Boolean network inference and robust solution identification applied to human embryology

Understanding human embryonic development is crucial for improving in vitro fertilization protocols, as only about 25% result in a viable pregnancy. A key step in this process is embryo implantation in the uterus. We therefore focus on the trophectoderm (TE), the outer cell layer of the blastocyst responsible for attachment to the endometrium. In particular, we investigate the medium and late stages of TE development using Boolean networks inferred from single-cell transcriptomic data. The SCIBORG framework was previously used to infer families of Boolean networks from scRNA-seq data combined with prior biological knowledge. Although the inference procedure is well defined and straightforward to apply, the impact of input parameters was not explored, and the space of near-optimal solutions remained unexamined.We extend the SCIBORG framework through four contributions: (i) an updated prior knowledge network (PKN) derived from Pathway Commons v14, (ii) a comparison of two PKN reduction functions, (iii) a gene-specific binarization strategy based on Gaussian Mixture Models (GMMs), replacing the previously used fixed threshold and (iv) an analysis of alternative Boolean network solutions close to optimality. Using an adapted Jaccard index and a stage-discrimination evaluation, we compared networks inferred with both binarization approaches and previously published models, retaining those with the best accuracy scores. The GMM-based approach yields fewer candidate networks while maintaining strong discrimination between developmental stages, with accuracy scores reaching 68%. Moreover, relaxing SCIBORG’s optimality criteria reveals near-optimal solutions that explain cell behaviors in test data with improved accuracy compared to previous models.