FormulationDE Predicts Drug Excipient Compatibility Using Interpretable Machine Learning and Chemical Rules

Pharmaceutical researchers have developed FormulationDE, an integrated artificial intelligence platform designed to predict whether active pharmaceutical ingredients and prospective excipients can remain stable when combined in medicine formulations. During the early preformulation phase of pharmaceutical development, assessing potential interactions between active drug molecules and excipient compounds is critical, as chemical or physical incompatibilities can diminish therapeutic efficacy, alter release kinetics, or produce toxic degradation products. Traditional compatibility assessments have historically depended upon laboratory testing techniques such as differential scanning calorimetry, Fourier-transform infrared spectroscopy, and high-performance liquid chromatography, which are notoriously time-consuming and resource-intensive. To overcome these experimental bottlenecks, investigators expanded the rule-based PharmDE expert system into a machine learning platform capable of rapid computational screening. By pairing predictive algorithmic classifiers with explicit chemical substructure descriptors and visual model interpretability, the newly introduced system offers pharmaceutical scientists a computational screening tool that highlights structural interaction risks before committing extensive physical resources to laboratory trials.

To establish a reliable computational framework, the investigators curated an experimentally verified dataset comprising 1,105 distinct drug-excipient pairs collected from more than 300 published scientific papers and the authoritative Handbook of Pharmaceutical Excipients. In contrast to earlier computational models that relied heavily on regulatory drug product labeling where inactive ingredients can occasionally interact despite market approval, this study restricted its data collection exclusively to studies verified by differential scanning calorimetry, Fourier-transform infrared spectroscopy, and high-performance liquid chromatography. Under these strict experimental criteria, a drug-excipient combination was labeled incompatible if calorimetric thermograms exhibited significant peak shifts, disappearances, or new thermal events, if infrared spectra demonstrated functional group alterations, or if chromatographic assays identified degradation impurities. The final cleaned dataset incorporated 532 entries previously documented in the PharmDE database alongside 573 newly extracted entries, yielding a balanced collection of 579 compatible and 526 incompatible pairs encompassing 194 distinct active pharmaceutical ingredients and 100 excipients.

Molecular representation required translating chemical structures into machine-readable mathematical features that preserve chemical interpretability during algorithmic decision-making. The research team retrieved simplified molecular-input line-entry system strings from the PubChem database for all 194 drugs and 100 excipients and utilized the open-source RDKit software library to generate 166-bit MACCS structural fingerprints. Because standard MACCS fingerprints map predetermined molecular fragments into binary bits, they provide direct chemical transparency regarding the specific functional groups that influence model predictions. Recognizing that standard fingerprints might omit specific functional motifs central to drug formulation science, the investigators supplemented the fingerprint profiles with 19 custom-designed chemical substructure descriptors denoted as DE-des. These supplemental descriptors were formulated using the SMARTS chemical pattern language to capture specific functional groups across drugs and excipients, enhancing the feature space to reflect interaction mechanisms while undergoing systematic deduplication to prevent redundant structural encoding.

To evaluate prospective predictive capability without artificial performance inflation, the researchers partitioned the 1,105 entries chronologically rather than utilizing a simple random split. All experimental records published before 2023, totaling 1,030 drug-excipient records divided between 520 compatible and 510 incompatible pairs, formed the dedicated model development set. Data collected across the 2023 to 2024 period was isolated completely as an independent prospective test set. This temporal split reflected practical pharmaceutical development conditions where algorithms must generalize to newly synthesized compounds and experimental observations without risk of training data leakage. To refine model construction and suppress overfitting, the investigators applied recursive feature elimination to identify optimal feature subsets, systematically pruning low-importance variables based on feature importance scoring to reduce computational complexity and eliminate irrelevant structural noise before finalizing the feature sets across the evaluated machine learning algorithms.

The researchers trained and benchmarked eight distinct machine learning algorithms using five-fold nested cross-validation, employing Bayesian optimization via BayesSearchCV to systematically fine-tune hyperparameters across prospective parameter regions. The candidate architectures included decision trees, logistic regression, k-nearest neighbors, support vector machines, random forests, LightGBM gradient boosting, artificial neural networks, and naive Bayes classifiers. Evaluation metrics spanned prediction accuracy, precision, recall, F1 score, receiver operating characteristic area under the curve, and Matthew’s correlation coefficient to measure balanced binary classification performance. Among all competing models, the LightGBM classifier demonstrated superior predictive capability, attaining an accuracy of 0.75, precision of 0.75, recall of 0.75, an F1 score of 0.74, a receiver operating characteristic area under the curve of 0.82, and a Matthew’s correlation coefficient of 0.50 during nested cross-validation, indicating robust discrimination between compatible and incompatible combinations.

