fuzzygene_config(
maxRegulators = 4,
selection = "prior-first",
minCorrelation = 0.3,
mode = "auto",
mfCount = 3,
mfKind = "gaussian",
maxEpoch = 100,
learningRate = 0.02,
pruneThreshold = 0.01,
maxGenes = 2000,
maxTrainRows = 0,
maxCandidateGenes = 0,
minBestCorrelation = 0,
noveltyQuota = 0,
rulePartition = "auto",
scatterRules = 48,
scatterInputs = 8,
scatterMinInputs = 2,
scatterZThreshold = 1,
scatterSeed = 0,
maxRules = 128,
maxConsequentParams = 512,
standardize = TRUE,
propagationSteps = 50,
relaxation = 0.5,
knockdownResidual = 0.2,
overexpressionSigmas = 3,
significanceZScore = 2,
degreeOfParallelism = 0,
minTimePoints = 3);
maxRegulators
how many candidate regulators (ANFIS inputs) are kept per target gene. The rule count of a grid partition is m^n, so this is the primary guard against the rule explosion; 3-6 is the practically useful range. Default 4. [as integer]
selection
regulator selection strategy: prior-first (default, prior edges by confidence then correlation), correlation (data driven only) or prior (prior edges only). [as string]
minCorrelation
minimum |Pearson| for a gene to be taken as a correlation based regulator. Default 0.3. [as double]
mode
fit route: auto (default, lagged when the matrix has at least minTimePoints distinct time points), static or lagged. [as string]
mfCount
number of membership functions per input. Default 3. [as integer]
mfKind
membership function shape: gaussian, bell, sigmoid, triangular or trapezoidal. Default gaussian. [as string]
maxEpoch
maximum training epochs per target gene. Default 100. [as integer]
learningRate
learning rate of the premise parameter optimizer. Default 0.02. [as double]
pruneThreshold
rules whose average normalized firing strength falls below this value are pruned after training; 0 disables pruning. Default 0.01. [as double]
maxGenes
upper bound on the number of modelled genes (highest variance first, prior TFs always kept); 0 keeps everything. Default 2000. [as integer]
maxTrainRows
maximum number of training rows per target gene; 0 uses every row. The dominant cost of the hybrid learning is the least squares normal equations, which scale as rows x P^2, so capping the rows is the cheapest large speedup: a few thousand rows are already far more than the few hundred consequent parameters need. Default 0 (all rows); 3000-5000 is the recommended setting for genome scale data. [as integer]
maxCandidateGenes
size limit of the candidate pool used by the correlation screening (prior TFs plus the highest variance genes); 0 screens every gene. The screening cost is O(genes^2 x rows), which at genome scale becomes comparable with the training itself. Default 0; 2000-4000 is the recommended setting for genome scale data. [as integer]
minBestCorrelation
skip a target gene when it has no prior regulator and its best |Pearson| against the candidates is below this value (such models have R2 close to 0). 0 disables the filter. [as double]
noveltyQuota
fraction (0-1) of the regulator slots reserved for genes that are not in the prior network of that target, i.e. for links discovered from the expression data. 0 (default) keeps the old "prior first" behaviour, where prior TFs can occupy every slot and the model can only repeat the prior. 0.2-0.4 is a sensible range: with K = 6 and 0.34 two slots are always given to data driven regulators. Ignored when selection = "prior-only". [as double]
rulePartition
rule generation strategy: auto (default), grid or scatter. grid places m membership functions per input and enumerates m^K rules, so the rule count explodes with the input count. scatter clusters the training samples (kmeans) and creates one rule per cluster, each with its own gaussian center/width per input - the number of rules no longer depends on K, which is what makes maxRegulators = 18 feasible. auto keeps grid while m^K <= maxRules and m^K*(K+1) <= maxConsequentParams, otherwise it switches to scatter. [as string]
scatterRules
number of clusters (= rules) of the scatter partition. Default 48. [as integer]
scatterInputs
maximum number of antecedents per scatter rule. Default 8. [as integer]
scatterMinInputs
minimum number of antecedents per scatter rule. Default 2. [as integer]
scatterZThreshold
an input joins a cluster rule when the cluster center deviates from the global mean by at least this many standard deviations. Default 1. [as double]
scatterSeed
random seed of the clustering, keeps the structure reproducible. Default 0. [as integer]
maxRules
rule budget per target gene. Default 128. [as integer]
maxConsequentParams
consequent parameter budget per target gene. Default 512. [as integer]
standardize
z-score the expression matrix before fitting. Default TRUE. [as boolean]
propagationSteps
maximum fixed point iterations of a perturbation. Default 50. [as integer]
relaxation
damping factor of the fixed point iteration. Default 0.5. [as double]
knockdownResidual
fraction of the baseline expression retained after a virtual knockdown. Default 0.2. [as double]
overexpressionSigmas
overexpression level in training standard deviations above the baseline. Default 3. [as double]
significanceZScore
|z| threshold used to flag a gene as significantly changed. Default 2. [as double]
degreeOfParallelism
per-gene training parallelism; 0 = all cores, 1 = serial. [as integer]
minTimePoints
distinct time points required to switch to the lagged route. Default 3. [as integer]
env
the R# runtime environment (unused, kept for error reporting symmetry). [as Environment]