(351b) Integrating Multiple Molecular Model Types With Stochastic Modeling

Molecular modeling has become a sprawling field, with models spanning
many length and time-scales, being phenomenological or physics-based,
and being biased or unbiased. In machine-learning, it has been
observed that integrating multiple models often leads to results
better than any one model. In molecular modeling, there is little
progress on this topic despite the plethora of model types and
simulation techniques. In this presentation, I'll discuss how to
integrate multiple models and data types within the framework of
graphical models, a type of stochastic modeling. Through the use of
graphical models, we've developed integrative models which integrate
QSAR models, sequence motif models, and simulation results. These
models are capable of excellent antimicrobial peptide activity
prediction while retaining model parameters that are easy to
interpret. An example graphical model is shown in Figure 1. These
integrative models are dynamic Bayesian networks and can encode
multiple data types and constraints. Despite their complexity, dynamic
Bayesian networks may be quickly built and trained due to the
generality of graphical model algorithms. I'll also discuss
preliminary results on combining simulation with experiments and other
simulations.

Figure 1: A graphical model which classifies sequences based on their
motifs and amino acid distributions. The classifier was trained on
antimicrobial peptides and performed with over 90% accuracy on
predicting the activity of unobserved peptides.