Predicting how a specific patient’s tumor will respond to a new drug is one of the holy grails of precision oncology. In the lab, researchers use "organoids"—3D clusters of patient-derived tumor cells—to test various drugs at different concentrations. However, these experiments are incredibly resource-intensive, meaning we can only ever test a fraction of the possible drug-patient combinations.
A new paper from Wesley Tansey and colleagues at Memorial Sloan Kettering and Columbia University introduces Bayesian Tensor Filtering (BTF), a framework designed to fill these gaps by sharing statistical strength across similar drugs and cell lines while respecting the rigid biological laws of dose-response.
TL;DR
The authors propose Bayesian Tensor Filtering (BTF), a hierarchical model that uses low-dimensional embeddings to impute missing dose-response curves in large-scale cancer studies. By combining structured shrinkage priors with a novel inference algorithm called Generalized Analytic Slice Sampling (GASS), the model accurately captures non-stationary, monotonic curves and identifies potential new biomarkers for drug sensitivity.
What Problem Does This Paper Address?
Exploratory cancer drug studies face three major hurdles:
- Sparsity: It is physically impossible to test every drug against every patient sample at every dose. This leaves a "missing data" problem that requires sophisticated imputation.
- Experimental Noise: The way cells are pipetted into plates often introduces correlated errors (e.g., a whole column of a plate might have fewer cells by chance), making it hard to distinguish a drug's effect from a technical fluke.
- Biological Constraints: We know that cancer drugs are cytotoxic—they kill cells, they don't grow them. Therefore, a dose-response curve must be monotonic (survival can only decrease as dose increases). Standard models often struggle to enforce this constraint while still allowing for the "sharp jumps" in toxicity often seen in real data.
Methods
The authors treat dose-response modeling as a tensor factorization problem. They represent the interaction between a cell line and a drug as a product of latent embeddings.
Bayesian Tensor Filtering (BTF)
To handle the "shape" of the response, BTF places a Bayesian group trend filtering prior on the drug embeddings. Unlike standard Gaussian Process priors that assume global smoothness, trend filtering uses horseshoe shrinkage priors. This allows the model to be mostly smooth but "jump" suddenly when the data suggests a threshold of toxicity has been crossed.
Generalized Analytic Slice Sampling (GASS)
The real technical "crux" is the inference. Because the model requires hard constraints (monotonicity and survival bounded between 0 and 1), standard MCMC methods like Metropolis-Hastings suffer from high rejection rates. The authors developed GASS, a variant of elliptical slice sampling that analytically restricts the proposal distribution to the feasible region defined by linear inequalities.
Datasets
The model was evaluated on:
- Internal Pilot Study: 35 drugs vs. 28 organoids (Columbia University).
- Landscape Study: 67 drugs vs. 284 organoids (Samsung Medical Center).
- Benchmarks: The Google Flu Trends dataset and a synthetic Poisson dynamical system were used to compare BTF against state-of-the-art factor models.
Key Insights and Ideas
- Sharing Strength: By factorizing the tensor, the model learns that if Drug A and Drug B have similar chemical structures, their response curves across different cell lines should also be similar.
- Non-Stationarity Matters: Dose-response isn't always a gentle slope. Sometimes a drug does nothing until a specific concentration, then kills everything. The trend filtering approach captures these "plateaus and jumps" much better than standard splines.
- Latent Biology: Even without being told where the tumors came from, the learned organoid embeddings naturally clustered by the "site of origin" (e.g., lung vs. brain vs. ovary), proving the model captures meaningful biological signals.
Key Results
- Superior Imputation: In the cancer drug studies, BTF significantly outperformed standard log-linear logistic models and Non-negative Matrix Factorization (NMF) in terms of negative log-likelihood on held-out data.
- Benchmark Performance: On the Google Flu Trends data, BTF outperformed the popular BNP-CovReg model, fitting the data more tightly without over-smoothing the peaks.
- Biomarker Discovery: The model identified that the EGFRvIII mutation is a strong biomarker for sensitivity to Trametinib, and PTEN loss is associated with sensitivity to Neratinib. These findings align with known biological pathways, providing a "sanity check" for the model's predictive power.
Limitations and Future Work
- Grid Assumptions: The current model assumes doses are on a regularly spaced grid. Extending this to irregular grids would make it more flexible for diverse experimental designs.
- Computational Cost: The model is intensive; the large-scale landscape study required several days on a compute cluster for hyperparameter tuning. The authors suggest variational inference as a potential path toward better scalability.
- Drug Features: Currently, the model doesn't explicitly use chemical features of the drugs (like SMILES strings). Incorporating this "side information" could help predict responses for entirely new drugs that haven't been tested at all.
For those interested in implementation, the authors have made their code available at https://github.com/tansey/functionalmf.
Summary updated