How to use vc_regression

vcregression provides a suite of tools for Bayesian statistical inference of sequence-function relationships using Gaussian process(GP) regression. The most important idea behind vcregression is to infer a prior corresponding to the proportions of variance due to each order of genetic interaction (the variance components) based on the empirical autocorrelation function observed in the data. The variance components provide a covariance structure over the sequence space, which allows us make various Bayesian statistical inferences.

Unlike most regression methods used in modeling sequence-function relationships that only consider genetic interactions (epistasis) up to the 2nd and occasionally the 3rd order, vcregression is capable to model epistasis among arbitrary number of sites, which makes it a powerful method for predicting the phenotypes of unsampled sequences.

The 4 major functionalities of vc_regression include:

  • estimation of the variance components from partially observed fitness landscapes
  • calculating the maximum a posterior estimate (MAP) using GP regression
  • calculating the posterior variance
  • posterior sampling using Hamiltonian Monte Carlo

Each functionality listed above correspond to a separate command line interface. For details, see Implementation.

Before using vcregressoin, make sure the input datafile is in the correct format. See Data preparation for guidelines.

For detailed implementations of vcregression, see Tutorial for an application to a high throughput dataset of human 5’ splice sites.

Notes on the upper limit of the size sequence space

vcregression leverages the isotropic property of the covariance matrix defined by the variance components. All fast calculations of vcregression relies on representing the covariance matrix using the Laplacian \(\mathbf{L}\) of the associated Hamming graph. As a result, most of the computational and memory cost comes from constructing and storing \(\mathbf{L}\), which is defined for the whole sequence space. So currently the largest sequence space vcregression can accomodate is on the order of several million. Therefore, before using vcregression on your own dataset, first calculate the size of the sequence space:

\[N = a^l,\]

and make sure \(N\) is no larger than several million.

If the sequence space associated with your dataset is much larger, a method using the dense covariance matrix is possible (but with computational complexity \(\mathcal{O}(n^3)\), if you have \(n\) data points). However, this functionality has not been implemented in vcregression. If you are interested in this alternative, please contact Juannan Zhou (Email: jzhou@cshl.edu).