Advancing antimicrobial polymer development: a novel database and accelerated design via machine learning

Abstract

The rapid growth of resistant microorganisms has caused serious public health issues and poses great pressure on the current healthcare system. In this environment, the necessity of new antibiotic materials is even more prominent. Antimicrobial polymers are a class of polymers that have the ability to eradicate or impede the proliferation of microbes on their surfaces or within their surrounding environment. The mechanism of action of antibacterial polymers also makes them a perfect fit for medical devices. Despite great growing needs, the design of new antibacterial polymers with desired antimicrobial properties is still challenging. In this work, we present the first open-source database for antimicrobial polymers which consists of 489 entries, with 177 unique polymers exhibiting diverse structures and properties. Multiple predictive models were also designed and trained to classify the antimicrobial properties of these polymers. The best-performing random forest model showed an average accuracy of 86.7% in a 10-fold cross-validation test. We also developed multiple guiding pipelines for the design of novel antimicrobial polymers.

---

1 Introduction

Infections and diseases caused by different harmful microorganisms including bacteria, fungi and viruses have increased significantly over the past 20 years, especially in the fields of medical devices, hospital surfaces, medicine, food packaging and dental equipment.^1–5 The emergence of drug-resistant microorganisms caused by the overuse of antibiotics complicates the situation and increases the pressure on the public health system.⁶ Currently existing antibiotics are unable to effectively kill drug-resistant microorganisms, which makes the treatment of many diseases difficult. Meanwhile, traditional disinfectants are also not effective in eliminating the proliferation of resistant microorganisms spread in the environment.⁷ As a result, new effective disinfectants are urgently needed.

In this context, antimicrobial polymers (AMPs) have gained attention from both academia and industry.^8–15 AMPs are a class of polymers that have the ability to kill or inhibit the growth of microbes on their surface or within their surrounding environment.¹⁶ AMPs have antibacterial properties themselves or can be used as a matrix filled with biocides. According to the different mechanisms of action, AMPs can be divided into three categories: (a) polymers with antibacterial properties themselves, (b) polymers with antibacterial properties obtained through chemical modification, and (c) polymers that do not have antibacterial properties themselves but can carry biocidal polymers.¹⁷ Due to their antimicrobial properties, AMPs are considered one of the primary candidates for new antibiotics. Unlike traditional low-molecular antibacterial agents, AMPs have no toxicity to the surrounding environment and have a longer service life.¹⁸ More importantly, AMPs destroy the membranes of microorganisms on the surface through electrostatic interactions, the hydrophobic effect and the chelate effect, and thus are less likely to cause resistance of microorganisms.¹⁹

Although there is a tremendous need for AMPs from industry and the medical system, the design of new AMPs with desired properties is challenging. This is mainly because the design of new polymers is often guided by intuitive and experimental experience. Such a trial-and-error method is time and labour intensive, while unable to stably produce products with certain target properties.^20–22 Moreover, polymers exhibit a higher degree of unpredictability in their structure–property relationships compared to small molecule compounds, primarily attributable to their larger molecular mass and intricate structures.²³

One possible solution is the employment of machine learning (ML) algorithms. In the last ten years, a rapidly increasing number of ML studies have been reported in materials science such as antimicrobial peptides,^24,25 ceramic materials,^26–28 and nanomaterials.^29–31 These applications have proved the ability of ML to provide precise prediction of materials properties, generate structures with desired properties, and thus accelerate material innovation. ML has also had successful achievements in the field of polymer design. Significant achievements have been made for diverse polymer structures and properties such as polymer electrolytes,³² polymers with high thermal conductivity³³ or charge transfer properties,³⁴ polymer–protein hybrids,³⁵ copolymers³⁶ and polymer-blend materials.³⁷ Polymers with the desired band gap, glass transition temperature, dielectric constant and other physicochemical properties have also been identified and synthesised with the aid of ML models.^38–43 However, the number of ML applications in the field of AMP is very limited despite the market demand and innovation capabilities.⁴⁴

ML is a data-driven method, and the lack of data is the main cause of the slow development of AMP using ML. Existing open-source polymer databases focus on the physical and chemical properties of polymers and no such AMP database has been built for this purpose. To address this problem, we have collected experimental AMP data from multiple resources which can be a critical foundation for the acceleration of AMP design with ML.

In this article, we will introduce this new AMP database and present our ML study focusing on the quantitative structure–activity relationship of AMPs.

2 Database

To ensure the quality of the database, we only collected experimental AMP data reported in peer-reviewed papers. A total of 489 data entries were collected, including 177 unique polymer structures. The properties of these polymers were summarized in a table format to include the following information: polymer structure, molecular weight (M_w), bacteria inhibitory function measurement, and type of bacteria. The data are presented in the ESI† and some polymer structures in the compiled database are demonstrated in Fig. 1.

Fig. 1 The chemical structure of eight polymer representatives of the database.

Molecular weight (M_w) refers to the average mass of the polymer molecules in a sample. It is a crucial parameter in polymer science and engineering because it directly affects the properties and behavior of polymers. In our database, the M_w values range from 0.19 to 4461.90 kg mol⁻¹, with the majority of the values being smaller than 50 kg mol⁻¹ as illustrated in Fig. 2 which shows the distribution of our data in terms of polymer M_w. In a few cases, the molecules have a very small number of repeating units, resulting in low M_w. These small molecules are part of a series where the number of repeating units gradually increases to a large value.

