Enzymes for Synthesis and Degradation of Bio‐Based Polyesters: Automatic In Silico Screening and Experimental Validation

Abstract

This study intends to contribute to the rational eco‐design of tailored new sustainable polymers with the aid of computational methods able to generate and screen new biocatalysts in silico for the synthesis or degradation of polyesters. Automatic computational workflows can be used for the rapid selection of enzymes that efficiently catalyze hydrolysis and/or synthesis of bio‐based polyesters. Proteins and protein–ligand interactions were studied by integrating specific software for molecular dynamics simulations (Gromacs) and docking (AutoDock) into a workflow which also included some geometric evaluation of productive binding poses. Such a pipeline, based on successive calls to freely available and well‐established software tools, was implemented in modeFRONTIER, a package equipped with a graphical user interface in which nodes, representing calls to external programs, are connected to form the sequential workflow. A series of bio‐based monomers and enzymes was screened, and the computational results were correlated with experimental data obtained for the synthesis direction as well as with the data available for hydrolysis. The pipeline, by prioritizing and ranking molecules, enables the identification of optimal catalysts for both synthetic and hydrolytic reactions, emphasizing efficient prioritization and ranking strategies. The method was validated on a pool of 6 proteins and 20 substrates, and the predicted enzyme–substrate matches were subsequently compared with both the experimental synthesis and hydrolysis data, thereby confirming the reliability of the predictions. Overall, the results indicate that the hydrolytic and synthetic activity of enzymes toward polyesters can be rationally predicted by integrating different computational tools, providing a cost‐effective alternative to time‐consuming experimental screening.

1. Introduction

The identification of the optimum enzyme for a given substrate remains an open problem in the field of biocatalysis because of the complexity of biological systems and the vast array of potential enzyme–substrate interactions. This task is even more challenging when the substrates are polymers such as polyesters. Several factors contribute to the complexity of this problem. First, enzymes are highly specific, and slight variations in substrate structure or conformation can significantly affect the catalytic efficiency. Additionally, the vast diversity of enzymes in nature, coupled with the huge number of potential substrates, makes exhaustive experimental exploration impractical.

Enzymes are one of the proposed solutions for tackling the environmental impact of plastic. They can overcome the limitations of conventional chemical strategies by catalyzing targeted modifications, synthesis, or degradation of polymers under highly selective and mild conditions [1]. Conversely, identifying the optimum biocatalyst for given substrate and vice versa are key aspects in the development of polymers, especially renewable polyesters. These bio‐based polymers are among the solutions offered by bioeconomy to mitigate the upstream impact of plastics and they can underpin a fruitful circular economy [2]. Several bio‐based polyesters have been commercialized in the last decades, such as polylactic acid (PLA), polyhydroxy‐alkanoates (PHAs) and, more recently, polybutylene succinate (PBS), but the range of monomers derived from renewable resources has expanded considerably in the last years [3].

Aliphatic polyesters emerge as a compelling choice for bio‐based plastics production, offering versatility by being synthesized with various building blocks of biological origin. Furthermore, the enzymatic route to aliphatic polyesters has paved the way for new materials with tailored innovative characteristics, which can find applications in medical and biomedical engineering [4], drug delivery systems, and artificial implants. Enzymes overcome the limitations of conventional chemical strategies by catalyzing, under highly selective and mild conditions, targeted modifications, synthesis, or degradation of polymers and, most importantly, bio‐based polymers [3]. This highlights the transformative impact of biocatalysis in enhancing daily lives through advanced and sustainable solutions for the entire life cycle of polymers and materials [5]. For instance, lipases and cutinases were successfully used for in vitro polycondensation of bio‐based diacids and polyols, leading to biodegradable polyesters with controlled structures [6, 7, 8]. On the other hand, cutinases were reported to be able to modify or degrade different synthetic polyesters [9].

However, the experimental screening of new enzymes could be very expensive and time‐consuming. In silico design and virtual screening of new enzymes can accelerate the optimization process by exploiting appropriate computational and bioinformatics tools, thus reducing the number of experiments [10, 11]. In recent years, new computational tools have been successfully used to select promising enzymes based on virtual screening of substrate–biocatalyst pairs [12, 13], opening up new perspectives on the correlation between the structural features of a polymer and its biodegradability.

