📄 bioengineering

Medicament identity rather than total loading governs the morphology of electrospun poly(vinylpyrrolidone) nanofibers for regenerative endodontics: a machine learning analysis of a failure-inclusive dataset

This study demonstrates that for electrospun poly(vinylpyrrolidone) nanofibers used in regenerative endodontics, the specific identity of the loaded medicament is a superior predictor of fiber morphology than total drug loading, as revealed by a machine learning analysis of a failure-inclusive dataset.

Original authors: Brimo, N., UYSAL, B., SERDAROGLU, D. C.

Published 2026-09-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Brimo, N., UYSAL, B., SERDAROGLU, D. C.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Inside a tooth that has lost its nerve, the root often remains a hollow, fragile tube with an open end, unable to grow any further. To save such a tooth, dentists must clean out the infection without killing the delicate stem cells needed to rebuild the root. The standard treatment uses a thick paste of antibiotics, but the very concentration required to kill the bacteria is often high enough to poison those same stem cells. A newer approach tries to solve this by using a mat of microscopic fibers to deliver the medicine. These fibers act like a sponge, holding the drug and releasing it slowly, which allows doctors to use a much smaller dose. However, making these fibers is a trial-and-error process. Scientists mix a polymer with a drug and spin it into fibers, but they often struggle to predict what the final product will look like. They assume that if they simply change the amount of drug in the mix, they can control the thickness of the fibers. A new study challenges this assumption, suggesting that the identity of the drug matters far more than the total amount.

Researchers set out to test this idea by creating a specific collection of twenty-nine different fiber recipes. They used a single type of plastic polymer dissolved in alcohol and kept the machine settings exactly the same for every batch. The only thing they changed was the medicine added to the mix. They tested four different drugs: metronidazole, ciprofloxacin, minocycline, and calcium hydroxide, sometimes using them alone and sometimes in combinations. Crucially, they did not discard the batches that failed. If a mixture produced no fibers at all, or if it created a string of beads instead of smooth threads, they recorded that failure just as carefully as the successes. This created a complete map of what works and what does not, rather than just a list of the best results.

The team then used computer models to analyze this data, asking a simple question: can you predict the thickness of the fiber if you only know the total weight of the medicine in the mix? The answer was a clear no. Models that looked only at the total amount of drug performed no better than a random guess. However, when the computer was told exactly which drug was present and in what specific amount, it could predict the fiber thickness with high accuracy. The study revealed that different drugs pull the fibers in opposite directions. One drug, ciprofloxacin, made the fibers thinner and thinner as its concentration increased. Another drug, metronidazole, made the fibers thicker until they became so unstable that they stopped forming fibers entirely, turning into a string of beads. Because these two drugs push the outcome in opposite ways, simply adding up their total weight hides the true effect. A mixture with a high total weight might spin perfectly if it contains the right drug, while a mixture with a lower total weight might fail completely if it contains the wrong one.

This finding changes how scientists should approach making these medical fibers. Instead of testing one drug at many different concentrations, it is more efficient to test a few different drugs at a few concentrations. The computer analysis showed that by understanding the specific behavior of each drug, researchers could find the best recipe with far fewer experiments. The study also confirmed that these fibers are not just a different way to carry the same paste; they are functionally superior. In tests, the fibers killed bacteria at concentrations eight times lower than the traditional paste, and they were much easier to remove from the tooth canal after treatment. While the computer models could not predict the outcome for a completely new drug they had never seen before, they proved highly effective at navigating the known options available for this specific medical need. By focusing on the identity of the medicine rather than just the quantity, the researchers have provided a clearer path to creating safer, more effective treatments for saving damaged teeth.

Technical Summary: Medicament Identity vs. Total Loading in Electrospun Nanofibers for Regenerative Endodontics

Problem Statement
In regenerative endodontics, electrospun nanofibers loaded with antibiotics or calcium hydroxide are developed as intracanal carriers to disinfect immature teeth while sparing stem cells. Current formulation strategies typically fix the polymer and machine settings while sweeping the total medicament concentration, operating under the assumption that total loading is the primary determinant of fiber morphology (diameter, uniformity, and spinnability). However, existing machine learning models for electrospinning are largely trained on literature-mined "success-only" datasets that pool diverse polymers and solvents, often reducing complex cargo compositions to a single concentration figure. This approach fails to capture the specific interactions between different actives and the polymer matrix, and it ignores the "failure boundary" where formulations produce beads or no fibers.

