobian has trace $-2$ and determinant $1.3820$. For a two-dimensional system, the eigenvalues solve $\lambda^2-\operatorname{tr}(J)\lambda+\det(J)=0$. Here,

\lambda^2+2\lambda+1.3820=0, \qquad \lambda=\frac{-2\pm\sqrt{4-4(1.3820)}}{2} \approx -1.000\pm0.618\,i.

The negative real parts confirm local stability, while the nonzero imaginary parts indicate damped oscillations near the equilibrium. This is a local prediction: it does not imply that all trajectories, including distant initial conditions, follow an oscillatory path.

Interpreting the Result and Checking the Model

A useful way to visualize steady states is to plot $r(P)$ and $\gamma P$ on the same axes. For negative feedback, the transcription curve is nonincreasing while the balance line rises, giving one intersection when transcription is positive at zero and the line eventually exceeds it. For positive feedback, the curve can bend upward over part of its range; it may cross the balance line once or several times. As emphasized in the earlier material on bistability, multiple intersections alone do not establish bistability: their local stability and the behavior between them also matter.

The stability result follows directly from the characteristic polynomial. For a $2\times2$ matrix, the eigenvalues have sum $\operatorname{tr}(J)$ and product $\det(J)$. When the trace is negative and determinant positive, either both eigenvalues are real and negative, or they are a complex-conjugate pair with negative real part. When the determinant is negative, their product is negative, so one eigenvalue is positive and one is negative. This justifies the classifications used above.

Several practical checks prevent common errors:

Check the slope by differentiating the full Hill function. For $P^2/(a+P^2)$, the derivative is $2aP/(a+P^2)^2$; omitting the factor $2a$ gives an incorrect stability threshold.
Check units before comparing terms. The terms $k_mk_p$ and $k_tr'(P^*)$ both have units of inverse time squared, so they can be compared in the determinant.
Check all nonnegative roots. Numerical solvers may return negative roots or miss roots near a tangency. A plot, a root scan, and substitution back into $r(P)=\gamma P$ provide useful checks.
Do not infer stability from the direction of feedback alone. Negative feedback is locally stabilizing in this model, but positive feedback can still be locally stable when its slope is below the threshold.
Do not infer global behavior from a local calculation. The Jacobian describes perturbations sufficiently close to one equilibrium, not the full trajectory from every initial condition.

The full two-state model is preferable when mRNA and protein turnover occur on comparable timescales or when the return dynamics matter. A reduced protein-only equation obtained by assuming mRNA rapidly equilibrates can be useful when that timescale separation is justified, but it may misrepresent transient behavior otherwise. Hill feedback is also a phenomenological approximation: it does not explicitly represent promoter states, binding kinetics, transcriptional delays, cell-to-cell variability, or stochastic expression. These factors may be important when interpreting experimental time courses.

Key takeaway: Find feedback-regulated steady states from $r(P^*)=\gamma P^*$, then assess each one with $\operatorname{tr}(J)=-(k_m+k_p)$ and $\det(J)=k_mk_p-k_tr'(P^*)$. The feedback slope at each equilibrium—not just whether feedback is positive or negative—determines the local stability classification.

Check Your Understanding

Answer each question, then compare with the solutions below.

For the model in this tutorial, derive the steady-state relation between $M^*$ and $P^*$, and the equation that determines $P^*$.
If $k_mk_p=2$, $k_t=1$, and $r'(P^*)=1.5$, classify the equilibrium locally using the trace and determinant.
Why can positive feedback produce multiple steady states, while a negative-feedback transcription curve in this model has a unique steady state?
For $g(P)=P^2/(a+P^2)$, derive $g'(P)$. Explain the error in using $P/(a+P^2)^2$ as its derivative.

Solutions. (1) Setting $dP/dt=0$ gives $M^*=(k_p/k_t)P^*$. Substitution into $dM/dt=0$ gives $r(P^*)=(k_mk_p/k_t)P^*=\gamma P^*$. (2) The determinant is $2-(1)(1.5)=0.5>0$, and the trace is $-(k_m+k_p)<0$, so the equilibrium is locally asymptotically stable. (3) A nonincreasing negative-feedback transcription curve intersects the increasing line $\gamma P$ once under the stated positive-rate conditions. Positive feedback can rise steeply over part of its range, allowing multiple intersections; each intersection must still be checked for stability. (4) The quotient rule gives $g'(P)=[2P(a+P^2)-2P^3]/(a+P^2)^2=2aP/(a+P^2)^2$. The incorrect expression omits the factor $2a$, which changes the slope and can lead to a wrong stability conclusion.