Novel computational approaches to addressing these challenges can benefit from integration with those structure‐based virtual screening methods, which are already well established as drug design tools and are designed to solve the problem of finding the optimum ligand for a given receptor, namely, a protein binding site. Given the large number of possibilities, virtual screening techniques are typically expected to identify promising candidates through a large number of docking attempts [14]. Such methods include a large set of options: structure‐based screening integrates protein structure data (when available); pharmacophore‐based screening identifies essential ligand features, while ligand‐based screening generally utilizes molecular descriptors and relies on the whole structure of known active compounds; finally, fragment‐based screening explores small molecular fragments, whereas shape‐based screening matches three‐dimensional structures. In addition, allosteric site screening can be developed based on the previously mentioned methods, but selecting an alternative/allosteric binding site as target. These methods collectively enhance the drug discovery process by accelerating candidate identification, offering diverse strategies to tackle the challenges of computing molecular interactions [15]. Similarly, virtual screening tools have been employed to identify potential substrates in chemoenzymatic reactions [16], also with the aid of QM/MM calculations [17].

It is important to stress that the problem of enzyme/substrate recognition has a few common points with typical of drug design, i.e., the selectivity of the candidate drug versus other off targets. In drug design, the ligand (very often an inhibitor) should fill the catalytic pocket to maximize existing binding patterns whereas, in the case of the enzymatic biodegradation of polymers, the substrate is oriented to favor the proceeding of the reaction, while maintaining transient, sufficiently weak interactions, with the target. Early‐stage screening often does not comprehensively evaluate all potential off‐target interactions, which is like the challenges faced in drug development. Both fields require the identification of specific molecular interactions, whether between enzymes and substrates or between drugs and targets, within a vast space of possible combinations. This highlights the critical need for thorough screening to optimize biocatalytic processes as well as drug efficacy.

Another important aspect to consider when modeling these systems is the environment in which the reactions is carried out. This aspect, not generally considered by docking methods, is appropriately addressed by molecular dynamics (MD) simulations, in which a system can be simulated along time in its environment at appropriate thermodynamic conditions. MD methods computationally simulate particle behavior using Newtonian mechanics and statistical mechanics. By calculating molecular movements, position, and by considering external factors, MD elucidates not only dynamics but also thermodynamic properties. For polymeric molecules, MD has successfully been used in molecular imprinting and simulation of polymer synthesis, revealing interactions and conformational changes between molecules. This makes it a powerful tool for understanding complex molecular systems [18].

The aim of the present work is to develop a virtual screening procedure to analyze and evaluate the ability of different hydrolases to interact and to accept aliphatic diesters, either for synthetic or degradative processes. To the best of our knowledge, this is the first time a virtual screening procedure is proposed in the framework of enzymatic catalysis for the synthesis of polymers.

Here, we capitalize on the above well‐established computational methods to fully automatize a screening pipeline for the identification of optimum substrate/catalyst matches. The main objectives are: (i) given a specific substrate, to be able to screen a set of biocatalysts, and rank them as optimum choices for both polymerization and hydrolysis; (ii) given a specific biocatalyst, to screen a set of substrates, and rank them to identify those likely to be recognized by the chosen enzyme, for catalyzing either polycondensation or hydrolysis and degradation.

The process was implemented in modeFRONTIER [19], a multidisciplinary and multiobjective optimization and automation software platform developed by ESTECO, designed for engineering and product design. ModeFRONTIER facilitates the integration of simulation tools, automates the execution of complex workflows, and performs optimization processes to enhance product performance, reduce costs, and accelerate the design process [20]. More specifically, the current workflow involves the implementation of the Substrate‐Imprinted Docking model, which was demonstrated to be a robust predictive method for modeling substrate specificity and enantioselectivity in lipases and esterases [21]. In the present work, this method was adapted and automatized for the evaluation of the specificity of hydrolases towards different bio‐based esters, which were used as ligands for the docking and were part of the MD simulations.

2. Methods

2.1. Protein Library Preparation

Crystallographic data. The crystal structure of the lipases Candida antarctica B (PDB identifier 1TCA/pdb_00001tca) and Thermomyces lanuginosus (PDB identifier 1DTE/pdb_00001dte) as well as the cutinases Humicola insolens (PDB identifier 4OYY/pdb_00004oyy), Fusarium solani (PDB identifier 1AGY/pdb_00001agy) and Thermobifida cellulosilytica (PDB identifier 5LUI/pdb_00005lui) were all retrieved from the Protein Data Bank [22] and were preprocessed by using Swiss‐PdbViewer [23]. All the lipases and cutinases were prepared by removing all the water molecules, the heteroatoms, any cocrystallized solvent or substrate, whereas polar hydrogens and partial charges were added to the structure using AutoDockTools (ADT version 1.5.6) [24].

2.2. Homology Modeling and MD

Since the crystal structure of the lipase from Pseudomonas stutzeri was not available on the Protein Data Bank, we built an homology model, by using the FASTA sequence (previously reported by Maraite et al. [25]) as input for the software SWISSMODEL and using 3D crystallographic structures from the PDB database as template (PDB identifier 1EX9/pdb_00001ex9), as detailed in the supplementary material. The resulting .pdb file was then used for the docking.

