Validation and Benchmarking of Falcon, a Robust and Efficient Method for Estimating Enzyme Complex Abundance and Metabolic Flux from Expression Data

The integration of gene expression with constraint-based models is designed to give more accurate flux predictions, particularly in multicellular organisms such as humans that are the focus of biomedical applications. We present a new algorithm in this area called FALCON (flux assignment with LAD convex objectives and normalization)¹. In contrast to most other methods, FALCON minimize the sum of the absolute deviation between enzyme expression and flux. FALCON also implements three improvements over previous work in this area^1.4: 1. It uses an improved method for partitioning gene expression among different protein complexes, based on transforming GPR rules to conjunctive normal form (CNF). 2. FALCON assigns reaction directions in batch, resulting in dramatic speedup while outputting similar flux estimates. 3. FALCON automatically selects a reaction to scale the flux distribution by. We show that together, these changes result in considerably better correlation of FALCON than other methods on a set of yeast exometabolic fluxes.⁴

            We also benchmark FALCON's performance in two other ways. First, we tested the assumption of uniform complex stoichiometry in FALCON, which is necessary due to limited current knowledge, by modifying the Whole Cell model of Karr et al³. The Whole Cell model contains experimentally measured complex stoichiometry. However, by modifying it to use uniform instead, we find that intracellular flux estimates between the two settings are remarkably well-correlated, justifying the assumption of FALCON¹. Second, we also predicted the consumption and release (CORE)² exometabolic profiles of the NCI-60 cancer cell lines using FALCON, and compared these to the results from FBA and a panel of previous gene expression-based methods. A previous comparison of methods in this area has focused only on comparing these methods against FBA in bacterial metabolism⁵. Our preliminary results indicate that, unexpectedly, FBA still performs better than most gene expression-based methods. However, FALCON's performance is at least comparable to most other expression-based methods. Furthermore, cancer cells may not be representative of all eukaryotic cells, due to their rapid growth. In future, we plan to also run FALCON on nongrowing tissue-specific expression datasets, and benchmark its performance there by comparison to manually curated tissue-specific models, and literature descriptions of tissue-specific metabolic activity.

1. Barker, Sadagopan, Wang, et al. "A robust and efficient method for estimating enzyme complex abundance and metabolic flux from expression data." arXiv:1404.4755. 2015.

2. Jain, Nilsson, et al. "Metabolite Profiling Identifies a Key Role for Glycine in Rapid Cancer Cell Proliferation." Science. 2012. 336:1040-1044

3. Karr, Sanghvi, et al. "A Whole-Cell Computational Model Predicts Phenotype from Genotype." Cell. 2012. 140:389-401

4. Lee, Smallbone, et al. "Improving metabolic flux predictions using absolute gene expression data." BMC Sys. Biol. 2012. 6:73

5. Machado and Herrgard. "Systematic Evaluation of Methods for Integration of Transcriptomic Data into Constraint-Based Models of Metabolism." PLoS Comp. Biol. 2014. 10: e1003580