GeneLens 0.1.3

Differential gene expression analysis via Monte-Carlo Machine Learning Inference and Network Analysis

GeneLens: Integrated DEG Analysis & Biomarker Prediction

Overview

GeneLens is a Python package for functional analysis of differentially expressed genes (DEGs) and biomarker prediction, integrating:

• Machine learning-based biomarker identification
• Graph-based prediction of gene function via protein-protein interaction networks analysis

Key applications:

• Identification of biomarkers
• Analysis of gene-gene networks

Features

Core Modules

1. FSelector

   • Machine learning pipeline for biomarker discovery
   • Features:
     • Automatic Monte Carlo simulation of stable models
     • Automated model training/tuning
     • Feature importance analysis
     • Customizable thresholds

2. NetAnalyzer

   • Implements graph-based algorithm (Osmak et al. 2020, 2021)
   • Predicts genes functions via topological analysis of molecular networks
   • Features:
     • Automated network construction
     • Pathway enrichment
     • Integration with Feature importance from FSelector

Additional Capabilities

• Standardized analysis pipelines
• Interactive network visualizations
• Support for multi-omics data integration

Installation

pip install genelens

Example of use

from genelens.fselector import FeatureSelector, get_feature_space, fsplot
from genelens import netanalyzer, enrichment
from matplotlib import pyplot as plt
import numpy as np
import pandas as pd
import networkx as nx
from importlib.resources import files

# data load
data = pd.read_csv(files("genelens").joinpath("data/exampl_data/train_test.csv"), index_col=0)

X = data.drop('index', axis=1)
y = list(map(int, data['index'] == 'HCM'))

print(X.shape)

(145, 14830)

# FeatureSelector initialization
FS_model = FeatureSelector(X, y,
                           C = None, 
                           C_space=np.linspace(0.0001, 1, 20), #bigger space -> more precision, more processor time
                           C_finder_iter=10,
                           cut_off_frac_estimation=True,
                           cut_off_frac_model=0,
                           cut_off_estim_params={'max_feature': 50}) # This parameter implements early stopping. Bigger feature space -> more precision, more processor time

The regularization coefficient was not specified, the search for the optimal C was started


processing: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 10/10 [02:20<00:00, 14.07s/it]


Optimal regularization coefficient (С) =  0.053

Prefit model for cutoff weight level estimation


fit model: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3000/3000 [12:38<00:00,  3.95it/s]


Prefit done
Serching cutoff level for feature weights... 0 

feature space analysis:  13%|██████████████████▌                                                                                                                               | 27/213 [00:07<00:51,  3.59it/s]


1 

feature space analysis:  13%|██████████████████▌                                                                                                                               | 27/213 [00:07<00:51,  3.60it/s]


2 

feature space analysis:  13%|██████████████████▌                                                                                                                               | 27/213 [00:07<00:51,  3.59it/s]


3 

feature space analysis:  13%|██████████████████▌                                                                                                                               | 27/213 [00:07<00:51,  3.59it/s]


4 

feature space analysis:  13%|██████████████████▌                                                                                                                               | 27/213 [00:07<00:51,  3.61it/s]


5 

feature space analysis:  13%|██████████████████▌                                                                                                                               | 27/213 [00:07<00:52,  3.57it/s]


6 

feature space analysis:  13%|██████████████████▌                                                                                                                               | 27/213 [00:07<00:51,  3.60it/s]


7 

feature space analysis:  13%|██████████████████▌                                                                                                                               | 27/213 [00:07<00:52,  3.57it/s]


8 

feature space analysis:  13%|██████████████████▌                                                                                                                               | 27/213 [00:07<00:51,  3.61it/s]


9 

feature space analysis:  13%|██████████████████▌                                                                                                                               | 27/213 [00:07<00:52,  3.56it/s]

optimal cut of weight level =  0.72

FS_model.fit(max_iter=2700, log=True, feature_resample=0) #more max_iter -> more precision, more processor time

fit model: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2700/2700 [11:20<00:00,  3.97it/s]

fsplots = fsplot(FS_model)
fsplots.plot_all(fontsize=25, labels=['a.', 'b.', 'c.', 'd.', 'e.', 'f.'], 
                left=0.1, right=0.9, top=0.9, bottom=0.1, hspace=0.5, wspace=0.5)
plt.show()

print(get_feature_space([FS_model], cut_off_level=0.75))

{'MYH6', 'RASD1'}

FS_model.best_features

{'RASD1': np.float64(0.9510623822037754),
 'MYH6': np.float64(0.8420449794132905)}

Network Enrichment Analysis

GenGenNetwork = netanalyzer.MainNet() #Load String db and create gene-gene interaction network
GenGenNetwork.get_LCC() #get the largest connected component from the network

LCC was extracted
Total connected components=146, LCC cardinality=9844

GenGenNetwork.minimum_connected_subgraph(FS_model.best_features)

RASD1 absent from LCC, excluded from further analysis
CDC42EP4 absent from LCC, excluded from further analysis

mst-graph was extracted
Initial core feature=1, mst-graph cardinality=0

Two of the three selected genes are missing from the version of the String database we are using. Therefore, it is not possible to construct an mst-graph. To continue the analysis, we will select the top 10 genes sorted by their Score

GenGenNetwork.minimum_connected_subgraph(dict(list(FS_model.all_features.items())[:10]))

RASD1 absent from LCC, excluded from further analysis
CDC42EP4 absent from LCC, excluded from further analysis
ZFP36 absent from LCC, excluded from further analysis

mst-graph was extracted
Initial core feature=7, mst-graph cardinality=17

pos = nx.circular_layout(GenGenNetwork.mst_subgraph)

nx.draw(
    GenGenNetwork.mst_subgraph,
    pos,
    with_labels=True,       
    node_color='skyblue',   
    edge_color='gray',      
    node_size=2000,         
    font_size=15            
)

# Показываем граф
plt.show()

enrich_res = enrichment.reactome_enrichment(list(GenGenNetwork.mst_subgraph.nodes()), species='Homo sapiens')
enrich_res = enrichment.reac_pars(enrich_res)
G_enrich = enrichment.get_net(enrich_res) #граф сигнальных путей

reactome_df, raw_res = enrichment.dendro_reactome_to_pandas(enrich_res, G_enrich)

enrichment.get_dendro(reactome_df, FS_model.all_features)

<Figure size 2400x2400 with 0 Axes>

The color gradient from gray to red in the signatures reflects the increase in the weight of genes according to their calculated Score. The redder the signature, the higher the weight.

More information can be found in our publications:

1. Pisklova, M., Osmak, G. (2024). Unveiling MiRNA-124 as a biomarker in hypertrophic cardiomyopathy: An innovative approach using machine learning and intelligent data analysis. International Journal of Cardiology, 410, 132220.
2. Osmak, G., Baulina, N., Kiselev, I., & Favorova, O. (2021). MiRNA-regulated pathways for hypertrophic cardiomyopathy: network-based approach to insight into pathogenesis. Genes, 12(12), 2016.
3. Osmak, G., Kiselev, I., Baulina, N., & Favorova, O. (2020). From miRNA target gene network to miRNA function: miR-375 might regulate apoptosis and actin dynamics in the heart muscle via Rho-GTPases-dependent pathways. International Journal of Molecular Sciences, 21(24), 9670.
4. Osmak, G. J., Pisklova, M.V. (2025). Transcriptomics and the “Curse of Dimensionality”: Monte Carlo Simulations of ML-Models as a Tool for Analyzing Multidimensional Data in Tasks of Searching Markers of Biological Processes. Molecular Biology, 59, 143-149.