(W0930-06-39) Can Molecular Descriptors Drive Nanosuspension Innovation? Leveraging Machine Learning and AI to Predict the Particle Size

Purpose: From the early product development stage, the major focus is on understanding the physicochemical properties of newly discovered molecules. Translating formulation techniques to different drugs is exhaustive and time-consuming process. The dynamic era of AI/ML begins with correlating the drug properties to critical formulation parameters. Nanosuspension has been a go-to approach post salt screening for enhancing solubility, dissolution rates, resulting in improved bioavailability, and better absorption. Several commercial nanosuspension products demonstrate their clinical importance: orally administered Rapamune^®(sirolimus) and Emend^®(aprepitant) showing improved bioavailability; injectable Invega Sustenna^®(paliperidone palmitate) providing extended-release properties; and topical products enabling enhanced skin penetration of poorly soluble actives such as antifungals and corticosteroids. Despite commercial success, nanosuspension development remains empirical, with scientists relying on resource-heavy trial-and-error. AI can transform formulation science by enabling predictive models that reduce experimentation. However, effective AI integration requires comprehensive datasets capturing relationships between molecular properties and formulation performance. Our research addresses this gap by generating structured datasets linking descriptors to outcomes across delivery platforms. Previously, we used ML to predict formulation strategies for poorly soluble drugs using marketed products, but these models were limited by small datasets and couldn’t explain formulation choices, as real-world decisions involve factors like manufacturing, stability, market needs, and patient use. This research represents a prototype, focusing on creating a systematic experimental dataset specifically designed to reveal the molecular drivers of nanosuspension behavior. Moreover, we have developed a unique design of five distinct parameters (NanoStabilityIndex, SizePredictionFactor, RedispersibilityIndex, HLB_Estimated, and PiStackingPotential) specifically designed to capture nanosuspension-relevant properties that traditional descriptors might miss. This is the first step in our broader initiative to develop predictive models across multiple formulation platforms, including amorphous solid dispersions, lipid-based delivery systems, and cyclodextrin complexation, ultimately creating a comprehensive framework to guide pharmaceutical formulation scientists through evidence-based decision-making.

Methods: Nanosuspensions were prepared using 40 poorly soluble drugs with diverse physicochemical properties, such as various log P ranges, molecular weights, routes of administration. Each was formulated into three systems: F1 (HPMC-SLS), F2 (Kollidon K17-Tween 80), and F3 (Albumin-Tween 80). Particle size distributions were measured, and D50 values were categorized as Low, Medium, High, or PPT (precipitation/aggregation). In model development, 185 molecular descriptors (RDKit/PaDEL) were found to be effective in prediction, and in addition, 5 custom descriptors were calculated-NanoStabilityIndex(combining charge characteristics, hydrogen bonding, hydrophobicity, and size), SizePredictionFactor(based on molecular weight, Log P, polar surface area, and flexibility), RedispersibilityIndex(polar properties and hydrogen bonding), HLB_Estimated (modified hydrophilic-lipophilic balance), and PiStackingPotential(aromatic interactions). Preprocessing involved removing zero-variance features, handling missing values through median imputation, and eliminating highly correlated features. After preprocessing, 113 molecular descriptors remained. Three separate machine learning models were developed to predict particle size categories for each formulation (D50_1, D50_2, and D50_3). Various algorithms, such as Random Forest, Gradient Boosting, SVM, XGBoost, and bagging classifiers, were used. Feature selection was performed using recursive feature elimination, model-based selection, and correlation-based methods. The models were evaluated using stratified cross-validation with F1 score as the primary performance metric. Additional validation with new compounds is planned to further enhance model robustness.

Results: The performance analysis revealed that Gradient Boosting performed best for F1 formulation prediction (F1 score: 0.5556), SVM for F2 formulation (F1 score: 0.6000), and Random Forest for F3 formulation (F1 score: 0.5714). Feature importance analysis identified distinct molecular descriptors influencing particle size across different formulations. For F1 (HPMC-SLS), BalabanJ (topological index), EstimatedLogS (solubility parameter), and SMR_VSA7 (molecular surface properties) were most significant. F2 (Kollidon K17-Tween 80) was predominantly influenced by PaDEL_VPC-6 (van der Waals surface properties), EState_VSA7 (electrotopological state), and NPR1 (normalized principal moments ratio). For F3 (Albumin-Tween 80), our custom NanoStabilityIndex proved highly predictive, along with PEOE_VSA6 (partial charge descriptors) and AcidGroupCount. Notably, only Crippen_LogP was common across all three formulations, confirming lipophilicity as universally important for nanosuspension behavior. Our custom descriptors, particularly NanoStabilityIndex and SizePredictionFactor, contributed significantly to prediction accuracy, validating their design. The distribution of particle size categories varied across formulations, with F3 showing the highest prevalence of precipitation (42.1%), suggesting challenges with albumin-based formulations for many compounds.

Conclusion: This study demonstrates that molecular descriptors- quantifiable chemical properties of drugs can effectively guide nanosuspension development when paired with machine learning. By analyzing three distinct excipient systems, we identified unique "fingerprints" (e.g., NanoStabilityIndex) that predict final particle size. These models eliminate guesswork in excipient selection and accelerate the design of stable nanosuspensions for poorly soluble drugs. By translating complex molecular data into actionable insights, this AI-driven approach offers a scalable, efficient pathway to innovate and optimize drug delivery systems, proving that predictive analytics, rooted in molecular science, can transform traditional trial-and-error pharmaceutical development.


Figure 1: Comprehensive framework for nanosuspension particle size prediction. The top panel illustrates the methodological workflow from data collection through evaluation, including molecular descriptor calculation and model selection. The bottom panel compares model performance for three formulation systems, showing significant F1 score improvements after hyperparameter fine-tuning: Gradient Boosting (25.3% improvement for HPMC-SLS), Support Vector Machine (70.1% improvement for Kollidon-Tween), and Random Forest (16.9% improvement for Albumin-Tween). Each model predicts particle size categories (Low, Medium, High, PPT) for its respective formulation system.


Figure 2: Learning Curves Across Multiple Machine Learning Algorithms for Nanosuspension Particle Size Prediction
Learning curves showing training (red) and cross-validation (green) F1 scores for different models before hyperparameter optimization. (A) D50_1 formulation (HPMC-SLS) models, where Gradient Boosting was ultimately selected for fine-tuning. (B) D50_2 formulation (Kollidon K17-Tween 80) models, where SVM was selected for final implementation. (C) D50_3 formulation (Albumin-Tween 80) models, with Random Forest chosen for optimization. The gap between training and validation scores indicates the challenge of generalization with limited training data


Figure 3: Particle size distribution histograms for three nanosuspension formulation systems. The panels show the frequency distributions of particle size (PS, top row), polydispersity index (PDI, middle row), and median particle diameter (D50, bottom row) for HPMC-SLS (left column), Kollidon K17-Tween 80(middle column), and Albumin-Tween 80(right column) formulations. Red and green dashed vertical lines represent Low/Medium and Medium/High classification thresholds, respectively. The distributions illustrate the varying particle size, PDI, and D50 across formulation systems, providing the basis for categorizing nanosuspensions into Low, Medium, High, and precipitation classes for predictive modeling.