GenSelTox: Genome-Informed AI-Driven Discovery and Experimental Validation of Inorganic Materials with Selective Antibacterial Action
The rational design of nanoparticles (NPs) with selective antimicrobial activity remains a major challenge in nanomedicine, particularly for combating antimicrobial resistance while preserving beneficial microbiota. In this work, we present an AI-driven discovery platform that integrates predictive modeling and generative design to prioritize selectively antibacterial inorganic NPs, which are subsequently validated experimentally. The platform combines a curated dataset of 2098 antibacterial activity measurements with enriched descriptors, including physicochemical properties, synthesis parameters, and Kyoto Encyclopedia of Genes and Genomes (KEGG) derived functional orthologs and metabolic pathways. Gradient boosting models trained on this gene informed dataset achieved high predictive performance, especially XGB regression acheived cross-validated R² score of 0.79, 0.83 and 0.80 for predicting minimal concentration (MC) with experimental parameters, functional ortholog features and metabolic pathways. Next, GenSelTox couples these models with a generative framework, integrating genetic algorithms, and reinforcement learning to explore NP formulation space for optimal selective toxicity. As proof-of-concept, one of top-ranked ZnO NP was synthesized and tested against pathogenic (Staphylococcus aureus, Pseudomonas aeruginosa) and non-pathogenic (Bacillus subtilis, Escherichia coli) strains. Experimental results confirmed species-selective inhibition consistent with model predictions, and interpretable features revealed key genomic drivers of susceptibility, including oxidative stress and DNA repair pathways. By uniting AI with mechanistic insight and empirical validation, GenSelTox advances the field toward precision nanotherapeutics. The platform offers a scalable, interpretable, and generalizable framework for selective nanoparticle design, with broad potential for adaptation to antifungal, antiviral, and anticancer applications.
The platform requires Python 3.8 or higher. We recommend using a virtual environment to manage dependencies. Begin by cloning the repository and navigating to the project directory:
git clone https://github.com/yourusername/GenSelTox.git
cd GenSelToxCreate and activate a virtual environment:
python -m venv venv
source venv/bin/activate # On Windows: venv\Scripts\activateInstall all required dependencies using the provided requirements file:
pip install -r requirements.txtThe platform operates on three primary dataset configurations. The base dataset MIC_df_preprocessed.csv contains core antibacterial activity measurements including nanoparticle properties such as size distribution (minimum, average, maximum), morphological characteristics (shape), surface modifications (coating), and synthesis parameters (method, temperature, duration, solvent). This dataset also includes experimental conditions, bacterial strain identifiers, and corresponding MIC values measured in µg/mL.
Two enriched versions extend this foundation with genomic information. The dataset MIC_df_ko.csv incorporates 11,273 binary features representing KEGG functional orthologs (K numbers), which encode the presence or absence of specific enzymatic functions and molecular processes in each bacterial strain. The dataset MIC_df_path.csv similarly enriches the base data with binary features corresponding to KEGG metabolic pathways, capturing higher-level functional organization of bacterial metabolism and cellular processes.
These genomic features are acquired through the KEGG REST API. Navigate to the data preprocessing directory and execute the provided Jupyter notebook:
cd "Data and Preprocessing"
jupyter notebook GET_KEGG_ko_path.ipynbThis notebook queries the KEGG database for bacterial genome annotations, retrieves functional ortholog assignments and pathway memberships for each strain in the dataset, and generates binary feature matrices that are subsequently merged with the experimental measurements. The resulting files kegg_ko_matrix.csv and kegg_pathway_matrix.csv serve as the genomic annotation layers for model training.
To prepare the raw experimental data for modeling, execute the preprocessing script located in the data processing directory. This can be accomplished by running:
cd "Data and Preprocessing"
python preprocessing.pyThe preprocessing pipeline performs several critical transformations on the raw measurements. Outlier removal is applied to MIC values (restricting the range to 0-5000 µg/mL), nanoparticle size parameters (filtering particles with average diameter below 200 nm and maximum diameter below 300 nm), and incubation conditions (limiting incubation periods to physiologically relevant ranges). Missing values in categorical features such as synthesis method and particle shape are imputed using modal values from the distribution, while numerical features including synthesis temperature and duration are imputed using mean values. Categorical features with insufficient representation (appearing in fewer than 0.5% of observations) are filtered to ensure statistical reliability. Finally, MIC values undergo logarithmic transformation to normalize their distribution and improve model performance.
After preprocessing the base experimental data, merge the genomic annotations by executing:
python merging.pyThis script integrates the KEGG-derived functional ortholog and pathway matrices with the cleaned experimental measurements, producing the three final datasets: MIC_df_preprocessed.csv containing only experimental features, MIC_df_ko.csv augmented with 11,273 functional ortholog features, and MIC_df_path.csv augmented with metabolic pathway annotations. These datasets represent different feature spaces for subsequent model training and evaluation.
Training predictive models on the prepared datasets can be performed using the standardized pipeline infrastructure. The following example demonstrates training on the functional ortholog-enriched dataset:
import polars as pl
from MLModelBuilding.BasePipeline import PipelineFIPI
from xgboost import XGBRegressor
from catboost import CatBoostRegressor
from lightgbm import LGBMRegressor
from sklearn.ensemble import RandomForestRegressor
df = pl.read_csv("MIC_df_ko.csv")
models = [
XGBRegressor(random_state=42),
CatBoostRegressor(random_state=42, verbose=False),
LGBMRegressor(random_state=42, verbose=0),
RandomForestRegressor(random_state=42)
]
pipeline = PipelineFIPI(
df=df,
dataset_name="MIC_KO_Dataset",
models=models,
cutoff_FI=0.95,
cutoff_PI=0.99
)
final_df = pipeline.run()This pipeline automatically handles data cleaning, dimensionality reduction through variance and correlation thresholds, ordinal encoding of categorical variables, standard scaling of numerical features, and train-test splitting with a fixed random seed for reproducibility. The PipelineFIPI class implements sequential feature selection using both feature importance and permutation importance. The cutoff_FI parameter specifies the cumulative importance threshold for retaining features based on model-intrinsic importance scores (0.95 retains features accounting for 95% of cumulative importance), while cutoff_PI applies the same logic to permutation-based importance estimates. Feature importance visualizations are automatically generated and saved during the pipeline execution.
