Augmenting Human\nExpertise in Weighted Ensemble Simulations\nthrough Deep Learning-Based Information Bottleneck

The weighted ensemble (WE) method stands out as a widely\nused segment-based\nsampling technique renowned for its rigorous treatment of kinetics.\nThe WE framework typically involves initially mapping the configuration\nspace onto a low-dimensional collective variable (CV) space and then\npartitioning it into bins. The efficacy of WE simulations heavily\ndepends on the selection of CVs and binning schemes. The recently\nproposed state predictive information bottleneck (SPIB) method has\nemerged as a promising tool for automatically constructing CVs from\ndata and guiding enhanced sampling through an iterative manner. In\nthis work, we advance this data-driven pipeline by incorporating prior\nexpert knowledge. Our hybrid approach combines SPIB-learned CVs to\nenhance sampling in explored regions with expert-based CVs to guide\nexploration in regions of interest, synergizing the strengths of both\nmethods. Through benchmarking on alanine dipeptide and chignolin systems,\nwe demonstrate that our hybrid approach effectively guides WE simulations\nto sample states of interest and reduces run-to-run variances. Moreover,\nour integration of the SPIB model also enhances the analysis and interpretation\nof WE simulation data by effectively identifying metastable states\nand pathways and offering direct visualization of dynamics.