Utils

RData Conversion

bioneuralnet.utils.rdata_to_df() converts an RData file to a CSV and loads it into a pandas DataFrame.

Logging

bioneuralnet.utils.get_logger() configures and returns a logger writing to bioneuralnet.log at the project root.

Graph Generation

This section details utility functions for generating networks from omics data matrices. All methods are industry-standard and open in the literature.

Methods and Examples

k-NN Cosine/RBF Similarity Graph

Computes either cosine similarity:

\[S_{ij} = \frac{x_i^\top x_j}{\|x_i\|\,\|x_j\|}\]

or Gaussian RBF kernel:

\[S_{ij} = \exp\bigl(-\|x_i - x_j\|^2 /(2\sigma^2)\bigr)\]

Sparsify by keeping top-(k) neighbors per node, optionally mutual.
```
  from bioneuralnet.utils.graphs import gen_similarity_graph
  A = gen_similarity_graph(X, k=15, metric='cosine', mutual=True)

Reference: Hastie et al., 2009 [Hastie2009_]
```

Pearson/Spearman Co-expression Graph

Compute correlation:

\[C_{ij} = \mathrm{corr}(x_i, x_j)\]

Then sparsify by top-(k) or threshold.

  from bioneuralnet.utils.graphs import gen_correlation_graph
  A = gen_correlation_graph(X, k=15, method='pearson')

Reference: Langfelder & Horvath, 2008 [Langfelder2008_]

Soft-Threshold Graph

Soft-threshold absolute correlation, similar to WGCNA:

\[W_{ij} = |C_{ij}|^\beta\]

Then top-(k) selection and normalization.

  from bioneuralnet.utils.graphs import gen_threshold_graph
  A = gen_threshold_graph(X, b=6.0, k=15)

Reference: Langfelder & Horvath, 2008 [Langfelder2008_]

Gaussian k-NN Graph

Gaussian kernel sparsified by k-nearest neighbors:

\[S_{ij} = \exp\bigl(-\|x_i - x_j\|^2 /(2\sigma^2)\bigr),\quad W = \text{Top}_k(S)\]
```
from bioneuralnet.utils.graphs import gen_gaussian_knn_graph
A = gen_gaussian_knn_graph(X, k=15, sigma=None)
```
Credit: adapts common practice from spectral clustering (Ng et al., 2002).

Mutual Information Graph

Estimate pairwise mutual information:

\[\mathrm{MI}_{ij} = I(x_i; x_j)\]

  from bioneuralnet.utils.graphs import gen_mutual_info_graph
  A = gen_mutual_info_graph(X, k=15)

Reference: Margolin et al., 2006 [Margolin2006_]

Preprocessing Utilities

A collection of data-cleaning and feature-selection functions for clinical and omics datasets.

Clinical Preprocessing

bioneuralnet.utils.preprocess.preprocess_clinical() Splits numeric and categorical features; replaces Inf/NaN; optionally scales numeric data (RobustScaler); encodes categoricals; drops zero-variance; and selects top-k features by RandomForest importance.

Example:
```
from bioneuralnet.utils.preprocess import preprocess_clinical
df_top = preprocess_clinical(X, y, top_k=10, scale=True)
```
bioneuralnet.utils.preprocess.clean_inf_nan() Replaces Inf with NaN, imputes medians, drops zero-variance columns, and logs counts.

Example:
from bioneuralnet.utils.preprocess import clean_inf_nan df_clean = clean_inf_nan(df)

Variance-Based Selection

bioneuralnet.utils.preprocess.select_top_k_variance() Cleans data, then picks the top-k numeric features by variance.

Example:
from bioneuralnet.utils.preprocess import select_top_k_variance df_var = select_top_k_variance(df, k=500)

Correlation-Based Selection

bioneuralnet.utils.preprocess.select_top_k_correlation() - Supervised: if you pass y, selects features by absolute Pearson correlation with y. - Unsupervised: if y=None, picks features with the lowest average inter-feature correlation (redundancy reduction).

Example:
from bioneuralnet.utils.preprocess import select_top_k_correlation df_sup = select_top_k_correlation(X, y, top_k=100) # supervised df_unsup = select_top_k_correlation(X, top_k=100) # unsupervised

RandomForest Feature Importance

bioneuralnet.utils.preprocess.select_top_randomforest() Requires numeric inputs; cleans data; drops zero-variance; fits RandomForest; and returns the top-k features by importance.

Example:
from bioneuralnet.utils.preprocess import select_top_randomforest df_rf = select_top_randomforest(X, y, top_k=200)

ANOVA F-Test Selection

bioneuralnet.utils.preprocess.top_anova_f_features() Runs ANOVA F-test (classification or regression), applies FDR correction, selects all significant features, and pads with next-best to reach max_features.

Example:
from bioneuralnet.utils.preprocess import top_anova_f_features df_anova = top_anova_f_features(X, y, max_features=100, alpha=0.05)

Network Pruning

bioneuralnet.utils.preprocess.prune_network() Prunes edges below a weight threshold, removes isolates, and logs before/after stats.

Example:
```
from bioneuralnet.utils.preprocess import prune_network
pruned = prune_network(adj_df, weight_threshold=0.1)
```
bioneuralnet.utils.preprocess.prune_network_by_quantile() Uses a quantile cutoff on edge weights, prunes accordingly, removes isolates, and logs stats.

Example:
from bioneuralnet.utils.preprocess import prune_network_by_quantile pruned_q = prune_network_by_quantile(adj_df, quantile=0.75)
bioneuralnet.utils.preprocess.network_remove_low_variance() Drops rows/columns in the adjacency matrix whose variance falls below a threshold.

Example:
from bioneuralnet.utils.preprocess import network_remove_low_variance filtered = network_remove_low_variance(adj_df, threshold=1e-5)
bioneuralnet.utils.preprocess.network_remove_high_zero_fraction() Removes rows/columns where the fraction of zero weights exceeds a threshold (default 0.95).

Example:
from bioneuralnet.utils.preprocess import network_remove_high_zero_fraction filtered_z = network_remove_high_zero_fraction(adj_df, threshold=0.95)

Data Summary Utilities

bioneuralnet.utils.data_summary.variance_summary() - Computes summary statistics for column variances.

Example:

from bioneuralnet.utils.data_summary import variance_summary
stats = variance_summary(df, low_var_threshold=1e-4)

bioneuralnet.utils.data_summary.zero_fraction_summary() - Computes statistics for the fraction of zeros per column.