For more granular control over the training process, the TrainablePipeline1 class enables custom model training, evaluation, and persistence:
from MLModelBuilding.TrainablePipeline import TrainablePipeline1
pipeline = TrainablePipeline1(
df=df,
dataset_name="MIC_Dataset",
models=[]
)
model = CatBoostRegressor(
iterations=1400,
learning_rate=0.093,
depth=4,
random_state=42,
verbose=False
)
r2, rmse = pipeline.fit(model)
pipeline.plot_top_features_gradient1(top_n=40, fname="feature_importance.png")
pipeline.save("trained_model.joblib")The fitted model along with its associated preprocessing transformations (encoders, scalers) is serialized to disk for subsequent use in generative design or prediction on new formulations.
Optimal hyperparameters for XGBoost models are identified through Bayesian optimization using Optuna. Navigate to the optimization directory and execute:
cd MLModelOptimization
python optimize_xgb.pyThis script performs 100 trials of hyperparameter search across a predefined space including tree architecture parameters (maximum depth from 3 to 8, minimum child weight from 1.0 to 10.0), learning dynamics (learning rate from 0.03 to 0.2 on a logarithmic scale, number of estimators from 300 to 1000), regularization terms (L1 alpha from 1e-4 to 0.3, L2 lambda from 0.5 to 3.0), and sampling strategies (subsample ratio from 0.6 to 0.9, column sampling from 0.5 to 0.8). Each trial is evaluated using 5-fold cross-validation with R² as the objective metric. The search employs tree-structured Parzen estimators to efficiently explore the hyperparameter landscape. Best parameters from multiple optimization runs are automatically saved to best_params_2.py with unique timestamps for version tracking.
Similarly, CatBoost hyperparameters are optimized through:
python optimize_catboost.pyThe CatBoost optimization space includes iteration count (300 to 2000), learning rate (0.01 to 0.3 logarithmically scaled), tree depth (4 to 10 levels), L2 regularization strength (1 to 10), random strength for split selection (0.5 to 3.0), and bootstrap strategy (Bayesian or Bernoulli). When Bayesian bootstrap is selected, the bagging temperature parameter is sampled from 0 to 1; when Bernoulli bootstrap is chosen, the subsample ratio is sampled from 0.6 to 1.0. Results are stored in best_params_cat.py with corresponding R² performance metrics. Multiple sequential runs are recommended to ensure convergence to global optima rather than local minima.
The platform implements two complementary approaches for discovering novel nanoparticle formulations with enhanced selective toxicity. The genetic algorithm framework evolves candidate formulations through iterative selection, crossover, and mutation operations guided by predicted antimicrobial selectivity.
To execute genetic algorithm-based discovery using the CatBoost predictor, navigate to the appropriate directory and run:
cd GeneticAlgorithm/Catboost
python ga_main.pyFor discovery using XGBoost predictions:
cd GeneticAlgorithm/XGBoost
python ga_main.pyPrior to execution, configure the target bacterial strains in ga_compd_generation.py by setting the pathogenic and beneficial bacteria identifiers. For example:
name_of_pathogenic_bacteria = 'Escherichia coli ATCC 25922'
name_of_good_bacteria = 'Pseudomonas aeruginosa nan'The genetic algorithm maintains a population of 100 candidate formulations per generation and evolves them over 100 generations using a mutation rate of 0.2 and crossover rate of 0.2. The crossover operator treats certain feature groups as linked units: nanoparticle size parameters (minimum, average, maximum) are transferred together to maintain realistic size distributions, synthesis-related parameters (method, precursor, capping agent) are co-inherited to preserve chemically feasible combinations, while independent features undergo individual crossover events with the specified probability. Fitness is evaluated as the difference in predicted log(MIC) between the beneficial and pathogenic strains, with higher values indicating greater selectivity favoring beneficial bacteria survival.
Output files are organized by generation in the output/ directory, with each generation's population saved as a CSV file containing all candidate parameters and their predicted antimicrobial activities against both target strains. Summary statistics tracking mean and maximum fitness across generations are saved for convergence analysis.
The reinforcement learning approach provides an alternative generative strategy using policy gradient methods. Execute the RL-based discovery through:
cd ReinforcementLearning
python rl_en.pyThe RL environment models nanoparticle design as a sequential decision process where the agent selects formulation parameters (particle size distribution, synthesis method, incubation conditions) to maximize the selectivity reward defined as the predicted MIC difference between target strains. The state space comprises all formulation parameters plus genomic features of the target bacteria, while the action space covers continuous ranges for size parameters and discrete choices for categorical variables like synthesis method. The default implementation uses Proximal Policy Optimization (PPO) with a learning rate of 1e-4, trained over 5120 timesteps with periodic checkpointing every 200 episodes. Alternative algorithms including Soft Actor-Critic (SAC) and Deep Deterministic Policy Gradient (DDPG) can be configured by modifying the model initialization in the script. Generated candidates and their predicted activities are continuously saved to result.csv for post-analysis and ranking.
Last Updated: March 2026
Version: 1.0.0