🧊 Welcome to the Polymer Property Prediction Platform

Discover our cutting-edge informatics platform that utilizes advanced machine learning algorithms to predict crucial polymer properties such as radius of gyration and surface accessible area.
Designed to support nanoplatform development, this tool helps researchers accelerate the design of smart drug and gene delivery systems through precise computational insights.
This AI web platform is part of the project: β€œAdaptive Design and Assembly of Polymer-Based Nanoplatforms for Smart Gene and Drug Delivery”.

Project Partners:

INCDTIM National Institute for Research and Development of Isotopic and Molecular Technologies – INCDTIM
UBB Faculty of Mathematics and Computer Science, BabeΘ™-Bolyai University – BBU

πŸ’‘ About the Platform

Our platform empowers researchers and developers by providing rapid, reliable predictions of polymer characteristics.
Leveraging pre-trained machine learning models built on extensive, curated datasets, users can efficiently evaluate potential materials, reducing the need for time-consuming experimental testing and streamlining the development pipeline.

🧩 Key Features

Property Prediction:
Instantly estimate vital polymer properties like radius of gyration and surface accessible area to inform nanoplatform design.

Pre-Built Machine Learning Models:
Utilize models trained on diverse polymer datasets, ensuring high accuracy and broad applicability across different chemistries.

User-Friendly Interface:
Easily input chemical structures or descriptors and receive immediate property predictions, enabling rapid screening and iterative design.

Customizable Parameters:
Tailor prediction settings or input specific features to meet your unique research requirements and hypotheses.

βš™οΈ How It Works

Input Data:
Select from predefined molecular descriptors within the platform.

Model Processing:
Our sophisticated machine learning algorithms analyze the input data to generate property predictions.

Results & Insights:
Review predicted valuesβ€”including radius of gyration and surface accessible areaβ€”to guide your material selection and nanoplatform design decisions.

πŸ“Š Properties Generator: Radius of Gyration

Polymer repeat unit
Polymer repeat unit

πŸ“Š Properties Generator: SASA

Polymer repeat unit
Polymer repeat unit

🧬 Application in Nanoplatform Design

Design Optimization:
Use predicted properties to identify polymers with ideal dimensions and surface characteristics for efficient nanoparticle assembly.

Personalized Medicine Development:
Customize polymer formulations to meet specific drug or gene delivery needs, minimizing experimental trial-and-error.

Accelerate R&D:
Expedite the discovery process by focusing on the most promising candidates early, reducing costs and development time.

🎯 Research & Innovation Objectives

Our platform aligns with key project goals:

Enabling adaptive design of polymer-based nanoplatforms (PBNs) tailored for targeted drug and gene delivery.

Developing multi-stage protocols for engineering stable, self-assembling polymer nanostructures.

Providing in silico validation tools to ensure candidate polymers are effective and stable for delivery applications.

πŸ“˜ Learn More

Gain deeper insights into how computational modeling is revolutionizing polymer nanomedicine.

Project Abstract: Advances in gene therapy and nanotechnology require efficient delivery vectors. This project introduces a machine learning-driven computational approach to design and validate polymer nanoplatforms for personalized medicine, aiming to reduce socio-economic burdens and improve therapeutic outcomes.

Objectives Overview: From adaptive design strategies to in silico validation, our research harnesses AI to innovate drug and gene delivery systems.

πŸ“š AdaptDelivery Documentation

This documentation provides a detailed scientific and technical overview of the AdaptDelivery platform. It explains the molecular modeling workflow, the structural descriptors used for prediction, the mathematical definitions of the computed properties, the machine learning framework, and the way users should interpret the platform outputs.

1. Platform Overview

AdaptDelivery is a polymer informatics platform that integrates molecular modeling, molecular dynamics simulations, machine learning and interactive web technologies to support the rapid prediction and visualization of polymer structural properties.

The platform was developed to reduce the computational effort associated with conventional molecular dynamics workflows. Instead of performing a new simulation for every polymer system and chain length, users can obtain machine learning-based predictions directly through a standard web browser.

The current version focuses on two key structural descriptors: radius of gyration (Rg) and solvent-accessible surface area (SASA). These properties provide information about polymer dimensions, compactness, conformational behavior, solvent exposure, hydration, and possible aggregation tendencies.

