Sonal Solanki
Research Subjects
Sonal Solanki has a total of 45 co-authors. Here are the top co-authors by number of shared publications:
- Nikunj Solanki
Hon. Shri Babanrao Pachpute Vichardhara Trust’s Group of Institutions, Faculty of Pharmacy, Kashti, Ta- Shrigonda, Ahilyanagar, Maharastra, India, 414071
8 shared publications - Sreekanth Thota
Vaagdevi College of Pharmacy Autonomous, Ramnagar, Hanamkonda, Telangana, India, 506001
5 shared publications - Sreenath Sriram
Vaagdevi College of Pharmacy Autonomous, Ramnagar, Hanamkonda, Telangana, India, 506001
3 shared publications - Yamsani Shravan
Vaagdevi College of Pharmacy Autonomous, Ramnagar, Hanamkonda, Telangana, India, 506001
3 shared publications - Jyoti Maithani Kalra
School of Pharmaceutical Sciences, Shri Guru Ram Rai University, Dehradun, Uttarakhand, India
2 shared publications - Kapil Kalra
Alpine College of Management and Technology Dehradun, Uttarakhand, India
2 shared publications
Latest Publications
- Open Access
- 2 June 2026
- Published by Springer Nature in Research Square
Descriptor-Guided Model-Zoo Learning of Nonlinear Quality Attributes in Daidzein-Loaded Hydroxyapatite Nanoparticles
Daidzein-loaded hydroxyapatite nanoparticles are promising carrier systems for improving the formulation performance of poorly soluble bioactive compounds; however, their critical quality attributes may show nonlinear and response-specific behaviour that is not always fully captured by conventional response surface modelling. This study implemented a descriptor-guided model-zoo learning framework to reanalyse a 17-run Quality by Design dataset of daidzein-loaded hydroxyapatite nanoparticles. Sixteen formulation-derived descriptors were generated from hydroxyapatite-to-daidzein ratio, dispersion volume and loading time, including quadratic, interaction, logarithmic, square-root and ratio-based terms. Descriptor-subset screening was then performed across multiple regression algorithms for response-wise prediction of particle size, entrapment efficiency and zeta potential. The reconstructed QbD/RSM model was most suitable for particle size, achieving internal leave-one-out cross-validation R² of 0.965 and RMSE of 0.725 nm. For entrapment efficiency, descriptor-guided Gaussian Process regression using loading time, squared carrier ratio and carrier-to-volume ratio improved prediction compared with QbD/RSM, giving LOOCV R² of 0.473. Descriptor Elastic Net using squared loading time, log-transformed dispersion volume, squared carrier ratio and carrier-to-volume ratio was selected for zeta potential, achieving LOOCV R² of 0.381 and low optimized-checkpoint error. Explainable AI analysis using permutation importance, SHAP and LIME identified response-specific descriptor contributions, while Lasso-based surrogate equations and symbolic regression provided interpretable equation-based approximations. These findings show that descriptor-guided model-zoo learning can extract additional predictive and interpretative value from small QbD formulation datasets while preserving formulation relevance and response-wise interpretability.
- 2 June 2026
- Published by Springer Nature in Research Square
Open AccessDescriptor-Guided Model-Zoo Learning of Nonlinear Quality Attributes in Daidzein-Loaded Hydroxyapatite Nanoparticles
Daidzein-loaded hydroxyapatite nanoparticles are promising carrier systems for improving the formulation performance of poorly soluble bioactive compounds; however, their critical quality attributes may show nonlinear and response-specific behaviour that is not always fully captured by conventional response surface modelling. This study implemented a descriptor-guided model-zoo learning framework to reanalyse a 17-run Quality by Design dataset of daidzein-loaded hydroxyapatite nanoparticles. Sixteen formulation-derived descriptors were generated from hydroxyapatite-to-daidzein ratio, dispersion volume and loading time, including quadratic, interaction, logarithmic, square-root and ratio-based terms. Descriptor-subset screening was then performed across multiple regression algorithms for response-wise prediction of particle size, entrapment efficiency and zeta potential. The reconstructed QbD/RSM model was most suitable for particle size, achieving internal leave-one-out cross-validation R² of 0.965 and RMSE of 0.725 nm. For entrapment efficiency, descriptor-guided Gaussian Process regression using loading time, squared carrier ratio and carrier-to-volume ratio improved prediction compared with QbD/RSM, giving LOOCV R² of 0.473. Descriptor Elastic Net using squared loading time, log-transformed dispersion volume, squared carrier ratio and carrier-to-volume ratio was selected for zeta potential, achieving LOOCV R² of 0.381 and low optimized-checkpoint error. Explainable AI analysis using permutation importance, SHAP and LIME identified response-specific descriptor contributions, while Lasso-based surrogate equations and symbolic regression provided interpretable equation-based approximations. These findings show that descriptor-guided model-zoo learning can extract additional predictive and interpretative value from small QbD formulation datasets while preserving formulation relevance and response-wise interpretability.
