Generative models for building a worm's mind

The structure and function of biological neural networks are the result of biological “construction rules”,
physical limits, stochasticity, developmental noise, and learning. The nature of these building rules and
the resulting architectures are fundamental to our understanding of neuro-developmental plans and the
structure and function of neural circuits. The detailed reconstruction of connectomes makes it possible
to study their design; Recent reconstruction of connectomes during different developmental stages in C.
elegans allows for characterization of their development. We present a class of generative models for the
development of connectomes, based on neuronal birth times, probability of connections between neuronal
types, synaptic pruning, and distance. We use these models to study the formation of the connectome of C.
elegans, and compare the importance of the features specified above. We find that the number of neuronal
types dominates models’ performance, and that a small number of types is sufficient for highly accurate
predictions of individual synapses. Moreover, our models reach Area Under the Receiver-Operator Curve
values of 0.85 (on cross-validated data) using 15 neuronal types. Surprisingly, we find that models that use
multiple growth epochs (with different growth parameters in each epoch), perform as well as models that
rely on a single growth epoch in predicting the connectome of the adult worm. However, analysis of the
developmental “trajectories” of the different models shows they follow disparate paths: Models that rely on
different growth epochs and include pruning are consistent with the experimentally observed connectomes
during development, whereas models that rely on a single developmental epoch, or lack pruning, are not.
Our results suggest a framework for studying the construction of connectomes and their underlying design
principles. For C. elegans, our models predict multiple developmental epochs and the existence of an
error-correction mechanism that may be implemented by minimally regulated synaptic pruning.