Introduction
In computational biology, researchers often develop complex, stochastic models—such as agent-based models (ABMs)—to simulate individual cell behaviors and understand systems like tumor growth or tissue development. These models are frequently "likelihood-free," meaning it is mathematically impossible or computationally prohibitive to derive a formula for the probability of observing specific data given a set of parameters.
Simulation-Based Inference (SBI) addresses this by using model simulations to estimate parameters. However, a growing divide exists between traditional statistical SBI (e.g., Approximate Bayesian Computation) and modern neural SBI (using deep learning). This paper provides a roadmap for navigating these choices, particularly when dealing with noisy, real-world biological data that often deviates from idealized models.
Methods and Approaches
The study evaluates four primary SBI approaches across two complex biological systems:
- Statistical SBI:
- Approximate Bayesian Computation (ABC): Specifically, an adaptive Sequential Monte Carlo (SMC) version that compares simulated and observed data using distance metrics.
- Bayesian Synthetic Likelihood (BSL): Approximates the likelihood of summary statistics using a multivariate Gaussian distribution.
- Neural SBI:
- Neural Posterior Estimation (NPE/SNPE): Uses normalizing flows to directly learn the mapping from data to the posterior distribution.
- Neural Likelihood Estimation (NLE/SNLE): Employs neural networks to learn the likelihood function itself.
The authors also examine "robust" variants (like RSNLE and RBSL) designed to mitigate errors when a model is misspecified—a common occurrence in biological research where models are necessarily simplifications of reality.
The Three-Stage Workflow
To help researchers choose the right tool for their specific problem, the authors propose a structured three-stage guideline.
- Stage 1: Pre-analysis: Assessing model compatibility and identifiability before running expensive simulations.
- Stage 2: SBI Stage: Selecting and executing the inference algorithm (ABC, SNPE, etc.) based on the simulation budget and data complexity.
- Stage 3: Uncertainty Quantification: Evaluating the reliability of the results and checking for model misspecification.
Biological Applications
The methods were tested using two distinct biological models: a tumor growth model and a cell invasion model. These models incorporate mechanisms such as cell proliferation and movement, which are often non-identifiable without high-quality data.
The study utilized both synthetic datasets (where ground truth is known) and real-world data, including pancreatic tumor growth time-series and image-based cell invasion data using FUCCI (Fluorescent Ubiquitination-based Cell Cycle Indicator) technology.
Key Results and Discussion
The Efficiency-Accuracy Trade-off
Neural SBI methods proved to be significantly more efficient, often requiring 5x to 7x fewer simulations than statistical methods. For example, in the tumor growth model, Neural SBI used approximately 100,000 simulations, while SMC ABC required between 500,000 and 700,000. However, this speed comes with a risk: neural networks are "amortized," meaning they learn from a specific training set. If real-world data falls "out-of-distribution" due to noise or model errors, these networks can produce confident but incorrect predictions.
Robustness of Statistical Methods
While slower, statistical methods like SMC ABC demonstrated greater robustness. They were more effective at converging to "pseudo-true" parameters even when the model was slightly misspecified. In the tumor growth example, even robust neural methods (RSNLE) struggled to produce accurate predictive intervals for noisy real-world data, whereas SMC ABC performed reliably across all cases.
The Importance of Summary Statistics
In the cell invasion model, the authors found that certain parameters, such as cell movement rates, were non-identifiable when using only cell density data. This highlights that the choice of summary statistics is just as critical as the choice of the inference algorithm itself.
Conclusion
While neural SBI offers impressive speed, traditional statistical methods remain the "gold standard" for robustness in the face of messy biological data. The authors conclude that model misspecification is the norm in biology, making rigorous pre-analysis and uncertainty quantification mandatory. Future work should focus on making neural SBI more resilient to out-of-distribution data by incorporating statistical checks into the neural training process.
Summary updated