MD simulations were conducted to optimize the geometry obtained by homology modeling: for this task, we used the all‐atom optimized potentials for liquid simulations OPLS‐AA force field available in GROMACS [26]. MD simulations were run for 100 ns at 300K and the system [27] was enclosed within a cubic box filled with water molecules modeled through a 3‐site rigid water model (namely TIP3P). Energy minimization by the steepest descent method was performed with a force constant of 1000.0 kJ/mol/nm. Subsequently, the system was equilibrated, first by conducting under an NVT canonical ensemble (NVT, 0.2 ns) and then by using an isothermal–isobaric ensemble (NPT, 0.2 ns) to ensure stability. The 100 ns production simulation was conducted for each MD system with GPU acceleration using the CINECA Leonardo supercomputer (pre‐exascale supercomputer hosted by CINECA) [28]. The RMSD (root mean square deviation), RMSF (root mean square fluctuation), Rg (radius of gyration), SASA (aolvent accessible surface area), and H‐bond (hydrogen bonds) values were determined by gmx rms, gmx rmsf, gmx gyrate, gmx hbond, and gmx sasa programs available in GROMACS, respectively. The resulting structures were finally visualized by using PyMOL software [29].

2.3. Substrate Library Preparation

A total number of 20 substrate molecules (dimers) were generated with ChemDraw [30] and minimized using MOPAC [31] (with the AM1 semiempirical method) prior to docking.

For substrates preparation, the industrially produced bio‐based polyols and diacids were selected. In this regard we selected four diols/polyols and five diacids. Selected diols/polyols are monoethylenglykol (MEG), glycerol (GLY), 1,4‐butandiol (BDO), and sorbitol (SRB). Selected diacids are succinic acid (SA), acidipic acid (AA), itaconic acid (ITA), sebacic acid (SBA), and dodecandioic acid (DDA). Thus, 20 corresponding esters were generated by exhaustively combining the two lists.

This selection was based on the availability of experimental reference data and on the methodological limitations of docking‐based approaches, such as AutoDock, which are best suited for small to moderately flexible ligands.

2.4. Docking

Dockings were performed for 20 heterodimers over the unequilibrated enzyme structures of the six hydrolases (five experimental structures and one homology model). For each system, the docking box was centered on the catalytic site, by using a cube with 15 Å side, finely described by using a grid spacing of 0.250 Å. In order to guarantee a very high number of docking poses, in the AutoDock software (version 4.2) [24] we set the Lamarckian genetic algorithm (GA) population size to 1500 and the maximum number of evaluations to 25,000,000. Thus, 500 docking generations, each one composed of 250 docking poses, were provided. Each pose was evaluated geometrically to define whether it could be considered as a “productive pose,” by comparing it with the ideal situation. In a productive pose, the substrate assumes a near‐attack conformation (NAC) that is geometrically compatible with the productive attack of the catalytic serine of the hydrolase on the carbon atom of the acyl group of the ligands (Figure 1A) (either carboxylic acid leading to chain elongation, or ester causing the reverse of the reaction). The criteria used for the selection of the productive poses were identified by studying the geometry of several complexes: we set to 3.2 Å the maximum distance between nucleophilic serine oxygen and the carbonyl carbon of the ligand; to 90° the maximum value for the angle of attack of the serine oxygen and the carbonyl carbon (Figure 1B); and up to 3 Å the distance between the oxygen of the carbonyl and the hydrogens of the residues responsible for the stabilization of the oxyanion [32].

FIGURE 1.

Schematic illustration of the criteria used to select the productive poses (A). Sorbitol adipate docked in the catalytic site of CAL B (B).

By employing the specified productive pose criteria, a systematic ranking of enzymes for diverse substrates was facilitated, contributing to a comprehensive understanding of their substrate specificity and a theoretical catalytic efficiency.

2.5. MD

We placed each system in a cubic box with a water layer of minimum 1.0 nm. Each system was neutralized with Cl^‐ ions and we used OPLS‐AA force field and TIP3P water. Substrates topologies were generated through the LigParGen server [33]. We performed a steepest descent minimization of the protein chain to be stopped either when the maximum force was lower than 1000.0 kJ/mol/nm or after 50,000 minimization steps (performed with 0.005 kJ/mol energy step size, with Verlet cutoff scheme, short‐range electrostatic cutoff, and van der Waals cutoff of 1.0 nm). NVT and NPT equilibrations were performed for 100 ps constraining the protein backbone. NPT production runs of the unconstrained systems were run at 320K (modified Berendsen thermostat) and 1 atm (Parinello–Rahman pressure coupling) for 500 ns. The iteration time step was set to 2 fs with the leap‐frog integrator and LINCS constraint. We used periodic boundary conditions. Representative conformations were generated by clustering simulations snapshots with the single linkage method, cutoff 0.1 nm, as implemented in GROMACS. All the simulations and their analysis were run as implemented in the GROMACS package v. 2019.6 on the CINECA Marconi supercomputer (M100).