- Open Access
- 2 June 2026
- Published by Springer Nature in Research Square
From Black-Box Prediction to White-Box Understanding: A Model-Zoo Comparison of Machine Learning Models for Predicting Rheological Properties of Emulgel Formulations
Rheological prediction is important in emulgel formulation development because viscosity directly influences physical stability, spreadability, residence time and product performance. However, artificial neural network (ANN)-based formulation models often provide limited transparency regarding the formulation and process variables responsible for prediction. In this study, a published Carbopol-based emulgel dataset was reanalysed using a multi-algorithm machine-learning workflow for forward viscosity prediction. The reconstructed dataset contained 11 formulation observations with Carbopol concentration, mixing time, stirring/mixing speed and experimental viscosity. Linear, regularized, ensemble and kernel-based regression models were compared using leave-one-out cross-validation. The selected ExtraTrees Regressor used Carbopol concentration, Carbopol² and stirring/mixing speed as predictors and achieved strong within-domain predictive agreement, with R² = 0.9783, MAE = 4.3867 Pa·s and RMSE = 4.9329 Pa·s. When applied to four reported available emulgel formulation cases, the model reduced MAE from 10.1568 to 4.9320 Pa·s and RMSE from 16.6718 to 5.8584 Pa·s relative to the source-study ANN-guided reference/target viscosity comparison. Predictor-level interpretation showed that viscosity was primarily associated with Carbopol concentration and its nonlinear squared term, while stirring/mixing speed contributed process-related information. This secondary computational reanalysis demonstrates that published emulgel datasets can be reused to obtain improved within-domain viscosity prediction and more transparent formulation-level interpretation, while prospective validation remains necessary before broader application.
- 2 June 2026
- Published by Springer Nature in Research Square
Open AccessFrom Black-Box Prediction to White-Box Understanding: A Model-Zoo Comparison of Machine Learning Models for Predicting Rheological Properties of Emulgel Formulations
Rheological prediction is important in emulgel formulation development because viscosity directly influences physical stability, spreadability, residence time and product performance. However, artificial neural network (ANN)-based formulation models often provide limited transparency regarding the formulation and process variables responsible for prediction. In this study, a published Carbopol-based emulgel dataset was reanalysed using a multi-algorithm machine-learning workflow for forward viscosity prediction. The reconstructed dataset contained 11 formulation observations with Carbopol concentration, mixing time, stirring/mixing speed and experimental viscosity. Linear, regularized, ensemble and kernel-based regression models were compared using leave-one-out cross-validation. The selected ExtraTrees Regressor used Carbopol concentration, Carbopol² and stirring/mixing speed as predictors and achieved strong within-domain predictive agreement, with R² = 0.9783, MAE = 4.3867 Pa·s and RMSE = 4.9329 Pa·s. When applied to four reported available emulgel formulation cases, the model reduced MAE from 10.1568 to 4.9320 Pa·s and RMSE from 16.6718 to 5.8584 Pa·s relative to the source-study ANN-guided reference/target viscosity comparison. Predictor-level interpretation showed that viscosity was primarily associated with Carbopol concentration and its nonlinear squared term, while stirring/mixing speed contributed process-related information. This secondary computational reanalysis demonstrates that published emulgel datasets can be reused to obtain improved within-domain viscosity prediction and more transparent formulation-level interpretation, while prospective validation remains necessary before broader application.
