Preprint: eight combinations of connection weight changes had the greatest effect on signals controlling movement in a worm model

Preprint: eight combinations of connection weight changes had the greatest effect on signals controlling movement in a worm model

A preprint published on September 7 examines how many internal parameter settings in a neural model can be distinguished from its motor output: the electrical signals from motor neurons that control movement in the simulation.

BAAIWorm, a model of the roundworm Caenorhabditis elegans, represents 136 neurons. The chemical and electrical connections between them have weights, numerical values that determine how strongly one cell influences another.

The author varied each of the 3 076 weights in turn and recorded the electrical signals of 80 motor neurons over time, producing 9 600 values in total. This produced a sensitivity map showing which coordinated changes across multiple connections have the greatest effect on motor output.

Eight directions of coordinated change accounted for 90% of the sensitivity, and 47 accounted for 99%. Here, a direction is a combination of changes across many weights. Changes to 3 072 individual connections produced a measurable response; several such changes could offset one another and leave motor output similar.

The author then tested this approach on cellular models with known original parameter settings. In simple neuronal models, fitting recovered the original values. In a model of the MAPK cascade, a chain of chemical signals within a cell, fitting closely reproduced the target oscillations despite a median parameter error of 40%. Different internal parameter settings can therefore produce similar observed trajectories.

In a separate test of 20 coarse-grained parameters, chemotaxis (movement toward food along a chemical gradient) and ordinary locomotion each distinguished four directions. Together, they distinguished six directions: tasks that require different functions from the system reveal different aspects of its internal parameter settings.

More precise tests of the internal workings of these simulators require trials in which the system performs different functions.