Researchers propose that AI for science retain alternative explanations of the same data and a record of model testing
Researchers propose that AI for science retain alternative explanations of the same data and a record of model testing
On November 6, Jakob Macke, Daniel Gedon and Stefan Wahl, researchers at the University of Tübingen and the ELLIS Institute Tübingen, published an essay on scientific modelling. Different models can reproduce a nerve impulse equally well, and the authors examine how to judge what those models tell us about the actual cell.
In 1951, the physiologist Andrew Huxley spent three weeks calculating how a nerve impulse propagates using a hand-operated mechanical calculator: the university computer was unavailable. Together with Alan Hodgkin, he was developing equations to describe the electrical activity of a nerve cell. The essay’s authors identify three tasks in their work: choosing the form of the equations, fitting their parameters to measurements and understanding what the calculations reveal about the actual cell.
A model specifies, for example, how quickly channels open. These channels are proteins that allow charged particles to pass through the cell membrane. Different combinations of channel parameters can sometimes produce almost identical nerve impulse recordings. To find suitable parameters, a neural network is shown many examples, each consisting of a set of model parameters and the impulse that a program calculated using them. Once trained, the network uses an actual recording to estimate the range of parameters consistent with the measurement.
In this search, the equations themselves are specified in advance. The same recording can also be explained by different models, for example models with a different set of channels. Given a description of the system, an AI agent can write code for these models, run it and revise the equations after comparing the results with measurements. The authors propose assessing how strongly the data support each explanation and generating new candidates from the more probable models. If a model can readily be fitted to a wide variety of data, its agreement with a particular set of measurements provides less evidence in its favour. In earlier work, the authors and their colleagues applied the ModelSMC model search method to neuronal recordings and data on the hormonal regulation of blood potassium levels.
If all the models considered that agree with the measurements imply the same feature of the cell’s underlying organisation, the data probably support that feature. Predictions of how the cell will respond to an intervention can also be compared. The authors offer a hypothetical example: blocking some of the cell’s sodium channels. Agreement between predictions would increase confidence in the expected response; disagreement would suggest an experiment that could distinguish between explanations that fitted the original recording equally well. In the MDA system for selecting experiments, an agent chose an electrical pulse and a channel-blocking substance to distinguish between explanations using simulated neurons.
The more neurons a model contains, the harder it becomes to understand simply by reading its equations. The authors therefore propose retaining a record of the search for scientists: which candidates were ruled out, which model properties were determined from measurements and which explanations still need to be distinguished. Hodgkin and Huxley remembered the alternatives they had discarded because they had built and revised the model themselves.
“A scientific model can provide an answer, but understanding comes through the process of building it,” the essay’s authors write.