Winkler, Burden, and Halley’s recursive neural network--an extension of Geard and Wiles’ work [1]--models the differentiation pathway in the C. elegans worm. The neural network models stem cell data and predicts gene expression behavior. The authors provide a more efficient Python implementation of Geard and Wiles’ original code, and add “a regularization step that enforces parsimony on the model.” They apply the resulting implementation to experimental observations in order to create a working model of embryogenesis in C. elegans.

The authors provide a detailed description of the model that includes a laudable discussion of its limitations, especially when applied to mammalian data. The expression is modeled as a binary event rather than a continuous biological reality, and the weight matrix governing the interactions in the model is an extreme simplification of the cell’s actual complex biological processes. Additionally, the model is limited to a small number of functional and regulatory genes. Although this limitation is negligible for small reference organisms such as C. elegans, it is possibly prohibitive for larger mammalian organisms.

The only criticism of this well-written paper is that it lacks a link to the source code, in order to recreate the analysis. The authors plan to apply this “modeling framework for embryogenesis and differentiation” to mammalian stem cell data.