Fig. 2 The molecular weight count of polymers in the compiled database.

It is noted that there are many different ways of reporting antimicrobial performance. In general, the experiments evaluate the antimicrobial performance of the materials by comparing the differences in colonies before and after the application of the antimicrobial agent. Table 1 summarizes the different types of measurements reported in the database.

Table 1 Explanation of four different measurements and their proportion in the database

Measurement	Explanation	Proportion
MIC(μg mL^−1)	Minimum inhibitory concentration (MIC) is defined as the lowest concentration of the antibiotic agent to inhibit the growth of 100% of the targeted microorganism	76.2%
% of bacteria killed	The percentage of the target microorganism that is killed	14.5%
MIC90(μg mL^−1)	Minimum inhibitory concentration (MIC) is defined as the lowest concentration of the antibiotic agent to inhibit the growth of 90% of the targeted microorganism	5.3%
Log reduction		4%
	Here ci/cj is the initial/final concentration. Log reduction of 1 equals 90%, 2 equals 99%, 3 equals 99.9% and so on

---

The bacterial target of antimicrobial polymers is crucial. AMPs have different antimicrobial capabilities against different microorganisms. Based on different purposes, AMPs can be designed to resist most microorganisms or have a high killing rate for certain types of microorganisms. In our database, the most commonly tested microorganisms, such as E. coli, S. aureus and B. subtilis, are included. The total number of data entries in the database against different targeting microorganisms are shown in Fig. 3.

Fig. 3 The total number of data entries against different targets in the compiled database.

3 Method

Our three-phase workflow for developing predictive ML models is summarized in Fig. 4.

Fig. 4 The three-phase machine learning workflow utilized in the study.

3.1 Preprocessing of the dataset

Polymers with an MIC value of 200 μg mL⁻¹ or less are categorized as active. Based on this threshold, the dataset was divided into a subset of 365 active polymers and a subset of 124 non-active polymers.

3.2 Polymer structure representation

As the molecular weight of AMPs is high and their structures contain many repeating units, it is unnecessary to calculate the descriptors for the whole molecule. It is a common approach to represent the whole polymer molecule using its repeating unit (monomer) and the end groups. In this study, the structure of each polymer is represented by its monomer and the two end groups and saved as a .mol file, a file format that stores information about the atoms, bonds and other information of a molecule. The corresponding files can be accessed through the supporting documents.

3.3 Descriptor generation

For simplicity, we used the simplified molecular-input line-entry system (SMILES) to represent the repeating units and end groups.⁴⁵ It is worth noting that the structures are C-capped to allow for the calculation of descriptors.

In total, over 5666 polymer descriptors including constitutional, topological, geometrical and other descriptors have been computed based on the SMILES strings of the AMPs using the AlvaDesc software.⁴⁶ For polymers with multiple repeating units, the descriptors for each repeating unit were calculated independently and summed up based on their weight in the polymer. Additionally, for both end groups, 50 constitutional descriptors were calculated separately and in tandem with the descriptors for the repeating units. The molecular weight M_w (kg mol⁻¹) reported in the original papers was also added as one descriptor. As a result, a total of 5767 (5666 + 50 × 2 + 1) descriptors were calculated for each polymer.

3.4 Feature selection

To reduce the dimensionality of the dataset, firstly, descriptors with over 90% of the entries as zeroes were removed. These features were regarded as carrying insignificant information. Descriptors with very high correlations were then excluded. The correlation analysis was done by calculating the correlation matrix of features, and for each feature pair with a correlation of 0.95 or higher, the second feature is removed. Highly correlated features tend to present similar information and provide non-additional insights. To further reduce the number of features, we innovatively applied 7 feature selection algorithms that cover the main methods used in past studies, including variance selection, correlation with the target, information gain, L1-regularized logistic regression, least absolute shrinkage and selection operator (LASSO), random forest and recursive feature elimination (RFE). These algorithms cover the three main categories of feature selection methods: filter, embedded and wrapper methods. Each algorithm was set to select features to a specific number. Features selected by most of the algorithms were finally selected.

3.5 Model implementation

Several classification ML models were implemented using the scikit-learn⁴⁷ libraries and Python 3.10.6: logistic regression, decision tree, random forest, support vector machine, K-nearest neighbor, Naïve Bayes and gradient boosting.⁴⁷

• Logistic regression⁴⁸ is a statistical model that estimates the probability of a binary outcome based on one or more predictor variables, utilizing a logistic function to make predictions.

• Decision tree⁴⁹ is a hierarchical structure where nodes represent features, branches represent decisions, and leaves represent outcomes, serving as a versatile tool for both classification and regression tasks.

• Random forest⁵⁰ is an ensemble learning method that constructs multiple decision trees during training and aggregates their predictions, reducing overfitting and improving accuracy by combining the results of multiple models.

• Support vector machine⁵¹ (SVM) is a supervised learning algorithm that finds the optimal hyperplane to separate data into different classes, relying on a subset of training data points called support vectors to define the decision boundary.

• K-nearest neighbor (KNN)⁵² is a non-parametric algorithm that classifies new data points based on the majority class of their K closest neighbors, making predictions by identifying the most similar instances in the feature space.

