Open questions on transcriptionally informed constraint-basedmodelling

Estimating metabolic fluxes is crucial for fully characterising the landscape of cancer cell metabolism. This challenge can be addressed through the use of constraint-based metabolic models, integrated with omics data, and computing the corresponding flux distributions through flux balance analysis or flux sampling. Regarding omics data integration, several computational methods have been developed to integrate transcriptomic data with genome-scale metabolic networks. In particular, the incorporation of transcriptomic data into flux constraints is typically used when the aim is to obtain a plausible flux distribution without necessarily creating a specific model. Since single-cell transcriptomics often have false (biological) zeros, which can lead to unrealistic flux distributions, denoising methods have recently been applied to improve flux clustering and restore biologically meaningful correlations. However, even if bulk transcriptomic data are generally more robust, with a lower number of zeros, applying denoising before integration can still substantially improve flux computations, in particular to mitigate zeros in reactions essential for fundamental cellular functions. We have studied this using transcriptomics data from the Cancer Cell Line Encyclopedia and found that applying a denoising step yielded higher silhouette scores for flux clusters and led to significantly better alignment with transcriptome-based clusters. With respect to flux sampling, it is important to be aware of false discoveries, a common issue in under-sampled analyses that has been noted to be present even for large sample sizes or thinning parameters, especially for genome-wide metabolic networks characterised by a high number of reactions and constraints. Taken together, these observations raise two open questions. The first is whether a tailored denoising algorithm should be implemented as a possible mandatory preprocessing step for transcriptome integration. The second is whether alternative or more recent sampling strategies, such as exploring the corners of the feasible space instead of the internal one, can be proposed to achieve a more accurate exploration of metabolic networks and reduce false discoveries. In this work, we will show that transcriptomic denoising improves flux computation robustness and highlight the need for more reliable sampling strategies to better explore metabolic flux spaces.