When assessed against the isolated 2023 to 2024 independent test set, FormulationDE confirmed generalizable predictive performance and outperformed previously published computational tools, including the stacking ensemble model designated as DECW introduced by Nguyen and colleagues. Confusion matrix evaluations on the recent experimental test set demonstrated that the LightGBM classifier maintained consistent discrimination between compatible and incompatible formulations on newly reported experimental data. The authors noted that prior platforms, such as DE-INTERACT based on artificial neural networks and DECW based on ensemble stacking, derived training records largely from approved drug labels. However, because marketed products occasionally tolerate minor incompatibilities mitigated by specialized manufacturing technologies, coating barriers, or negligible clinical risk, models trained on label data struggle with laboratory-level chemical incompatibility. Grounding FormulationDE in direct analytical chemistry observations provided more consistent alignment with experimental preformulation outcomes.

To explain the chemical reasoning driving individual algorithmic predictions, the researchers coupled the trained LightGBM model with Shapley Additive Explanations analysis and an established expert rule base. SHAP summary plots ranking the top twenty influential features revealed that specific functional groups strongly dictated prediction directionality. Features shifting predictions toward compatibility included phenolic hydroxyl groups, imino functional groups, and alkaline earth elements. Conversely, the presence of six-membered ring structures, aromatic motifs, and oxygen-containing moieties significantly increased algorithmic predictions of incompatibility. These empirical findings aligned with sixty mechanistic interaction rules preserved from the PharmDE expert system. For instance, excipient aldehyde groups readily undergo condensation reactions such as Schiff base formation with primary or secondary drug amines. Similarly, basic or catalytic excipient amines can induce drug amino moiety instability and dimerization, while tertiary drug amines are highly vulnerable to oxidation by excipient peroxide impurities to form degradation N-oxides.

To translate these algorithmic capabilities into accessible laboratory workflows, the research team deployed FormulationDE as an interactive, publicly accessible web platform hosted on the Streamlit cloud infrastructure. Formulators can input chemical representations as SMILES strings or search excipients directly by common name, evaluating up to ten candidate excipient compounds simultaneously against a single active pharmaceutical ingredient. The computational pipeline processes the chemical inputs, calculates probability scores reflecting the likelihood of compatibility or incompatibility, and outputs downloadable prediction tables in CSV format. Crucially, the web interface generates dynamic SHAP force plots alongside every numerical probability score, visually highlighting which molecular substructures increased or decreased compatibility risks for each specific pairing. Furthermore, the platform integrates direct references to the sixty chemical interaction rules from the PharmDE knowledge base, providing pharmaceutical scientists with mechanistic chemistry-oriented warnings alongside quantitative probability values.

The research team conducted a targeted retrospective case study evaluating FormulationDE against published high-throughput screening data generated by Nicole Wyttenbach and colleagues, who tested acetylsalicylic acid and fluoxetine hydrochloride across 96-well plates stored under accelerated environmental conditions. In that historical study, drug-excipient blends prepared by solvent evaporation were held at 40 or 50 degrees Celsius under 10 or 75 percent relative humidity for one or four weeks. FormulationDE correctly predicted nine out of fourteen experimental outcomes, demonstrating acceptable predictive utility while exposing key computational limitations. The model incorrectly predicted compatibility for acetylsalicylic acid mixed with specific excipients because its static structural descriptors did not capture environmental humidity and local pH shifts that drive ester hydrolysis into salicylic and acetic acids. Similarly, the model misjudged fluoxetine hydrochloride interactions with Primojel because it could not model dynamic sodium ion exchange forming soluble complexes, nor could it capture the crystallization inhibition effects of PVP 30.

Despite its demonstrated predictive strengths, the authors emphasized several operational boundaries and outlined essential directions for future software iterations. The investigators plan to incorporate expanded real-world datasets, deep learning architectures capable of capturing non-linear chemical interactions, and incremental learning methodologies to assimilate continuous industrial manufacturing feedback. By integrating experimental high-throughput screening pipelines with adaptive computational algorithms, future iterations of FormulationDE aim to advance computational pharmaceutics and the emerging Quality by Computational Design paradigm, transitioning pharmaceutical formulation from empirical trial-and-error toward predictive, transparent digital design.

Subject of Research: Artificial intelligence and machine learning for predicting drug-excipient compatibility in pharmaceutical formulation

Article Title: FormulationDE: an updated artificial intelligence system for drug-excipient compatibility prediction

Article References: Jia, S., Wang, N., Yang, R., Zhao, X., Ding, Y., & Ouyang, D. (2026). FormulationDE: an updated artificial intelligence system for drug-excipient compatibility prediction. AAPS Open, 12(1), Article 20. https://doi.org/10.1186/s41120-026-00156-4

Image Credits: AI Generated

DOI: 10.1186/s41120-026-00156-4

Keywords: FormulationDE, drug-excipient compatibility, machine learning, LightGBM, SHAP analysis, MACCS fingerprints, SMARTS descriptors, preformulation, PharmDE, differential scanning calorimetry, Fourier-transform infrared spectroscopy, high-performance liquid chromatography, active pharmaceutical ingredients, recursive feature elimination, BayesSearchCV, chemical substructures, molecular descriptors, computational pharmaceutics