• Naïve Bayes⁵³ is a probabilistic classifier that applies Bayes’ theorem with the assumption of feature independence, making it computationally efficient and particularly well-suited for text classification tasks.

• Gradient boosting⁵⁴ is an ensemble technique where weak learners, typically decision trees, are added sequentially to correct errors made by the previous models, resulting in a powerful predictive model with high accuracy.

• eXtreme gradient boosting⁵⁵ (XGBoost) is a powerful, scalable machine learning library for gradient boosting, optimized for speed and performance. It supports various objective functions, efficient handling of missing data, and parallel processing. XGBoost is widely used for classification and regression tasks, especially for small datasets.

• Artificial neural networks (ANNs) consist of interconnected layers of neurons whose weights are adjusted during training to learn patterns from data. Through multiple training runs and evaluations, ANNs can achieve high accuracy and adaptability across diverse applications.⁵³

Models obtained using these methods were then optimized through a Grid Search algorithm, together with a 10-fold cross-validation (GSCV) algorithm for evaluation.

4 Results and discussion

4.1 Identification of significant features

From the large number of descriptors, zero-like and highly correlated features were excluded, resulting in a set of 494 features remaining. The number of descriptors selected by each of the seven feature selection algorithms, as shown in Table 2, was set to 100 ± 10. The main reason for this approach is to avoid the bias of each algorithm.

Table 2 Seven algorithms applied for feature selection

Algorithm	Number of features selected	Category
Variance selection	100	Filter method
Correlation with target	100	Filter method
Information gain	95	Filter method
L1-regularized logistic regression	103	Embedded method
LASSO	100	Filter method
Radom forest	94	Filter method
RFE	100	Wrapper method

---

Subsequently, features selected by the seven algorithms were divided into 7 groups based on the numbers of algorithms identifying them as relevant descriptors, as shown in Fig. 5. Three descriptors chosen by all seven algorithms were ‘P_VSA_m_2’, ‘P_VSA_ppp_L’ and ‘P_VSA_charge_7’. Seven descriptors chosen by six different algorithms were ‘P_VSA_LogP_2’, ‘P_VSA_LogP_6’, ‘P_VSA_MR_6’, ‘P_VSA_s_4’, ‘P_VSA_charge_9’, ‘CATS 2D_05_LL’ and ‘AMR’.

Fig. 5 The count of descriptors selected by different feature selection algorithms.

In this study, 32 descriptors selected by five or more algorithms were used to form the final set of input descriptors for the ML models. Most of the descriptors are highly independent of each other and their relationships are visualized in a heat map showing the correlation coefficients (see Fig. 6).

Fig. 6 The heatmap of the correlation coefficient between descriptors selected. Red color indicates the high correlation coefficient values, and blue low correlation coefficient values.

Herein, we provide a detailed explanation of the 32 selected descriptors.

4.1.1 M _w (kg mol⁻¹). Polymers are materials with large molecular weights and, not surprisingly, this molecular weight descriptor was found to play an important role in affecting the antimicrobial properties of such materials.

4.1.2 P_VSA-like. P_VSA-like descriptors are based on the sum of atomic contributions to the van der Waals surface area, based on the atoms having a property in a defined range of values:⁵⁶

g_ij = min{max{|R_i − R_j|, b_ij}, (R_i + R_j)} (2)

b_ij = r_ij− c_ij (3)

In eqn (1), VSA_i is the van der Waals surface area of the i-th atom, R_i is the atomic van der Waals radius of the atom i, nAT is the number of atoms, and a_ij represents the elements of the adjacent matrix.

g _ij is the max value of |R_i − R_j| and b_ij and (R_i + R_j).

b _ij denotes the ideal length of the bond formed by atoms i and j.

r _ij is the reference bond length and c_ij is a correlation term related to the bond multiplicity: 0 for single, 0.1 for aromatic, 0.2 for double, and 0.3 for triple bonds. R_i and c_ij are pre-defined values, which can be accessed online at https://www.talete.mi.it/help/dproperties_help/index.html?p_vsa_like_descriptors.htm.

Based on equations 1, 2, and 3, P_VSA-like descriptors can be computed using eqn (4) in which δ is a Kronecker delta function which is equal to one for atoms with property values in the specified range, and zero otherwise. w_i denotes one of the weighting schemes including: ‘log P’ for log P (octanol/water), ‘MR’ for molar refractivity, ‘m’ for atomic mass, ‘e’ for Sanderson electronegativity, ‘ppp’ for potential pharmacophore points and ‘charge’ for partial charge. k is the bin number, indicating a pre-defined range.

4.1.3 O%. This descriptor calculates the relative occurrence frequency of the O atom.

4.1.4 ALOGP. ALOGP is the Ghose–Crippen–Viswanadhan octanol–water partition coefficient, defined as:
where n_i is the atom of type i and h_i is the corresponding hydrophobicity contribution.⁵⁷

4.1.5 CATS 2D. CATS represents Chemically Advanced Template Search descriptors. CATS 2D descriptors are a particular class of autocorrelation descriptors and are defined as:
where u and v represent two atom types, δ(i; u), δ(j; v) and δ(d_ij; k) are three Kronecker delta functions equal to one if atom i is of type u, atom j is of type v, and the topological distance d_ij is equal to k, respectively, zero otherwise. In CATS 2D descriptors, the atom-type definition is related to the concept of potential pharmacophore points (PPPs). PPP is a generalized atom type defined considering some physicochemical aspects. CATS 2D descriptors are calculated based on the topological distance varying from 0 to 9 and any atom of the molecule can be assigned to none, one, or two atom types (DD, DA, DP, DN, DL, AA, AP, AN, AL, PP, PN, PL, NN, NL, and LL) resulting in a vector of 150 frequencies.

