Bayesian Tensor Filtering: Smooth, Locally-Adaptive Factorization of Functional Matrices

Wesley Tansey, Christopher Tosh, David M. Blei

Introduction

In the high-stakes field of oncology, identifying the optimal drug for a specific patient is a primary objective. Biologists conduct extensive exploratory studies, testing hundreds of pharmacological agents against various tumor cell lines or "organoids" across a range of concentrations. The resulting "dose-response curves" provide a critical map of drug efficacy. However, these studies are frequently compromised by missing data, experimental noise—such as technical "pipetting errors"—and the complex, non-linear nature of biological systems.

The Challenge of Modeling Dose-Response

A significant challenge in this domain is the "cold-start" problem: predicting the response for a (sample, drug) pair that has not been experimentally tested. Standard statistical models often rely on Gaussian noise assumptions, which fail to capture the skewed technical errors inherent in laboratory data. Furthermore, biological systems impose a strict monotonicity constraint; a cytotoxic drug should ideally only decrease cell survival as concentration increases. Most conventional machine learning models struggle to enforce these fundamental biological rules.

Methodology: Bayesian Tensor Filtering (BTF)

To address these limitations, Tansey et al. proposed Bayesian Tensor Filtering (BTF), a hierarchical framework that integrates factor analysis with adaptive smoothing. The BTF pipeline utilizes several key innovations:

  1. Functional Factorization: Unlike standard matrix factorization, BTF treats drug attributes as evolving across dose levels, enabling the model to "share statistical strength" between similar drugs and cell lines.
  2. Group Trend Filtering: The model employs "horseshoe priors" on the differences between dose levels. This encourages global smoothness while allowing for sharp transitions or "cliffs" where a drug's effect suddenly intensifies.
  3. Generalized Analytic Slice Sampling (GASS): This novel sampling algorithm handles hard linear inequality constraints (like monotonicity) and non-Gaussian likelihoods more efficiently than traditional MCMC methods.
  4. Empirical Bayes Likelihood: To account for technical noise, the researchers utilized a mixture-of-Gammas likelihood, providing greater robustness against outliers and systematic errors.
Figure 2: Dose-response curve for Organoid 1 and Drug 2, illustrating the model's ability to capture sharp efficacy thresholds while maintaining monotonicity.

Results and Insights

The BTF model was rigorously evaluated using both public benchmarks and large-scale biological datasets, such as the Samsung Medical Center "Landscape" study.

  • Predictive Accuracy: BTF outperformed state-of-the-art models like BNP-CovReg on time-series benchmarks. In cancer studies, it achieved significantly lower Negative Log-Likelihood (NLL) than standard logistic factor models.
  • Adaptive Smoothness: The model effectively balances smoothness with the need to capture sudden biological shifts, as seen in Figure 2.
  • Biomarker Discovery: BTF successfully identified known and novel biomarkers, linking specific mutations (e.g., EGFRvIII, PTEN loss) to drug sensitivities (e.g., Trametinib, Neratinib).
Figure 3: Dose-response curve for Organoid 2 and Drug 3, demonstrating the model's performance on non-responsive samples.

Conclusion and Future Directions

Bayesian Tensor Filtering offers a powerful approach to modeling the "messy reality" of cancer drug response. By incorporating biological constraints directly into the inference process and utilizing adaptive smoothing, it provides more accurate and biologically grounded predictions. While currently computationally intensive, future developments in Variational Inference and the incorporation of raw chemical features may further enhance the scalability and utility of this framework in precision oncology.

Summary updated

BibTeX
@article{tansey2019bayesian,
  author = {Wesley Tansey and Christopher Tosh and David M. Blei},
  title = {{Bayesian Tensor Filtering: Smooth, Locally-Adaptive Factorization of Functional Matrices}},
  journal = {The Annals of Applied Statistics},
  year = {2019},
}
Figure