- Open Access
- 1 June 2026
- Published by Springer Nature in Research Square
From QbD to Explainable AI: Predictive Design Space Mapping of Lecithin/Chitosan Nanoparticles
Quality by Design (QbD) studies generate structured formulation datasets, but response surface models may not fully describe nonlinear or response-specific behavior in nanoparticle systems. This study reanalyzed a published silymarin-loaded lecithin/chitosan nanoparticle QbD dataset to determine whether independent explainable machine learning could add predictive and interpretive value to conventional QbD outputs. Twenty-nine formulation runs, comprising six formulation/process inputs and three critical quality attributes—particle size, polydispersity index (PDI), and entrapment efficiency—were extracted and modeled. A model-zoo of 105 regression variants per response was evaluated using leave-one-out cross-validation, followed by optimized-checkpoint comparison with reported QbD/RSM predictions, permutation feature importance, partial dependence analysis, and predictive design space mapping. Model performance was response dependent: XGBoost provided the best global particle-size prediction (R² = 0.926; RMSE = 27.52 nm), Extra Trees was best for PDI (R² = 0.695; RMSE = 0.0379), and distance-weighted KNN was best for entrapment efficiency (R² = 0.938; RMSE = 1.72%). At the optimized checkpoint, Gaussian Process regression predicted particle size with lower error than RSM (0.789 vs. 8.804 nm), and checkpoint-compatible Random Forest predicted PDI more closely than RSM (0.0004 vs. 0.032 absolute error), whereas RSM remained slightly closer for entrapment efficiency. Explainable AI identified lecithin:chitosan ratio and drug amount as recurring formulation drivers, while the predicted acceptable region was concentrated within intermediate formulation ranges. These findings support explainable machine learning as a complementary layer for QbD-based nanoparticle development rather than as a replacement for response surface modeling.
- 1 June 2026
- Published by Springer Nature in Research Square
Open AccessFrom QbD to Explainable AI: Predictive Design Space Mapping of Lecithin/Chitosan Nanoparticles
Quality by Design (QbD) studies generate structured formulation datasets, but response surface models may not fully describe nonlinear or response-specific behavior in nanoparticle systems. This study reanalyzed a published silymarin-loaded lecithin/chitosan nanoparticle QbD dataset to determine whether independent explainable machine learning could add predictive and interpretive value to conventional QbD outputs. Twenty-nine formulation runs, comprising six formulation/process inputs and three critical quality attributes—particle size, polydispersity index (PDI), and entrapment efficiency—were extracted and modeled. A model-zoo of 105 regression variants per response was evaluated using leave-one-out cross-validation, followed by optimized-checkpoint comparison with reported QbD/RSM predictions, permutation feature importance, partial dependence analysis, and predictive design space mapping. Model performance was response dependent: XGBoost provided the best global particle-size prediction (R² = 0.926; RMSE = 27.52 nm), Extra Trees was best for PDI (R² = 0.695; RMSE = 0.0379), and distance-weighted KNN was best for entrapment efficiency (R² = 0.938; RMSE = 1.72%). At the optimized checkpoint, Gaussian Process regression predicted particle size with lower error than RSM (0.789 vs. 8.804 nm), and checkpoint-compatible Random Forest predicted PDI more closely than RSM (0.0004 vs. 0.032 absolute error), whereas RSM remained slightly closer for entrapment efficiency. Explainable AI identified lecithin:chitosan ratio and drug amount as recurring formulation drivers, while the predicted acceptable region was concentrated within intermediate formulation ranges. These findings support explainable machine learning as a complementary layer for QbD-based nanoparticle development rather than as a replacement for response surface modeling.