4.1.6 DELS. The molecular electrotopological variation (DELS) is calculated as:

where nSK is the number of non-H atoms. The DELS index could be considered as a measure of the total charge transfer in the molecule.⁵⁸

4.1.7 J_B(i). This descriptor is a Balaban-like index from the Burden matrix weighted by ionization potential. It is also highly correlated with multiple 2D matrix-based descriptors extracted from matrixes such as adjacency, topological distance, Laplace and Chi.

4.1.8 nCsp3_endgroup1. nCsp3 stands for the number of sp³ hybridized carbons. This descriptor indicates the count of sp³ hybridized carbons in one of the end groups (end group 1). This descriptor is highly correlated with the following: Se (sum of atomic Sanderson electronegativities, scaled on a carbon atom), Si (sum of first ionization potentials, scaled on a carbon atom), nAT (total number of atoms), RBN (number of rotatable bonds) and nH (number of hydrogen atoms).

4.2 Modelling results

To ensure the acquisition of generalized ML models, ten separate runs were performed for each ML algorithm and the average of accuracies was used to evaluate the performance of the obtained ML models. The variance of ten accuracies was also calculated as a metric to quantify the spread or dispersion of the accuracy values. Six out of seven models achieved an average accuracy of around 85% and only the Naïve Bayes model had a low accuracy of 67.2%, as shown in Table 3. The receiver operating characteristic (ROC) curves for all algorithms are provided in Fig. S2 of the ESI.† For the dataset in this work, the Random Forest method is among those that achieved the highest average accuracy while also having the lowest variance, indicating its excellent predictive capability and a steady performance across different data sets.

Table 3 Statistical results of different ML models for antimicrobial polymers

ML model	Average accuracy	Variance
Logistic regression	0.850	0.0015
Decision tree	0.858	0.0012
Random forest	0.877	0.0005
Support vector machine	0.875	0.0010
K-nearest neighbor	0.836	0.0026
Naïve Bayes	0.666	0.0047
Gradient boosting	0.863	0.0009
XGBoost	0.867	0.0008
ANN	0.854	0.0015

---

4.3 Antimicrobial design principles

Since the trained models can provide accurate predictions and reveal the relationships between descriptors and polymer properties, we analysed the modelling results and determined guidelines for the design of new AMPs.

4.3.1 Logistic regression. Logistic regression is a statistical model used for binary classification tasks, predicting the probability of an event occurring based on input features. The weights in logistic regression represent the strength and direction of the relationship between each feature and the log-odds of the event happening, allowing us to interpret the impact of each feature on the antimicrobial properties.

As shown in Fig. 7, a number of descriptors such as ‘P_VSA_charge_2’, ‘P_VSA_charge_7’, and ‘P_VSA_ppp_hal’ showed a negative impact on the antimicrobial property. When the value of these descriptors decreases, the antimicrobial property tends to be stronger. On the other side, the increase of the descriptors with positive impact such as ‘J_B(i)’, ‘M_w (kg mol⁻¹)’, ‘AMR’, and ‘P_VSA_i_2’ would contribute to higher antimicrobial activity.

Fig. 7 The impact of each descriptor on antimicrobial properties.

4.3.2 Decision tree. The decision tree model is a popular machine learning algorithm that mimics human decision-making by partitioning the data into subsets based on feature values. It iteratively selects the best features to split the data and creates a tree-like structure of decisions, making it interpretable and easy to understand. The decision tree obtained from this study is shown in Fig. S1.† Based on the decision boundary and classification accuracy, we can summarize the conditions that the descriptors need to meet to result in a higher probability of active antimicrobial activities. It is worth noting that the guidelines consider multiple descriptors and will only work when the conditions for all of these descriptors are met.

Fig. 8–10 illustrate different parts of the decision tree (subtrees 1 to 3) separately for clarification, and the classification is highlighted.

Fig. 8 Decision boundary from a decision tree (subtree 1).

Fig. 9 Decision boundary from a decision tree (subtree 2).

Fig. 10 Decision boundary from a decision tree (subtree 3).

In the decision tree, each node corresponds to a set of polymers. When the majority of polymers in each node (more than 50%) are active, the node has a blue color, otherwise yellow. The number at the bottom of the node shows the exact number of active/non-active polymers in the node. For example, in the first node of subtree 1 (Fig. 8), there are 329 active polymers out of a total number of 441 polymers. The number at the top of the node shows the condition for the descriptor values. Polymers satisfying the condition will be transferred to the lower left node of the subtree and otherwise to the lower right one. When most of the polymers in a node are classified correctly, there will be no subsequent conditions. Following the different branches of conditions, we can find different sets of conditions for active and non-active polymers.

