Add interactions to a trajectory model — add

This function adds significant interactions to a given trajectory model if they do not already exist. It identifies significant interactions based on the provided threshold and updates the model features with logistic features derived from these interactions. The trajectory model is then re-learned with the new features.

Usage

add_interactions(
  traj_model,
  interaction_threshold = 0.001,
  max_motif_n = NULL,
  max_add_n = NULL,
  max_n = NULL,
  lambda = 1e-05,
  alpha = 1,
  seed = 60427,
  interactions = NULL,
  ignore_feats = c("TT", "CT", "GT", "AT", "TC", "CC", "GC", "AC", "TG", "CG", "GG",
    "AG", "TA", "CA", "GA", "AA"),
  force = FALSE,
  logist_interactions = FALSE,
  use_cv = FALSE,
  nfolds = 10,
  family = "binomial",
  rescale_pred = FALSE,
  only_sig_motifs = FALSE,
  only_sig_add_motifs = TRUE,
  interaction_scale_factor = 1,
  min_signal_correlation = NULL
)

Arguments

traj_model: The trajectory model object.
interaction_threshold: threshold for the selecting features to create interactions. IQ learns a linear model on the features and selects the features with coefficients above this threshold. Default: 0.001
max_motif_n: maximum number of motifs to consider for interactions. If NULL, all motifs above the interaction_threshold will be considered. Default: NULL
max_add_n: maximum number of additional features to consider for interactions. If NULL, all additional features above the interaction_threshold will be considered. Default: NULL
max_n: maximum number of interactions to consider. If NULL, all interactions will be considered. If set, the interactions will be selected based on correlation with the signal in the training data. Default: (motif models + additional features) * 10
lambda: The lambda value to use for relearning. If NULL, the lambda value from the trajectory model is used.
alpha: The elastic net mixing parameter for glmnet. alpha=1 is the lasso penalty, alpha=0 is the ridge penalty. Default: 1
seed: random seed for reproducibility.
interactions: A precomputed interaction matrix. If provided, the function will not compute the interactions. Default: NULL
ignore_feats: A character vector of features to ignore when creating interactions. Default: dinucleotides
force: If TRUE, the function will add interactions even if they already exist. Default: FALSE
logist_interactions: Logical indicating whether to transform interactions to logistic functions. Default: FALSE
use_cv: Logical indicating whether to use cross-validation for lambda selection. Default is FALSE.
nfolds: Number of folds for cross-validation. Default is 10.
family: The family to use for the glmnet model. Either "binomial" (default) or "gaussian" for linear regression.
rescale_pred: Logical indicating whether to rescale the predicted values. Default is TRUE.
only_sig_motifs: Logical indicating whether to only consider significant motifs for interactions. Default: FALSE
interaction_scale_factor: A multiplier applied to the normalized interaction matrix before it is joined into the model features. Default: 1 (no scaling). Values >1 make interactions more prominent relative to motif features; values <1 down-weight them.
min_signal_correlation: If non-NULL, after the max_n correlation-based top-N cap, drop any interaction column whose absolute correlation with the training diff_score is below min_signal_correlation * max(|cor|). For example, min_signal_correlation = 1/8 keeps only interactions whose |cor| exceeds 1/8 of the best interaction's |cor|, mirroring Akhiad's manual post-hoc filter. Default: NULL (no filter).

Value

The updated trajectory model with added interactions.