In addition to numerical predictions, AdaptDelivery provides interactive property-versus-chain-length graphs and browser-based three-dimensional visualization of generated polymer structures, allowing users to inspect both predicted values and the corresponding molecular representation within the same interface.

Platform objective:  to provide fast, accessible, and reproducible polymer property predictions through a unified online environment for real-time prediction and interactive visualization.

2. Computational Workflow

The AdaptDelivery workflow integrates molecular modeling, molecular dynamics simulations, structural descriptor extraction, machine learning model development, and web-based deployment within a unified computational framework.

Representative synthetic and natural polymer systems are first constructed and solvated in water. Molecular dynamics simulations are then performed to capture polymer conformational behavior and polymer–solvent interactions over time.

The resulting molecular trajectories are analyzed to calculate two key structural descriptors: the radius of gyration (Rg) and the solvent-accessible surface area (SASA). These values are used to generate the datasets required for training and evaluating the artificial neural network models.

After training and validation, the predictive models are integrated into the AdaptDelivery online platform. Users can select a polymer and chain length to obtain real-time property predictions, interactive graphs, and a browser-based three-dimensional representation of the generated polymer structure.

Workflow:  Polymer construction β†’ Molecular dynamics simulation β†’ Structural descriptor extraction β†’ ANN training and validation β†’ Real-time prediction β†’ 2D and 3D visualization

3. Supported Polymer Systems

The current version of the AdaptDelivery platform supports both synthetic and natural polymer systems. The polymers are organized through a hierarchical selection workflow based on Polymer Class, Polymer Type, and Polymer Name.

Synthetic polymers

The synthetic polymer dataset includes representative systems from several chemical families, including acrylamides, amines, amides, imines, ketones, olefins, oxides, and vinyl-based polymers.

The synthetic polymers currently supported by the AdaptDelivery framework are: Polyethylene (PE), Polyethylene oxide (PEO), Polyethyleneimine (PEI), Polyacrylamide (PAM), Poly(allyamine), isotactic R configuration (PAAMR), Poly(N-isopropylacrylamide), isotactic R configuration (PNIPAAMR), and Poly(vinyl methyl ketone), isotactic R configuration (PVMKR).

Natural polymers

The natural polymer class currently includes representative lignin-based units: the Guaiacyl unit (GUAI), the p-Hydroxyphenyl unit (PHP), and the Syringyl unit (SYR), modeled within the CHARMM force-field framework (Vermaas et al. 2019).

Polymer selection hierarchy:  Polymer Class β†’ Polymer Type β†’ Polymer Name

4. Molecular Dynamics Simulations

Molecular dynamics simulations are used to investigate the conformational behavior of polymer chains in aqueous environments and to generate the reference data required for machine learning model development.

The molecular topology and parameter files for each polymer are generated using CHARMM-GUI (Choi et al. 2021; Jo et al. 2008). The resulting polymer systems are subsequently simulated with NAMD 3.0 (Phillips et al. 2020) using the CHARMM36 all-atom force field (Huang and MacKerell 2013; Vanommeslaeghe et al. 2009).

Each polymer system is first equilibrated for 3 ns, followed by a 200 ns production simulation using a 2 fs integration timestep. The simulations are performed at a temperature of 300 K and a pressure of 1 atm.

Temperature is controlled using a Langevin thermostat with a damping coefficient of 1 psβˆ’1, while pressure is maintained using a Langevin piston. Bonds involving hydrogen atoms are constrained with the SHAKE algorithm. Long-range electrostatic interactions are treated using the Particle Mesh Ewald method with a grid spacing of 1 Γ… and a cutoff distance of 12 Γ….

Periodic boundary conditions are applied to mimic an infinite system. During the simulations, the energy and temperature of each system are monitored to ensure stability and support the generation of high-quality, reproducible data.

Simulation protocol:  CHARMM-GUI preparation β†’ NAMD 3.0 β†’ CHARMM36 all-atom force field β†’ 3 ns equilibration β†’ 200 ns production run β†’ 2 fs timestep β†’ 300 K β†’ 1 atm

5. Radius of Gyration

The radius of gyration, denoted by Rg, is a fundamental descriptor of polymer size and conformational organization. It quantifies the spatial distribution of the polymer atoms around the center of mass and is commonly used to characterize the compactness and spatial extent of a polymer chain in solution:

\[ R_g = \sqrt{ \frac{1}{N} \sum_{i=1}^{N} \left\|\mathbf{r}_i-\mathbf{r}_{\mathrm{cm}}\right\|^2 }. \]

Here, N is the number of selected polymer atoms, ri is the position vector of atom i, and rcm is the position vector of the polymer center of mass:

\[ \mathbf{r}_{\mathrm{cm}} = \frac{ \sum_{i=1}^{N} m_i\mathbf{r}_i }{ \sum_{i=1}^{N} m_i }. \]

For each molecular dynamics trajectory frame, the center of mass is calculated using atomic masses, after which the radius of gyration is evaluated as the root-mean-square distance of the selected atoms from that center.