Methodology
The authors constructed ENDOSPIN-29, a dataset of 29 poly(vinylpyrrolidone) (PVP) formulations produced under a single, fixed process backbone (8% w/w PVP in ethanol, fixed flow rate, voltage, and distance). The dataset includes:

Actives: Metronidazole (MET), ciprofloxacin (CIP), minocycline, and calcium hydroxide, used alone and in combination (double and triple antibiotic mixtures).
Failure Inclusion: Unlike typical literature datasets, this collection explicitly retains five formulations that failed to produce submicron fibers (producing beads or micron-scale structures instead).
Feature Sets: The study compares two input representations for machine learning models: (1) Total Loading (sum of all medicament weight fractions) and (2) Per-Medicament Composition (individual weight fractions of each active).
Models: Nine regression models (including Gaussian Processes, Gradient Boosting, and Support Vector Regressors) were trained using leave-one-out cross-validation (LOOCV) to predict the logarithm of mean fiber diameter, uniformity (coefficient of variation), distribution span, and asymmetry. Classification models were also tested for spinnability.
Analysis: The study includes dose-response inference, class extrapolation tests (holding out entire medicament classes), and a retrospective sequential design simulation to evaluate how efficiently a Gaussian Process could navigate the design space compared to random selection.

Key Results

Composition Governs Morphology: Models using per-medicament composition significantly outperformed those using total loading. The best model (Gradient Boosting) achieved a leave-one-out coefficient of determination ( $R^2$ ) of 0.84 and a median relative error of 18% for fiber diameter prediction. In contrast, models using only total loading performed at or below the baseline (predicting the mean), with some yielding negative $R^2$ values.
Opposing Dose Responses: Different actives drive morphology in opposite directions despite similar total loadings:
- Ciprofloxacin: Thins fibers monotonically (from 406 nm to 257 nm) as concentration increases.
- Metronidazole: Thickens fibers initially but destroys fiber formation (creating beads) above 10% w/w.
- Triple Antibiotic: Produces large diameters (>1.8 µm) at 7% of each component, whereas ciprofloxacin alone at 7.5% produces sub-300 nm fibers.
Uniformity and Spinnability: Composition-based models also predicted fiber uniformity ( $R^2$ = 0.54) and spinnability (AUC = 0.83) significantly better than loading-only models. However, asymmetry (skewness of the diameter distribution) remained unpredictable by either feature set.
Generalization Limits: The advantage of composition-based features is limited to interpolation within the known drug panel. When an entire medicament class (e.g., ciprofloxacin) was held out from training, the model's predictive power collapsed ( $R^2$ dropped to -0.06), indicating the models lack the molecular descriptors necessary to predict the behavior of unseen drugs.
Sequential Design Efficiency: A Gaussian Process-guided sequential design algorithm identified the optimal formulation in the pool in approximately 7.5 runs, compared to 14.3 runs for random selection, demonstrating a 1.7-fold improvement in efficiency under a payload-first objective.
Cross-Laboratory Variance: While the model transferred reasonably well between two laboratories for most formulations, a discrepancy was observed with 5% calcium hydroxide: one lab produced no fibers, while the other produced 1600 nm fibers. The authors attribute this to unrecorded variables (e.g., particle size, dispersion) affecting the suspension-based calcium hydroxide.

Significance and Claims
The paper claims that medicament identity is a more critical variable than total loading for governing the morphology of electrospun nanofibers in this specific context. The central methodological contribution is the inclusion of failure data, which allows the model to learn the boundaries of the feasible region rather than just the interior of successful runs.

The authors argue that standard formulation sweeps (varying one drug while holding others constant) traverse the "least informative" axis. Instead, sampling across different actives at fewer concentration levels would map the design space more efficiently. However, the paper is modest regarding its scope:

The models are interpolative, not extrapolative; they cannot predict the behavior of novel drugs without molecular descriptors.
The dataset is small (29 formulations), meaning the reported accuracy is a lower bound and confidence intervals are wide.
The findings are specific to the PVP/ethanol system and the specific actives tested, which are the standard drugs for regenerative endodontics.

The study concludes that for defined panels of actives in regenerative endodontics, machine learning models trained on failure-inclusive, composition-specific data can significantly reduce the experimental burden of formulation development, provided the models are not expected to generalize to entirely new chemical classes.

Technical Summary: Medicament Identity vs. Total Loading in Electrospun Nanofibers for Regenerative Endodontics

More like this