As shown in subtree 1 (Fig. 8), there are 111 active polymers that have a P_VSA_charge_14 of no more than 5.495 and a P_VSA_charge_7 of no more than 24.214. In contrast, polymers with a P_VSA_charge_7 greater than 24.214 and a P_VSA_ppp_L greater than 162.2 are suggested to be non-active and should be avoided in the design process. Subtree 1 also shows that polymers with a P_VSA_charge_14 of no more than 5.495, but P_VSA_charge_7 and P_VSA_pp_L greater than 24.214 and 162.2, respectively, are more likely to be non-active.

Going down the subtree, it can also be seen that most of the polymers (54/61 or 88.5%) with P_VSA_pp_L no greater than 162.2 are classified as active and all of the remaining 10 polymers with M_w values (kg mol⁻¹) smaller than 3.4 are also active. Polymers with larger M_w values could still be active if they have CATS2D_04_LL smaller than 5.0 and Wap smaller than 188.

Following the same pattern, we can also find conditions for designing active polymers using subtrees 2 and 3. We summarise all the conditions for active polymers or non-active polymers in Table 4.

Table 4 Conditions for designing antimicrobially active and non-active polymers using the decision tree algorithm

Conditions for active polymers	Condition 1: P_VSA_charge_14 ≦ 5.495 & P_VSA_charge_7 ≦ 24.214
	Condition 2: P_VSA_charge_14 ≦ 5.495 & P_VSA_charge_7 > 24.214 & P_VSA_pp_L ≦ 162.2 & Mw (kg mol^−1) ≦ 3.4
	Condition 3: P_VSA_charge_14 ≦ 5.495 & P_VSA_charge_7 > 24.214 & P_VSA_pp_L ≦ 162.2 & Mw (kg mol^−1) > 3.4 & CATS2D_04_LL ≦ 5.0 & Wap ≦ 188
	Condition 4: P_VSA_charge_14 > 5.495 & CATS2D_05_LL ≦ 3.5 & P_VSA_i_2 > 160.468 & P_VSA_charge_7 ≦ 112.738 & Wap ≦ 1816
	Condition 5: P_VSA_charge_14 > 5.495 & CATS2D_05_LL ≦ 3.5 & P_VSA_e_2 ≦ 183.036 & P_VSA_charge_7 ≦ 86.956
	Condition 6: P_VSA_charge_14 > 5.495 & CATS2D_05_LL ≦ 3.5 & P_VSA_e_2 > 183.036 & O% ≦ 9.997 & DELS ≦ 30.107 & Mw (kg mol^−1) ≦ 50.7
Conditions for non-active polymers	Condition 1: P_VSA_charge_14 ≦ 5.495 & P_VSA_charge_7 > 24.214 & P_VSA_pp_L > 162.2
	Condition 2: P_VSA_charge_14 > 5.495 & CATS2D_05_LL ≦ 3.5 & P_VSA_i_2 ≦ 160.468 & P_VSA_LogP_2 ≦ 19.47
	Condition 3: P_VSA_charge_14 > 5.495 & CATS2D_05_LL ≦ 3.5 & P_VSA_i_2 ≦ 160.468 & P_VSA_LogP_2 > 19.47 & P_VSA_e_2 > 96.552 & DELS > 47.128
	Condition 4: P_VSA_charge_14 > 5.495 & CATS2D_05_LL ≦ 3.5 & P_VSA_i_2 > 160.468 & P_VSA_charge_7 > 112.738
	Condition 5: P_VSA_charge_14 > 5.495 & CATS2D_05_LL ≦ 3.5 & P_VSA_e_2 ≦ 183.036 & P_VSA_charge_7 > 86.956
	Condition 6: P_VSA_charge_14 > 5.495 & CATS2D_05_LL ≦ 3.5 & P_VSA_e_2 > 183.036 & O% > 9.997 & P_VSA_charge_14 ≦ 10.787

---

4.3.3. Random forest. Random Forest is an ensemble learning method that constructs a multitude of decision trees during training and outputs the mode of the classes for classification or the average prediction for regression.⁵⁹ Different from a single decision tree, Random Forest mitigates overfitting and enhances accuracy by combining predictions from multiple decision trees trained on bootstrapped samples and using random subsets of features for each tree. Based on the best-performing Random Forest model, the influence of each descriptor on the polymers’ antimicrobial properties can be quantified and ranked. These top-ranking descriptors could be critical factors in the antimicrobial properties and can assist with the future design of AMPs. In Table 5, we present seven most relevant descriptors that could be used as a guideline for AMP design. The full list can be found in Table S1 of the ESI.†

Table 5 Top ranking descriptors by the Random Forest algorithm and their feature importance values

Descriptor	Importance	Description	Type
M w (kg mol^−1)	0.121	Molecular weight	Constitutional indices
P_VSA_LogP_2	0.060	P_VSA-like on log P, bin 2	P_VSA-like descriptors
O%	0.043	Percentage of O atoms	Constitutional indices
ALOGP	0.036	Ghose–Crippen octanol–water partition coefficient (log P)	Molecular properties
CATS 2D_05_LL	0.034	CATS 2D lipophilic–lipophilic at lag 05	Pharmacophore descriptors
DELS	0.033	Molecular electrotopological variation	Topological indices
J_B(i)	0.013	Balaban-like index from Burden matrix weighted by the ionization potential	2D matrix-based descriptors

---

4.4 Evaluation of the design principles

As presented above, the relationship between the polymer structure and the antimicrobial activity is complex and multidimensional. Inverse engineering requires careful consideration of various factors. In this section, we will illustrate how the design principles obtained using different ML techniques can be reflected in the compiled database using four polymers, two of which have very low MIC values (high activity) (polymers 1 and 2, Fig. 11) and 2 with very high MIC values (low activity) (polymers 3 and 4, Fig. 13).