Calculating Rg independently for every trajectory frame makes it possible to analyze conformational fluctuations and the temporal evolution of polymer size. In general, lower Rg values correspond to more compact conformations, whereas higher values indicate a more spatially extended polymer chain.

Platform interpretation:  AdaptDelivery predicts the radius of gyration as a function of the selected polymer system and monomer number.

6. Solvent-Accessible Surface Area

The solvent-accessible surface area, denoted by SASA, quantifies the extent of the polymer surface that is accessible to solvent molecules, typically water. It is an important structural descriptor for characterizing polymer–solvent interactions, hydration, folding, conformational changes, and aggregation behavior:

\[ A_{\mathrm{SASA}} = \sum_i A_i. \]

Here, Ai represents the solvent-accessible surface area contribution associated with atom i.

SASA is calculated by rolling a spherical solvent probe over the molecular surface. In the AdaptDelivery molecular dynamics workflow, a probe radius of 1.4 Γ… is used to represent a water molecule:

\[ r_{\mathrm{probe}} = 1.4\ \text{Γ…}. \]

The calculation is performed independently for each frame of the molecular dynamics trajectory using the VMD software package (Humphrey et al. 1996). The selected polymer atoms are analyzed while excluding the surrounding solvent, and the accessible surface contributions are summed to obtain the total SASA value.

This frame-by-frame analysis makes it possible to monitor changes in solvent exposure and polymer conformation over time. In general, larger SASA values indicate greater exposure of the polymer surface to the surrounding solvent.

Platform interpretation:  AdaptDelivery predicts the solvent-accessible surface area as a function of the selected polymer system and monomer number.

7. Structural Dataset Generation

Reference values for the radius of gyration and solvent-accessible surface area were calculated from molecular dynamics simulations for representative polymer chain lengths:

\[ n \in \left\{5,\,8,\,10,\,12,\,16,\,25,\,50,\,100,\,150,\,200\right\}. \]

To construct a continuous relationship between monomer number and polymer property, the tabulated molecular dynamics values were fitted using a power-law function:

\[ f(x;a,b)=a\,x^b. \]

Here, x represents the monomer number, while a and b are fitting parameters determined through nonlinear least-squares fitting using the curve_fit function from the SciPy computational library (Virtanen et al. 2020).

The fitted power-law relationship was evaluated for monomer numbers up to 1000 and used to generate 20,000 pairs consisting of a monomer number and the corresponding polymer property value.

Dataset generation workflow:  Molecular dynamics reference values β†’ Power-law fitting β†’ Evaluation up to 1000 monomers β†’ 20,000 property pairs β†’ Training, validation and test datasets

8. Machine Learning Models

AdaptDelivery uses artificial neural networks to learn the relationship between polymer chain length and structural properties derived from molecular dynamics simulations. Once trained, the models can generate predictions on a substantially shorter timescale than running a new molecular dynamics simulation for each input.

The monomer number is used as the input variable, while the output of the neural network is the predicted polymer property:

\[ x \xrightarrow{\mathrm{ANN}} \widehat{P}(x). \]

Separate models are used for the prediction of radius of gyration (Rg) and solvent-accessible surface area (SASA), because these properties lie on different numerical scales.

Radius of gyration model

For the Rg prediction model, the study reports an architecture containing an input layer, one fully connected hidden layer with a channel size of 16, and an output layer that returns the predicted radius of gyration. Several alternative architectures were also investigated.

Solvent-accessible surface area model

The SASA prediction model uses a deeper fully connected architecture with three hidden layers containing 64, 32, and 16 channels, respectively.

Nonlinearity is introduced through the rectified linear unit activation function:

\[ \operatorname{ReLU}(z)=\max(0,z). \]

Here, z represents the linear activation of a neuron, obtained from the weighted input values and the corresponding bias term. The ReLU activation introduces nonlinearity into the neural network models.

