3.3 Sensitivity Testing and "Sensible" fitting

"Simulation Controls"

Testing the sensitivity (steepness) of simulation outputs to perturbations in all parameters is essential for professional modelers. Such tests can be considered as equivalent to controls in experiments where for example, one tries to find flaws in the hypothesis under consideration. Their use gives confidence in the validity of conclusions for the model and enhances acceptability to the experimental community. I have, on occasion, observed considerable reluctance in acceptance of (and even hostility to) the results of simulations presented to audiences of experimenters. I believe that much of this comes about from the lack of attention to appropriate sensitivity testing by the author.

"Sensible" fitting

The meaning of the term "sensible" is rather subjective but Hodgkin and Huxley give numerous examples of the high quality of judgement implied therein. One example was in the scaling of their data to "correct" for considerable degradation of physiological condition with time. Another was to use independent data on the changes in dynamics induced by heating or cooling to adjust for the differences in temperature existing among the data sets and arrive at a better composite representation of all of their data. Otherwise it may have taken them another squid season or two at Plymouth to collect sufficient data under a variety of conditions. Yet another example was the choice of the 4th power instead of 6th for "n" in their expression for the potassium conductance. This represented a very sensible balance of calculation time vs accuracy because the higher power produced a " better fit" only at very short times after the depolarizing step change in voltage. Now, when computers are so fast that the use of a higher power for the "n" variable would require and insignificant increase in computation time, one might consider refitting their potassium conductance data with a 6th or higher power expression. Nevertheless, their equations have been so useful for so long that the slight improvement in the match to their experimental data would seem to be of only second-order or third-order importance.

Tight coupling between experiment and simulations

Probably a major reason that the contributions of Hodgkin and Huxley are so remarkable and enduring results from involvement in the whole process of developing their model: In such a setting the investigator is aware of the problems and limitations through out the experiment-simulation cycle and can arrive at the most "sensible" use of their time and efforts to arrive at an understanding of the system under investigation.

I have also found that simulations of my instruments individually and coupled to the experimental preparation extraordinarily useful. Such simulations of our voltage clamp experiments: (a) provided a unique way to evaluate the quality of data which in turn helped in knowing "sensible" limits of their interpretation, (b) showed that one should not take for granted the accuracy of the record shown on an oscilloscope, (c) showed sources and magnitudes of the errors in those records, and (d) gave valuable information as to which errors were amenable to compensation or correction.

Back to: 3.1 Fitting
Back to: 3.2 Manual Iterative Searches
Back to: Top of 3.3 Sensitivity Testing
Forward to: 3.4 Numerical Accuracy