- Open Access
- 15 May 2026
- Published by Springer Nature in Research Square
Explainable Machine Learning Using Taguchi-QbD Data for Digital Design Space Mapping of Pregabalin Extended-Release Tablets
This study applied explainable machine learning to published Taguchi-QbD data for pregabalin extended-release tablets to support formulation understanding and digital design space mapping. No new experimental batches were prepared; the analysis used formulation-response data from a previously developed pregabalin extended-release tablet system. A secondary data set containing 16 Taguchi formulations was prepared using six formulation and process variables: lipid material type, lipid quantity, hydrophilic matrix type, hydrophilic matrix quantity, filler type, and compression force. Hardness, friability, and dissolution at 1, 2, 4, and 8 h were selected as modeling responses. Ridge regression, random forest regression, extra trees regression, and Gaussian process regression were evaluated using leave-one-out cross-validation. Model-independent explainability was performed using permutation feature importance, partial dependence analysis, and surrogate decision-tree rules. A desirability-based virtual formulation screening approach was then used to identify high-desirability regions within the original experimental design limits. Dissolution at 2 h and 4 h showed the most useful cross-validated predictive performance, with mean absolute errors of 10.25% and 8.10%, respectively. Explainability analysis identified hydrophilic matrix quantity as the dominant variable controlling early and intermediate drug release, whereas lipid type and lipid quantity were more influential for tablet hardness and late-stage release behavior. Among the original Taguchi formulations, F15 showed the highest observed desirability, although it did not fully satisfy all dissolution criteria. Virtual screening identified high-desirability in silico candidates characterized mainly by high hydrophilic matrix loading and adequate lipid content. The published optimized formulation was positioned within a high-desirability AI-predicted region, supporting consistency between the explainable ML workflow and the original QbD optimization. These findings indicate that small-data explainable machine learning can complement QbD-based formulation analysis by extracting formulation rules and prioritizing design-space regions; however, AI-predicted candidates require experimental confirmation.
- 15 May 2026
- Published by Springer Nature in Research Square
Open AccessExplainable Machine Learning Using Taguchi-QbD Data for Digital Design Space Mapping of Pregabalin Extended-Release Tablets
This study applied explainable machine learning to published Taguchi-QbD data for pregabalin extended-release tablets to support formulation understanding and digital design space mapping. No new experimental batches were prepared; the analysis used formulation-response data from a previously developed pregabalin extended-release tablet system. A secondary data set containing 16 Taguchi formulations was prepared using six formulation and process variables: lipid material type, lipid quantity, hydrophilic matrix type, hydrophilic matrix quantity, filler type, and compression force. Hardness, friability, and dissolution at 1, 2, 4, and 8 h were selected as modeling responses. Ridge regression, random forest regression, extra trees regression, and Gaussian process regression were evaluated using leave-one-out cross-validation. Model-independent explainability was performed using permutation feature importance, partial dependence analysis, and surrogate decision-tree rules. A desirability-based virtual formulation screening approach was then used to identify high-desirability regions within the original experimental design limits. Dissolution at 2 h and 4 h showed the most useful cross-validated predictive performance, with mean absolute errors of 10.25% and 8.10%, respectively. Explainability analysis identified hydrophilic matrix quantity as the dominant variable controlling early and intermediate drug release, whereas lipid type and lipid quantity were more influential for tablet hardness and late-stage release behavior. Among the original Taguchi formulations, F15 showed the highest observed desirability, although it did not fully satisfy all dissolution criteria. Virtual screening identified high-desirability in silico candidates characterized mainly by high hydrophilic matrix loading and adequate lipid content. The published optimized formulation was positioned within a high-desirability AI-predicted region, supporting consistency between the explainable ML workflow and the original QbD optimization. These findings indicate that small-data explainable machine learning can complement QbD-based formulation analysis by extracting formulation rules and prioritizing design-space regions; however, AI-predicted candidates require experimental confirmation.