Current modeling strategy:  Separate property-specific ANN models β†’ Rg prediction β†’ SASA prediction β†’ Real-time deployment within AdaptDelivery

9. Training Procedure

The artificial neural networks are trained using supervised learning. For each input monomer number, the corresponding reference value of the polymer property is known from the generated dataset and is used as the target value during model optimization.

Before training, the complete dataset is randomly shuffled and divided into three subsets:

Dataset partition:  80% training set β†’ 10% validation set β†’ 10% test set

The training set is used to optimize the neural-network parameters. The validation set is evaluated after each training cycle to monitor model performance and detect possible overfitting, while the test set is reserved for the final evaluation of the fully trained model.

Training samples are processed in mini-batches of 128. For each mini-batch, the difference between the predicted property values and the corresponding reference values is calculated using the mean absolute error:

\[ E_1 = \frac{1}{m} \sum_{i=1}^{m} \left| P_i^{\mathrm{predicted}} - P_i^{\mathrm{label}} \right|. \]

Here, m = 128 is the mini-batch size, Pipredicted is the property value returned by the neural network, and Pilabel is the corresponding reference value.

The objective of the training process is to minimize this error with respect to the network parameters. Optimization is performed using the Adam algorithm (Kingma and Ba 2014), which applies stochastic gradient-based updates. The error gradient is propagated backward through the network, and the model parameters are updated iteratively.

Training cycle:  Mini-batch input β†’ Forward propagation β†’ Property prediction β†’ Error calculation β†’ Backpropagation β†’ Parameter update

10. Model Performance and Validation

The predictive performance of the models was evaluated using the test set reserved during dataset partitioning. Model accuracy was assessed separately for each polymer system and for each predicted structural property.

The principal evaluation metric was the mean absolute error (MAE), which measures the average absolute difference between the predicted property values and the corresponding reference values:

\[ \operatorname{MAE} = \frac{1}{N} \sum_{i=1}^{N} \left| P_i^{\mathrm{predicted}} - P_i^{\mathrm{label}} \right|. \]

In addition to MAE, the mean predicted test-set value (MPTS) and the mean label test-set value (MLTS) were calculated to provide a direct comparison between model predictions and expected values.

The results showed strong agreement between predicted and reference values across the investigated synthetic and natural polymer systems. The reported MAE values were small relative to the corresponding mean property values, supporting the predictive accuracy of the models.

As an additional comparison with previously published molecular dynamics results, the model predicted a radius of gyration of approximately 8.97 Γ… for PEO with 18 monomers, compared with the 9.53 Γ… value reported by (Lee et al. 2008).

Evaluation strategy:  Test set β†’ Polymer-specific evaluation β†’ Property-specific MAE β†’ Comparison of predicted and reference values

11. Online Platform Usage

AdaptDelivery provides a structured workflow for generating polymer property predictions directly through the web interface. Users begin by opening the Properties Generator and selecting the desired prediction module: Radius of Gyration (Rg) or Solvent-Accessible Surface Area (SASA).

The prediction and visualization modules can be accessed either through a registered user account or through the Guest access option, which allows the platform to be used without mandatory registration. Registered users can additionally access Prediction History, where previous predictions can be reviewed and deleted.

The polymer system is selected through a hierarchical input structure based on Polymer Class, Polymer Type, and Polymer Name. After a polymer is selected, the platform displays its full name and corresponding two-dimensional chemical representation, providing visual confirmation before the prediction is generated.

The user then enters a single integer representing the polymer chain length or monomer number. The currently accepted prediction range is:

\[ 1 \leq n \leq 1000. \]

This interval corresponds to the monomer-number range covered by the machine learning datasets. Although the integrated PDB generators can produce longer polymer chains, predictions are restricted to the same range used for model training and evaluation.

After the input is submitted, the platform automatically standardizes and scales the numerical data according to the requirements of the trained model. The corresponding artificial neural network is then executed to generate the requested polymer property prediction.

The resulting output includes the predicted property value, an interactive property-versus-chain-length graph, and a browser-based three-dimensional visualization of the generated polymer structure.

The platform backend is implemented in Python using Flask, while pre-trained neural network models are loaded for inference using PyTorch. Interactive property plots are generated using Plotly.js and dynamically generated PDB structures are rendered directly in the browser using 3Dmol.js. PostgreSQL supports user registration, authentication and platform data management.

