Source-linked AI summary
Deep learning for in vitro prediction of pharmaceutical formulations
Yilong Yang, Zhuyifan Ye, Yan Su, Qianqian Zhao, Xiaoshan Li, Defang Ouyang
TL;DR
Pharmaceutical formulation development remains dependent on laborious, costly trial-and-error work and limited formulation data. This study developed deep-learning models for two dosage forms, combined automatic data splitting with pharmaceutics-specific evaluation criteria, and compared them with six machine-learning methods. Both deep neural networks achieved accuracies above 80% and outperformed the other machine-learning models, supporting a shift toward data-driven formulation research.
Problem
Formulation development relies on laborious, time-consuming, and costly trial-and-error work, while reliable and standard formulation data are limited.
Method
The study developed deep-learning models for two dosage forms, used an automatic data-splitting approach and pharmaceutics-specific evaluation criteria, and compared them with six machine-learning methods.
Results
Both deep neural networks achieved accuracies above 80% and higher accuracy than the other machine-learning models.
Takeaways & Limitations
Deep learning was successfully developed for pharmaceutical-formulation prediction and may support a shift from experience-dependent to data-driven research.
Takeaways & Limitations
ANN formulation-prediction accuracy is relatively low when experimental data are limited, and reliable standard formulation data remain difficult to obtain.
Abstract
from arXiv · showhide
Current pharmaceutical formulation development still strongly relies on the traditional trial-and-error approach by individual experiences of pharmaceutical scientists, which is laborious, time-consuming and costly. Recently, deep learning has been widely applied in many challenging domains because of its important capability of automatic feature extraction. The aim of this research is to use deep learning to predict pharmaceutical formulations. In this paper, two different types of dosage forms were chosen as model systems. Evaluation criteria suitable for pharmaceutics were applied to assessing the performance of the models. Moreover, an automatic dataset selection algorithm was developed for selecting the representative data as validation and test datasets. Six machine learning methods were compared with deep learning. The result shows the accuracies of both two deep neural networks were above 80% and higher than other machine learning models, which showed good prediction in pharmaceutical formulations. In summary, deep learning with the automatic data splitting algorithm and the evaluation criteria suitable for pharmaceutical formulation data was firstly developed for the prediction of pharmaceutical formulations. The cross-disciplinary integration of pharmaceutics and artificial intelligence may shift the paradigm of pharmaceutical researches from experience-dependent studies to data-driven methodologies.
1. Introduction
Pharmaceutical formulation development remains laborious, costly, and dependent on individual experience, motivating systematic data-driven methods. Machine learning, especially deep learning, is presented as a way to automate feature extraction and improve formulation prediction.
- Motivation: These limitations motivate efficient and systematic formulation-development methods that can keep pace with pharmaceutical-industry requirements.The introduction frames systematic development as necessary under pressure to reduce healthcare costs.
- Motivation: Formulation development still relies on trial-and-error studies based on individual pharmaceutical scientists’ experiences.The approach is described as laborious, time-consuming, and costly, with difficulty achieving optimum formulations.
- Motivation: Machine learning can use existing experimental data to make data-driven predictions for more efficient formulation development.The paper associates well-designed machine learning methods with faster development, formulation optimization, cost savings, and preserved expert knowledge.
- Prior methods: Expert systems struggle to convert vague pharmaceutical expertise into rules that accurately predict formulation performance.The limitation is attributed to the difficulty of extracting experts’ experiences into explicit rules.
- Prior methods: ANNs can address problems that standard expert systems cannot, but they require expert-designed feature extractors.Their formulation-prediction accuracy is also described as relatively low because experimental data are limited.
2. Methods
The study assembled formulation data for two dosage forms, represented formulation inputs and dosage-form characteristics, and trained models using separate training, validation, and test sets. It compared six conventional machine-learning methods with deep neural networks and evaluated predictions using pharmaceutics-relevant criteria.
- Pharmaceutical data: The dataset contains 131 oral fast-dissolving-film formulations and 145 sustained-release-matrix-tablet formulations.The two dosage forms use formulation composition, process parameters, and in vitro characteristics as data sources or targets.
- Pharmaceutical data: The prediction targets included dosage-form characteristics such as oral-fast-dissolving-film disintegration time and sustained-release-matrix-tablet dissolution profiles.The sustained-release profiles were measured cumulatively at 2, 4, 6, and 8 hours.
- Pharmaceutical data: Molecular descriptors represented API properties, while excipient encodings and process parameters represented formulation inputs.Descriptors included molecular weight, XLogP3, hydrogen-bond counts, rotatable bonds, polar surface area, heavy atoms, complexity, and logS.
- Data splitting: Each dosage-form dataset was divided into training, validation, and test subsets, with 20 formulations assigned to both validation and test sets.The validation set tuned hyperparameters, while the test set assessed prediction on unknown data.
- Modeling: Six regression methods—MLR, PLSR, SVM, ANNs, RF, and k-NN—were compared with deep neural networks.Deep networks were trained as feed-forward architectures, using 10 layers for oral fast-dissolving films and 9 layers for sustained-release matrix tablets.
- Evaluation: Evaluation used pharmaceutics-specific success criteria in addition to conventional regression measures.A cumulative dissolution prediction was successful when f2 was at least 50, while disintegration-time prediction required an error no greater than 10 seconds.
3. Results and Discussion
The study addresses formulation prediction with small, imbalanced datasets by combining deep learning, pharmaceutics-specific evaluation criteria, and automated representative-data selection. Across OFDF and SRMT, MD-FIS improved dataset splitting, while deep neural networks generally outperformed conventional machine-learning models.
- Dataset challenges: Each dosage form had only around 140 formulations, and nearly half of the APIs included fewer than four formulations.These small, imbalanced datasets make representative training, validation, and test selection important.
- Dataset splitting: Random splitting produced accuracy variation exceeding 40%, with average accuracy below 60%.The method could select APIs with very few formulations, reducing prediction reliability.
- Dataset splitting: 83.46% for OFDF and 78.85% for SRMT were the validation accuracies obtained with the original maximum dissimilarity algorithm.The algorithm preferentially selected small API groups, boundary formulations, and extraordinary formulations.
- Model comparison: All deep neural-network accuracies exceeded 80%, and deep learning achieved the highest training, validation, and test accuracies for both OFDF and SRMT.The conventional machine-learning methods did not achieve sufficient performance, with SRMT accuracies ranging from 25% to 55%.
- Model evaluation: Experimental and deep-learning-predicted values were very close, supporting satisfactory prediction performance for the OFDF and SRMT test evaluations.The paper compares predicted disintegration times and cumulative drug-release curves with experimental results.
4. Conclusion
Deep learning models were developed for pharmaceutical formulation prediction and showed good generalization and effective prediction of key characteristics. The authors suggest that integrating pharmaceutics with artificial intelligence may support a shift from experience-dependent to data-driven research.
- Deep learning models were successfully developed to predict pharmaceutical formulations.
- The models showed good generalization performance on external datasets and effectively predicted key characteristics in regression problems.
- The models performed better than those trained by other machine learning methods, which the authors attribute to finding complex correlations in the data.
- Machine learning could help predict in vivo and in vitro characteristics, assist experiment design, and support product-quality control across the product cycle.
- The integration of pharmaceutics and artificial intelligence may shift pharmaceutical research from experience-dependent studies toward data-driven methodologies.