- Open Access
- 14 May 2026
- Published by Springer Nature in Research Square
Machine Learning Modeling of D-Optimal Design Data for Metformin Hydrochloride Orally Disintegrating Tablets
Metformin hydrochloride orally disintegrating tablets require an appropriate balance between mechanical strength and rapid disintegration, particularly because the drug is administered at a relatively high dose and has unfavorable taste characteristics. The present study applied machine learning to published D-optimal design data for metformin hydrochloride orally disintegrating tablets to evaluate whether formulation and process variables could predict powder-flow and tablet-performance responses. The dataset contained 30 experimental runs with four inputs: co-processed metformin hydrochloride ratio, co-processed excipient ratio, compression pressure, and excipient brand. Five responses were modeled separately: angle of repose, Hausner’s ratio, Carr’s index, hardness, and disintegration time. Polynomial Ridge Regression, Random Forest, Extra Trees, Gradient Boosting, Support Vector Regression, Gaussian Process Regression, and k-nearest neighbors were compared using leave-one-out cross-validation. Model performance was assessed using mean absolute error, root mean square error, and cross-validated R². Gaussian Process Regression was the best model for Hausner’s ratio, Carr’s index, hardness, and disintegration time, while Support Vector Regression performed best for angle of repose. Tablet hardness and disintegration time were predicted with the strongest performance, with R² values of 0.840 and 0.858, respectively. In contrast, angle of repose was weakly predicted, with an R² of 0.158. Permutation importance and partial dependence analysis indicated that compression pressure and excipient brand were the main contributors to tablet-level responses. The findings show that machine learning can support interpretation of D-optimal formulation data, especially for tablet-performance attributes, while predictions remain limited by dataset size and absence of new external validation.
- 14 May 2026
- Published by Springer Nature in Research Square
Open AccessMachine Learning Modeling of D-Optimal Design Data for Metformin Hydrochloride Orally Disintegrating Tablets
Metformin hydrochloride orally disintegrating tablets require an appropriate balance between mechanical strength and rapid disintegration, particularly because the drug is administered at a relatively high dose and has unfavorable taste characteristics. The present study applied machine learning to published D-optimal design data for metformin hydrochloride orally disintegrating tablets to evaluate whether formulation and process variables could predict powder-flow and tablet-performance responses. The dataset contained 30 experimental runs with four inputs: co-processed metformin hydrochloride ratio, co-processed excipient ratio, compression pressure, and excipient brand. Five responses were modeled separately: angle of repose, Hausner’s ratio, Carr’s index, hardness, and disintegration time. Polynomial Ridge Regression, Random Forest, Extra Trees, Gradient Boosting, Support Vector Regression, Gaussian Process Regression, and k-nearest neighbors were compared using leave-one-out cross-validation. Model performance was assessed using mean absolute error, root mean square error, and cross-validated R². Gaussian Process Regression was the best model for Hausner’s ratio, Carr’s index, hardness, and disintegration time, while Support Vector Regression performed best for angle of repose. Tablet hardness and disintegration time were predicted with the strongest performance, with R² values of 0.840 and 0.858, respectively. In contrast, angle of repose was weakly predicted, with an R² of 0.158. Permutation importance and partial dependence analysis indicated that compression pressure and excipient brand were the main contributors to tablet-level responses. The findings show that machine learning can support interpretation of D-optimal formulation data, especially for tablet-performance attributes, while predictions remain limited by dataset size and absence of new external validation.
