Kowshika Sarker - Research

Kowshika Sarker, Ruoqing Zhu, Hannah D. Holscher, and ChengXiang Zhai. Prior-guided longitudinal metabolomic analysis. BIBM 2024.

We conduct the first study of integrating prior knowledge from a genome-scale metabolomic model (GEM) regarding metabolic reactions and subsystems with experimental longitudinal metabolomes to yield longitudinal heuristic reaction- and subsystem-level features. For each sample in the metabolomic dataset, we construct a directed knowledge graph representing the causal mechanism of the changes in the activity of different metabolic entities (metabolite → subsystem → reaction → metabolite) using the GEM. As node features, we use concentrations at the metabolite nodes and zero at the reaction and subsystem nodes. We apply a graph convolutional recurrent neural network(GCRNN) approach to this knowledge graph for predictive modeling of the metabolite concentrations at the next timestamp from the current timestamp. The predictions at the reaction and subsystem nodes contribute to the predictions at the neighboring metabolite nodes and the GCRNN model parameters are shared across all nodes of a graph. We hypothesize that if the model learns to understand the metabolite changes, the predicted changes at reactions and subsystems are informative regarding the participation of these entities in the metabolite concentration changes. We evaluate our approach by applying the predicted heuristic reaction and subsystem
representative features to downstream clustering and classification and show that these novel features preserve biological signals, and improve downstream tasks.

Kowshika Sarker, Ruoqing Zhu, Hannah D. Holscher, and ChengXiang Zhai. Augmenting nutritional metabolomics with a genome-scale metabolic model for assessment of diet intake. ACM-BCB 2023.

We integrate prior domain knowledge with metabolomic profiles and propose multiple methods such as knowledge network analysis with random walk with restart algorithm to engineer novel predictive features representing heterogeneous unmeasured metabolic entities - namely metabolic reactions and metabolic subsystems. Our empirical experiments show that such domain-guided analyses can offer novel hypotheses on underlying metabolic dynamics and improve downstream empirical performance.

Workflow

Prior knowledge is integrated with experimental metabolomic profiles to engineer reaction and subsystem-level novel features. Downstream tasks demonstrate utility of the engineered features.

Knowledge Network

One of the proposed feature engineering methods constructs a heterogeneous network merging prior knowledge with metabolome.

We analyze the network with a random walk with restart algorithm to engineer reaction and subsystem-level features, which can offer fine-grained novel hypotheses and improve downstream performance.