Fig. 11 Chemical structures and descriptor values of two most active polymers in the database (polymer 1 (top) and polymer 2 (bottom)).

As shown in Fig. 12, polymer 1 satisfies condition 3 of the decision tree guidelines while polymer 2 satisfies condition 4. These are classified by the model to be antimicrobially active.

Fig. 12 Decision tree active conditions that are met by polymers 1 (top) and 2 (bottom).

Fig. 13 Chemical structures and descriptor values of two least active polymers in the database (polymer 3 (top) and polymer 4 (bottom)).

It should also be noted that although the structure of polymer 2 is similar to that of polymer 1, polymer 2 has higher M_w and O% but lower CATS2D_05_LL, SHED_LL, P_VSA_LogP_2 and ALOGP. Logistic regression and random forest algorithm ranking suggests that polymer 2 has a higher MIC value than polymer 1.

Similarly, as shown in Fig. 14, polymer 3 satisfies the non-active condition 1, while polymer 4 satisfies the non-active condition 2. They are classified by the decision tree model to be non-active polymers.

Fig. 14 Decision tree non-active conditions that are met by polymers 3 (top) and 4 (bottom).

5 Conclusion

We have compiled and presented the first AMP database with experimental data from multiple peer-reviewed articles. This database is expected to contribute to the rapid, data-driven development of AMP for both research advancement and industry applications. We applied innovative algorithms and identified 32 significant descriptors based on 7 different feature selection approaches. We also proposed multiple ML models with high predictive accuracy (around 85%) for antimicrobial properties. Furthermore, we determined the impact and importance of descriptors affecting the antimicrobial properties of polymers. A guideline was proposed for the design of highly active AMPs. We hope this database and the discovery of influential descriptors can provide a solid and informative foundation for researchers in the field to explore new AMPs in the future.

Author contributions

Yuankai Zhao: conceptualization, investigation, data curation, visualization, coding, model development, and writing – original draft. Shadi Houshyar: reviewing & editing and supervision. Roger J Mulder: data validation, reviewing & editing, and supervision. Daniel J. Eyckens: data validation and reviewing & editing. Tu C. Le: conceptualization, data validation, writing, review & editing, project administration, and supervision.

Data availability

The authors declare that the data supporting the findings of this study are available within the paper and its ESI files.† Should any raw data files be needed in another format they are available from the corresponding author upon reasonable request.

Conflicts of interest

The authors declare no conflicts of interest.

Acknowledgements

Yuankai Zhao acknowledges the CSIRO-RMIT scholarship program.