- Open Access
Artificial Intelligence and Circulating microRNA Signatures for Early Breast Cancer Detection: A Systematic Review and Meta-Analysis
Background: Early breast cancer detection remains central to improving clinical outcomes, yet conventional screening pathways, particularly mammography, have recognized limitations in sensitivity, specificity, and performance in dense breast tissue. Circulating microRNAs (miRNAs) have emerged as promising minimally invasive biomarkers, while artificial intelligence and machine learning (AI/ML) offer powerful tools for identifying diagnostically relevant multi-marker patterns within complex biomarker datasets. This systematic review and meta-analysis evaluated the diagnostic performance of AI/ML-based circulating miRNA signatures for early breast cancer detection. Methods: A systematic search of PubMed/MEDLINE, Scopus, and Web of Science Core Collection was conducted from database inception to 31 December 2025. Studies were eligible if they were original human investigations evaluating circulating miRNAs using an AI/ML-based diagnostic model for breast cancer detection and reporting extractable diagnostic performance metrics. Study selection followed PRISMA 2020 and PRISMA-DTA guidance. Methodological quality was assessed using QUADAS-2. Pooled sensitivity and specificity were synthesized using a bivariate random-effects model, and overall diagnostic performance was summarized using a hierarchical summary receiver operating characteristic framework. Results: Seven studies met the inclusion criteria for qualitative synthesis, with eligible studies contributing to the quantitative analysis depending on data availability. Across the pooled analysis, AI/ML-based circulating miRNA models demonstrated good overall diagnostic performance, with a pooled AUC of 0.905 (95% CI: 0.890–0.921), pooled sensitivity of 81.3% (95% CI: 76.8%–85.2%), and pooled specificity of 87.0% (95% CI: 82.4%–90.7%). Heterogeneity was moderate for AUC (I² = 42.3%) and sensitivity (I² = 38.7%) and low for specificity (I² = 28.4%). Risk-of-bias assessment showed overall low-to-moderate methodological concern, with patient selection representing the most variable domain. Deeks’ funnel plot asymmetry test showed no significant evidence of publication bias (p = 0.34). Conclusions: AI/ML-based circulating miRNA signatures show promising diagnostic accuracy for early breast cancer detection and may have value as non-invasive adjunctive tools within imaging-supported diagnostic pathways. However, the evidence base remains limited by methodological heterogeneity, variable validation rigor, and the predominance of retrospective case-control designs. Prospective, standardized, and externally validated studies are needed before routine clinical implementation can be justified.
Open AccessArtificial Intelligence and Circulating microRNA Signatures for Early Breast Cancer Detection: A Systematic Review and Meta-Analysis
Background: Early breast cancer detection remains central to improving clinical outcomes, yet conventional screening pathways, particularly mammography, have recognized limitations in sensitivity, specificity, and performance in dense breast tissue. Circulating microRNAs (miRNAs) have emerged as promising minimally invasive biomarkers, while artificial intelligence and machine learning (AI/ML) offer powerful tools for identifying diagnostically relevant multi-marker patterns within complex biomarker datasets. This systematic review and meta-analysis evaluated the diagnostic performance of AI/ML-based circulating miRNA signatures for early breast cancer detection. Methods: A systematic search of PubMed/MEDLINE, Scopus, and Web of Science Core Collection was conducted from database inception to 31 December 2025. Studies were eligible if they were original human investigations evaluating circulating miRNAs using an AI/ML-based diagnostic model for breast cancer detection and reporting extractable diagnostic performance metrics. Study selection followed PRISMA 2020 and PRISMA-DTA guidance. Methodological quality was assessed using QUADAS-2. Pooled sensitivity and specificity were synthesized using a bivariate random-effects model, and overall diagnostic performance was summarized using a hierarchical summary receiver operating characteristic framework. Results: Seven studies met the inclusion criteria for qualitative synthesis, with eligible studies contributing to the quantitative analysis depending on data availability. Across the pooled analysis, AI/ML-based circulating miRNA models demonstrated good overall diagnostic performance, with a pooled AUC of 0.905 (95% CI: 0.890–0.921), pooled sensitivity of 81.3% (95% CI: 76.8%–85.2%), and pooled specificity of 87.0% (95% CI: 82.4%–90.7%). Heterogeneity was moderate for AUC (I² = 42.3%) and sensitivity (I² = 38.7%) and low for specificity (I² = 28.4%). Risk-of-bias assessment showed overall low-to-moderate methodological concern, with patient selection representing the most variable domain. Deeks’ funnel plot asymmetry test showed no significant evidence of publication bias (p = 0.34). Conclusions: AI/ML-based circulating miRNA signatures show promising diagnostic accuracy for early breast cancer detection and may have value as non-invasive adjunctive tools within imaging-supported diagnostic pathways. However, the evidence base remains limited by methodological heterogeneity, variable validation rigor, and the predominance of retrospective case-control designs. Prospective, standardized, and externally validated studies are needed before routine clinical implementation can be justified.