Systematic Bayesian posterior analysis guided by Kullback-Leibler divergence facilitates hypothesis formation.

Area of Science:

• Computational Biology
• Biostatistics
• Systems Biology

Background:

• Bayesian inference generates posterior distributions for model parameters, crucial for hypothesis formation and refinement.
• Existing methods for identifying alternative model hypotheses from posteriors are often qualitative and unsystematic.
• Previous approaches typically halt at hypothesis formation, leaving further investigation incomplete.

Purpose of the Study:

• To introduce a quantitative, systematic framework for accelerating Bayesian hypothesis formation.
• To develop a method for investigating generated hypotheses and deriving novel biological insights.
• To utilize Kullback-Leibler (KL) divergence for ranking model parameters based on information gain from experimental data.

Main Methods:

• Implemented KL divergence to rank model parameters by their information gain from experimental data.
• Prioritized examination of parameters with the highest information gain for evidence of alternative hypotheses.
• Applied the KL divergence ranking approach to two distinct examples, including a computational model of prolactin-induced JAK2-STAT5 signaling.

Main Results:

• The KL divergence ranking successfully identified parameters associated with alternative hypotheses in both established and novel applications.
• In the JAK2-STAT5 signaling model, a bimodal posterior was discovered within the top-ranked parameters for the prolactin receptor degradation rate.
• The approach facilitated model refinement and the determination of the most plausible degradation rate, yielding biologically significant insights.

Conclusions:

• The KL divergence-based ranking provides a novel and generalizable quantitative framework for Bayesian hypothesis formation.
• This method systematically uncovers alternative model hypotheses and facilitates deeper investigation, leading to new biological discoveries.
• The approach's effectiveness is contingent on appropriate prior formulation and posterior distribution characteristics.