References

1. T. U. Berendonk, C. M. Manaia, C. Merlin, D. Fatta-Kassinos, E. Cytryn, F. Walsh, H. Bürgmann, H. Sørum, M. Norström, M.-N. Pons, N. Kreuzinger, P. Huovinen, S. Stefani, T. Schwartz, V. Kisand, F. Baquero and J. L. Martinez, Tackling antibiotic resistance: the environmental framework, Nat. Rev. Microbiol., 2015, 13, 310–317.
2. F. Siedenbiedel and J. C. Tiller, Antimicrobial polymers in solution and on surfaces: overview and functional principles, Polymers, 2012, 4, 46–71.
3. H. M. Lode, Clinical impact of antibiotic-resistant Gram-positive pathogens, Clin. Microbiol. Infect., 2009, 15, 212–217.
4. M. J. Garcia-Fernandez, L. Martinez-Calvo, J. Ruiz, M. R. Wertheimer, A. Concheiro and C. Alvarez-Lorenzo, Loading and Release of Drugs from Oxygen–rich Plasma Polymer Coatings, Plasma Processes Polym., 2012, 9, 540–549.
5. S. I. Hay, P. C. Rao, C. Dolecek, N. P. J. Day, A. Stergachis, A. D. Lopez and C. J. L. Murray, Measuring and mapping the global burden of antimicrobial resistance, BMC Med., 2018, 16, 78.
6. A. Muñoz-Bonilla and M. Fernández-García, Polymeric materials with antimicrobial activity, Prog. Polym. Sci., 2012, 37, 281–339.
7. H. Takahashi, I. Sovadinova, K. Yasuhara, S. Vemparala, G. A. Caputo and K. Kuroda, Biomimetic antimicrobial polymers—Design, characterization, antimicrobial, and novel applications, Wiley Interdiscip. Rev.: Nanomed. Nanobiotechnol., 2023, 15, e1866.
8. P. Pham, S. Oliver and C. Boyer, Design of antimicrobial polymers, Macromol. Chem. Phys., 2022, 2200226.
9. M. Haktaniyan and M. Bradley, Polymers showing intrinsic antimicrobial activity, Chem. Soc. Rev., 2022, 51, 8584–8611.
10. K.-S. Huang, C.-H. Yang, S.-L. Huang, C.-Y. Chen, Y.-Y. Lu and Y.-S. Lin, Recent advances in antimicrobial polymers: a mini-review, Int. J. Mol. Sci., 2016, 17, 1578.
11. M. R. E. Santos, A. C. Fonseca, P. V. Mendonça, R. Branco, A. C. Serra, P. V. Morais and J. F. J. Coelho, Recent developments in antimicrobial polymers: a review, Materials, 2016, 9, 599.
12. M. M. Konai, B. Bhattacharjee, S. Ghosh and J. Haldar, Recent progress in polymer research to tackle infections and antimicrobial resistance, Biomacromolecules, 2018, 19, 1888–1917.
13. J. R. Smith and D. A. Lamprou, Polymer coatings for biomedical applications: a review, Trans. IMF, 2014, 92, 9–19.
14. T. Potta, Z. Zhen, T. S. P. Grandhi, M. D. Christensen, J. Ramos, C. M. Breneman and K. Rege, Discovery of antibiotics-derived polymers for gene delivery using combinatorial synthesis and cheminformatics modeling, Biomaterials, 2014, 35, 1977–1988.
15. G. N. Tew, R. W. Scott, M. L. Klein and W. F. DeGrado, De novo design of antimicrobial polymers, foldamers, and small molecules: from discovery to practical applications, Acc. Chem. Res., 2010, 43, 30–39.
16. H. B. Kocer, I. Cerkez, S. D. Worley, R. M. Broughton and T. S. Huang, Polymeric Antimicrobial N -Halamine Epoxides, ACS Appl. Mater. Interfaces, 2011, 3, 2845–2850.
17. A. Jain, L. S. Duvvuri, S. Farah, N. Beyth, A. J. Domb and W. Khan, Antimicrobial Polymers, Adv. Healthcare Mater., 2014, 3, 1969–1985.
18. E.-R. Kenawy, S. D. Worley and R. Broughton, The Chemistry and Applications of Antimicrobial Polymers: A State-of-the-Art Review, Biomacromolecules, 2007, 8, 1359–1384.
19. L. Timofeeva and N. Kleshcheva, Antimicrobial polymers: mechanism of action, factors of activity, and applications, Appl. Microbiol. Biotechnol., 2011, 89, 475–492.
20. D. J. Audus and J. J. de Pablo, Polymer Informatics: Opportunities and challenges, ACS Macro Lett., 2017, 6, 1078–1082.
21. L. Chen, G. Pilania, R. Batra, T. D. Huan, C. Kim, C. Kuenneth and R. Ramprasad, Polymer informatics: current status and critical next steps, Mater. Sci. Eng., R, 2021, 144, 100595.
22. Danishuddin and A. U. Khan, Descriptors and their selection methods in QSAR analysis: paradigm for drug design, Drug Discovery Today, 2016, 21, 1291–1302.
23. G. E. Wnek, Structure–Property Relationships of Small Organic Molecules as a Prelude to the Teaching of Polymer Science, J. Chem. Educ., 2017, 94, 1647–1654.
24. G. J. Gabriel, A. E. Madkour, J. M. Dabkowski, C. F. Nelson, K. Nüsslein and G. N. Tew, Synthetic mimic of antimicrobial peptide with nonmembrane-disrupting antibacterial properties, Biomacromolecules, 2008, 9, 2980–2983.
25. J. Xu, F. Li, A. Leier, D. Xiang, H.-H. Shen, T. T. Marquez Lago, J. Li, D.-J. Yu and J. Song, Comprehensive assessment of machine learning-based methods for predicting antimicrobial peptides, Briefings Bioinf., 2021, 22, 1–22.
26. Y.-C. Hsu, C.-H. Yu and M. J. Buehler, Using Deep Learning to Predict Fracture Patterns in Crystalline Solids, Matter, 2020, 3, 197–211.
27. R. W. K. Li, T. W. Chow and J. P. Matinlinna, Ceramic dental biomaterials and CAD/CAM technology: State of the art, J. Prosthodontic Res., 2014, 58, 208–216.
28. J. R. Jones, Review of bioactive glass: From Hench to hybrids, Acta Biomater., 2013, 9, 4457–4486.
29. M. J. Kratochvil, A. J. Seymour, T. L. Li, S. P. Paşca, C. J. Kuo and S. C. Heilshorn, Engineered materials for organoid systems, Nat. Rev. Mater., 2019, 4, 606–622.