User workflow:  Select prediction module β†’ Select polymer system β†’ Enter monomer number β†’ Generate prediction β†’ Explore numerical, graphical, and structural results

12. Current Limitations

The current machine learning models support polymer chain lengths from 1 to 1000 monomers. This restriction corresponds to the interval covered by the generated training, validation, and test datasets. Predictions are therefore restricted to the same range used for model training and evaluation.

The current version of AdaptDelivery predicts two structural descriptors: radius of gyration (Rg) and solvent-accessible surface area (SASA). Separate property-specific neural networks are used for these outputs.

The three-dimensional structure displayed in the browser is a dynamically generated, strictly linear polymer chain in PDB format. It is produced on demand using the custom Python-based polymer generator library and rendered directly in the browser.

Predictions are available only for polymer systems that have already been included in the scientific workflow. Adding a new polymer requires the generation of molecular dynamics data, extraction of structural descriptors, construction of the corresponding dataset, model training and validation, and integration into the platform backend.

Current scope:  Supported polymer systems β†’ 1–1000 monomers β†’ Rg and SASA prediction β†’ Linear PDB visualization β†’ Property-specific ANN models

13. Citation and Availability

AdaptDelivery was developed as an online polymer informatics platform for the real-time prediction and visualization of structural polymer properties. The platform integrates molecular dynamics simulations, data-driven modeling, artificial neural networks, and interactive web technologies within a unified computational environment.

The scientific work associated with the AdaptDelivery platform is cited below. Researchers who use AdaptDelivery in scientific studies, presentations, teaching materials, or publications are encouraged to cite this work.

Recommended citation:
Parajdi L.G., Toth I., Farcas A.-A., Balint Z., Kiraly A., Farcas A. AdaptDelivery: A Polymer Informatics Platform for Predicting Structural Properties of Polymers via Machine Learning. Manuscript submitted for publication, 2026.

According to the associated study, all data generated or analyzed during the research are included in the paper mentioned above.

14. References

  • Choi Y.K., Park S.J., Park S., Kim S., Kern N.R., Lee J., Im W. CHARMM-GUI Polymer Builder for modeling and simulation of synthetic polymers. Journal of Chemical Theory and Computation, 17(4):2431–2443, 2021.
  • Huang J., MacKerell A.D. CHARMM36 all-atom additive protein force field: validation based on comparison to NMR data. Journal of Computational Chemistry, 34(25):2135–2145, 2013.
  • Humphrey W., Dalke A., Schulten K. VMD: Visual Molecular Dynamics. Journal of Molecular Graphics, 14(1):33–38, 1996.
  • Jo S., Kim T., Iyer V.G., Im W. CHARMM-GUI: a web-based graphical user interface for CHARMM. Journal of Computational Chemistry, 29(11):1859–1865, 2008.
  • Kingma D.P., Ba J. Adam: a method for stochastic optimization. arXiv preprint, arXiv:1412.6980, 2014.
  • Lee H., Venable R.M., MacKerell A.D., Pastor R.W. Molecular dynamics studies of polyethylene oxide and polyethylene glycol: hydrodynamic radius and shape anisotropy. Biophysical Journal, 95(4):1590–1599, 2008.
  • Phillips J.C., Hardy D.J., Maia J.D.C., et al. Scalable molecular dynamics on CPU and GPU architectures with NAMD. The Journal of Chemical Physics, 153(4):044105, 2020.
  • Vanommeslaeghe K., Hatcher E., Acharya C., et al. CHARMM general force field: A force field for drug-like molecules compatible with the CHARMM all-atom additive biological force fields. Journal of Computational Chemistry, 31(4):671–690, 2009.
  • Vermaas J.V., Petridis L., Ralph J., Crowley M.F., Beckham G.T. Systematic parameterization of lignin for the CHARMM force field. Green Chemistry, 21(1):109–122, 2019.
  • Virtanen P., Gommers R., Oliphant T.E., et al. SciPy 1.0: fundamental algorithms for scientific computing in Python. Nature Methods, 17(3):261–272, 2020.

πŸ“© Contact

Begin predicting polymer properties today and accelerate your nanoplatform development. For inquiries or collaboration opportunities, contact us at contact@adaptdelivery.org.