Estimating Cellular Goals from High-Dimensional Biological Data

Laurence Yang, Michael A. Saunders, Jean-Christophe Lachance, Bernhard O. Palsson, José BentoKDD 2019 · DOI

TL;DR

This paper introduces BIG-BOSS, a novel computational method that can automatically infer the "cellular goal" – the underlying optimization problem a cell is trying to solve – from high-dimensional biological data. This is crucial for building more accurate models of cellular metabolism, enabling better predictions for applications like drug discovery and bioengineering.

What Problem Does This Paper Address?

Understanding how cells function at a systems level is critical for many applications, from developing new drugs to engineering microbes for industrial purposes. A key tool for this is constraint-based modeling (like COBRA), which represents cellular metabolism as a network of reactions. However, these models require knowing the cell's "goal" – what it's optimizing for (e.g., maximizing growth rate). While this is relatively straightforward for simple microbes, it's much harder for complex organisms like human cells. Existing methods for inferring these goals are limited to small models and struggle with the complexity of real-world biological systems. This paper tackles the challenge of inferring these cellular goals from data, specifically by learning new, uncharacterized biochemical reactions that contribute to the cell's overall objective, and doing so for large-scale, realistic models.

Methods

The core of this work is BIG-BOSS, a new algorithm designed to estimate new biochemical reactions (represented as columns in the stoichiometric matrix S) that best explain observed cellular behavior. It formulates this as a bilevel optimization problem, which is then transformed into a single-level optimization problem. The key innovation is the use of the Alternating Direction Method of Multipliers (ADMM), a powerful optimization technique that allows for:

  • Scalability: BIG-BOSS can handle large metabolic models (up to ~6,600 reactions and ~4,400 metabolites), a significant improvement over previous methods.
  • Parallelism: ADMM's structure enables efficient parallel computation across multiple CPU cores, speeding up the analysis.
  • Sparsity Induction: The algorithm incorporates a sparsity-inducing regularization term, encouraging the inference of simpler, more interpretable reactions.
  • Adaptive Penalty and Preconditioning: These techniques are employed to improve ADMM's convergence and reduce the need for manual parameter tuning.

The paper details the proximal operators for various components of the ADMM algorithm, including quadratic, sparse, bilinear, and boundary constraints, and discusses the use of robust numerical solvers for these operations.

Datasets

The authors evaluated BIG-BOSS using 75 large-scale metabolic network models from the BiGG database, covering a range of organisms including prokaryotes, eukaryotes, and mammalian cells. These models vary significantly in size, from 95 to 10,600 reactions and 72 to 5,835 metabolites.

Synthetic flux data was generated for training and testing by simulating cellular behavior under different nutrient environments. The "ground truth" cellular goal reaction was hidden from the algorithm during inference. The study also explored the impact of using protein concentration data in addition to metabolic fluxes, leveraging existing datasets and computational models.

Key Insights and Ideas

  • Inferring "Unknowns": The most significant insight is the ability to infer not just the importance of known reactions (like in some previous methods) but to discover entirely new biochemical reactions that are crucial for explaining cellular behavior. This moves beyond simply optimizing existing models to actively improving them.
  • ADMM as a Scalable Solution: The paper demonstrates that ADMM, when carefully implemented with adaptive penalties and efficient proximal operators, is a highly effective framework for tackling the complex, non-convex optimization problems inherent in inferring cellular goals from large biological networks.
  • Flux Prediction is Robust: Even when the inferred "goal reaction" coefficients aren't perfectly identical to the true ones, the resulting metabolic flux predictions can still be highly accurate. This suggests that cellular metabolism has a degree of robustness, and accurate flux prediction is achievable even with imperfect goal inference.
  • Multi-omics Integration Potential: The study hints at the power of integrating different data types. While metabolic fluxes alone provided limited predictive power, combining them with protein concentration data significantly improved the accuracy of predicting cellular states.

Key Results

  • Scalability: BIG-BOSS successfully processed metabolic models with up to ~6,600 reactions and ~4,400 metabolites, a significant leap from previous methods limited to ~70 reactions.
  • Prediction Accuracy: BIG-BOSS achieved high accuracy in predicting metabolic fluxes, with median R² values of 0.993 for training data and 0.977 for test data across 75 models.
  • Goal Recovery: The inferred goal reactions showed a median Pearson correlation of 1.0 with the true goal reactions used for simulation, indicating good recovery of the underlying cellular objective.
  • Robustness to Missing Data: While missing flux measurements reduced accuracy, BIG-BOSS could still achieve reasonable flux predictions (R² > 0.90) in some cases, especially when important fluxes were measured.
  • Multi-core Speedup: The fine-grained implementation of BIG-BOSS demonstrated significant speedups, achieving up to 8x speedup with 64 CPU cores for certain problem sizes.
  • Protein Data Integration: Combining metabolic fluxes with protein concentration data led to a substantial improvement in prediction accuracy (R² = 0.567) compared to using fluxes alone (R² = 0.281).

Limitations and Future Work

  • Non-convexity: The problem of inferring cellular goals remains non-convex, meaning that while ADMM can find good solutions, global optimality is not always guaranteed.
  • Data Quality: The accuracy of inferred goals is dependent on the quality and coverage of the input data. Missing crucial measurements can significantly impact performance.
  • Computational Cost: While scalable, inferring goals for very large models can still be computationally intensive, especially without preconditioning or adaptive penalty tuning.
  • Future Work: The authors suggest exploring the use of more data types (e.g., transcriptomics), developing more sophisticated methods for handling missing data, and further optimizing the ADMM implementation for even greater efficiency. The potential to use BIG-BOSS for analyzing large gene expression datasets, particularly in areas like cancer research, is highlighted.

Summary updated

BibTeX
@inproceedings{yang2019estimating,
  author = {Laurence Yang and Michael A. Saunders and Jean-Christophe Lachance and Bernhard O. Palsson and José Bento},
  title = {{Estimating Cellular Goals from High-Dimensional Biological Data}},
  booktitle = {ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2019},
  doi = {10.1145/3292500.3330775},
}
Figure