30. Q. Liu, S. Zheng, K. Ye, J. He, Y. Shen, S. Cui, J. Huang, Y. Gu and J. Ding, Cell migration regulated by RGD nanospacing and enhanced under moderate cell adhesion on biomaterials, Biomaterials, 2020, 263, 120327.
31. S. Mitragotri and J. Lahann, Physical approaches to biomaterial design, Nat. Mater., 2009, 8, 15–23.
32. Y. Wang, T. Xie, A. France-Lanord, A. Berkley, J. A. Johnson, Y. Shao-Horn and J. C. Grossman, Toward Designing Highly Conductive Polymer Electrolytes by Machine Learning Assisted Coarse-Grained Molecular Dynamics, Chem. Mater., 2020, 32, 4144–4151.
33. S. Wu, Y. Kondo, M. Kakimoto, B. Yang, H. Yamada, I. Kuwajima, G. Lambard, K. Hongo, Y. Xu, J. Shiomi, C. Schick, J. Morikawa and R. Yoshida, Machine-learning-assisted discovery of polymers with high thermal conductivity using a molecular design algorithm, npj Comput. Mater., 2019, 5, 66.
34. S. Ye, N. Meftahi, I. Lyskov, T. Tian, R. Whitfield, S. Kumar, A. J. Christofferson, D. A. Winkler, C.-J. Shih, S. Russo, J.-C. Leroux and Y. Bao, Machine learning-assisted exploration of a versatile polymer platform with charge transfer-dependent full-color emission, Chem, 2023, 9, 924–947.
35. M. J. Tamasi, R. A. Patel, C. H. Borca, S. Kosuri, H. Mugnier, R. Upadhya, N. S. Murthy, M. A. Webb and A. J. Gormley, Machine Learning on a Robotic Platform for the Design of Polymer–Protein Hybrids, Adv. Mater., 2022, 34, 2201809.
36. C. Kuenneth, W. Schertzer and R. Ramprasad, Copolymer Informatics with Multitask Deep Neural Networks, Macromolecules, 2021, 54, 5957–5961.
37. Z. Liang, Z. Li, S. Zhou, Y. Sun, J. Yuan and C. Zhang, Machine-learning exploration of polymer compatibility, Cell Rep. Phys. Sci., 2022, 3, 100931.
38. Y. Zhao, R. J. Mulder, S. Houshyar and T. C. Le, A review on the application of molecular descriptors and machine learning in polymer design, Polym. Chem., 2023, 14, 3325–3346.
39. H. Doan Tran, C. Kim, L. Chen, A. Chandrasekaran, R. Batra, S. Venkatram, D. Kamal, J. P. Lightstone, R. Gurnani, P. Shetty, M. Ramprasad, J. Laws, M. Shelton and R. Ramprasad, Machine-learning predictions of polymer properties with Polymer Genome, J. Appl. Phys., 2020, 128, 171104.
40. A. Mannodi-Kanakkithodi, G. Pilania, T. D. Huan, T. Lookman and R. Ramprasad, Machine learning strategy for accelerated design of polymer dielectrics, Sci. Rep., 2016, 6, 20952.
41. J. Liang, S. Xu, L. Hu, Y. Zhao and X. Zhu, Machine-learning-assisted low dielectric constant polymer discovery, Mater. Chem. Front., 2021, 5, 3823–3829.
42. L. A. Miccio and G. A. Schwartz, From chemical structure to quantitative polymer properties prediction through convolutional neural networks, Polymer, 2020, 193, 122341.
43. P. R. Duchowicz, S. E. Fioressi, D. E. Bacelo, L. M. Saavedra, A. P. Toropova and A. A. Toropov, QSPR studies on refractive indices of structurally heterogeneous polymers, Chemom. Intell. Lab. Syst., 2015, 140, 86–91.
44. S. M. McDonald, E. K. Augustine, Q. Lanners, C. Rudin, L. Catherine Brinson and M. L. Becker, Applied machine learning as a driver for polymeric biomaterials design, Nat. Commun., 2023, 14, 4838.
45. D. Weininger, SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules, J. Chem. Inf. Comput. Sci., 1988, 28, 31–36.
46. A. Mauri, in Ecotoxicological QSARs, ed. K. Roy, Springer US, New York, NY, 2020, pp. 801–820.
47. F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss and V. Dubourg, Scikit-learn: Machine learning in Python, J. Mach. Learn. Res., 2011, 12, 2825–2830.
48. Z. Bursac, C. H. Gauss, D. K. Williams and D. W. Hosmer, Purposeful selection of variables in logistic regression, Source Code Biol. Med., 2008, 3, 17.
49. S. R. Safavian and D. Landgrebe, A survey of decision tree classifier methodology, IEEE Trans. Syst. Man Cybern., 1991, 21, 660–674.
50. G. Biau and E. Scornet, A random forest guided tour, TEST, 2016, 25, 197–227.
51. C. Cortes and V. Vapnik, Support-vector networks, Mach. Learn., 1995, 20, 273–297.
52. T. Cover and P. Hart, Nearest neighbor pattern classification, IEEE Trans. Inf. Theory, 1967, 13, 21–27.
53. S. J. Russell and P. Norvig, Artificial intelligence: a modern approach, Prentice Hall/Pearson Education, Upper Saddle River, N.J., 2nd edn, 2003.
54. J. H. Friedman, Greedy Function Approximation: A Gradient Boosting Machine, Ann. Stat., 2001, 29, 1189–1232.
55. J. H. Friedman, Stochastic gradient boosting, Comput. Stat. Data Anal., 2002, 38, 367–378.
56. P. Labute, A widely applicable set of descriptors, J. Mol. Graphics Modell., 2000, 18, 464–477.
57. A. K. Ghose, V. N. Viswanadhan and J. J. Wendoloski, Prediction of Hydrophobic (Lipophilic) Properties of Small Organic Molecules Using Fragmental Methods: An Analysis of ALOGP and CLOGP Methods, J. Phys. Chem. A, 1998, 102, 3762–3772.
58. P. Gramatica, M. Corradi and V. Consonni, Modelling and prediction of soil sorption coefficients of non-ionic organic pesticides by molecular descriptors, Chemosphere, 2000, 41, 763–777.
59. L. Breiman, Random Forests, Mach. Learn., 2001, 45, 5–32.

---

Footnote

† Electronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d4py00736k

---