The minimization did not account for the presence of the surrounding solvent environment in which the protein and substrate are immersed. The minimized complexes have been stripped of the associated ligand. The protein structures thus obtained were used for a second run of docking experiments, with the respective ligands in the presence of which the equilibration took place. Docking parameters of this second run were unchanged.

2.6. Cheminformatics Molecular Description

For the molecular description, the VolSurf³ program was used (version 1.1.0b12) with the default options. The probes utilized for the calculation of the descriptors were as follows: DRY (hydrophobic probe), OH₂ (water probe), O (carbonyl probe), and N1 (nitrogen), as previously described [34]. Data from VolSurf³ were integrated with data from the other steps by means of heatmaps, which were generated with the ComplexHeatmap R/Bioconductor package [35, 36].

2.7. Enzymatic Synthesis of the Oligoesters

In a 100 mL flask, adipic acid and 1,4‐butanediol/glycerol were added at a molar ratio of 1:1. The reaction was started by the addition of lipases 50 U/mmol substrate. The reactions were performed as previously described [37] under vacuum in a rotary evaporator at 70°C and 70 mbar for 72 hr. At the end of reaction, the product was recovered using dichloromethane, and the solvent was removed by vacuum‐drying at 40°C and characterized by ¹H‐NMR spectroscopy.

3. Results and Discussion

3.1. Automated In Silico Screening Workflow

The pipeline we implemented aims to suggest, from a pool of candidates, which substrate/biocatalyst combinations are worth to focus for experimental validation and further testing. The dataset used included 6 hydrolases (3 lipases and 3 cutinases, Figure 2) and a set of 20 substrates, considered as dimers (formed by combining one molecule of diacid and one molecule of diol). The substrates were obtained by exhaustively combining four diols (MEG, GLY, BDO, and SRB) and five diacids (SA, AA, ITA, SBA and DDA), as detailed in Figure 3, where substrates are grouped by diols.

FIGURE 2.

The 3D structures of the lipases Candida antarctica B (PDB identifier 1TCA), Thermomyces lanuginosus (PDB identifier 1DTE), cutinases Humicola insolens (PDB identifier 4OYY), Thermobifida cellulosilytica (PDB identifier 5LUI), Fusarium solani (PDB identifier 1AGY), retrieved from the Protein Data Bank. For P. stutzeri we used a homology model, obtained as described in the Methods section.

FIGURE 3.

Chemical structures of the dimers considered as substrates for the computational studies.

The virtual screening pipeline receives as (required) input the three‐dimensional structures of both substrates (with their MOL representation) and biocatalysts (with their PDB structures, obtained either experimentally or built by homology modeling of P. stutzeri). It consists of subsequent calls to widely employed state‐of‐the‐art simulation packages. The overall pipeline was implemented in modeFRONTIER (Figure 5), which allows to connect different actions through a graphic user interface, ensuring the execution of the corresponding code and/or software with a proper and interconnected management of the input/output flow from the mentioned actions. The computational procedure is divided into three main steps and graphically summarized in Figure 4. It is also detailed below in dedicated subsections.

FIGURE 5.

Compressed view of the overall pipeline was here implemented in modeFRONTIER.

FIGURE 4.

Automatic computational procedure for in silico enzyme screening. From left to right: substrates and enzymes preparation, docking, MD, docking, and data analysis.

Figure 6 depicts the “Modelling_Step_1” subprocess.

FIGURE 6.

Extended view of the automatic computational procedure for in silico enzyme screening. The Modelling_Step_1 subprocess is depicted. From left to right: substrates and enzymes preparation, docking, postdocking, and substrate transfer to the second step.

3.1.1. Preliminary Docking

The first step of the pipeline is based on the docking software Autodock. Input files (.pdbqt) require 3D coordinates, partial charges, and atom types for all the atoms of the system. The process begins with the minimization of all the ligands and the processing of all the (macro)molecules to obtain their corresponding pdbqt files. Each binding pocket, described by a set of files reporting the energy maps obtained through AutoGrid, undergoes docking calculations for each ligand–macromolecule pair with AutoDock, by means of a specific bash script, included as supplementary materials. The generated poses (500 for each ligand–macromolecule pair) are then evaluated in terms of distance and angles, to assess whether each pose can be categorized as productive, as described below. At this point, a check point is included in which the limit of the productive poses was set up as 5%. More precisely, if in this step the number of productive poses was less than 5% of the total, the substrate would not be considered for the following steps of the flow. If the number of productive poses was above 5%, the pose closest to the defined NAC in the enzymatic active site was chosen, regardless of which reactive carbons were involved in the interaction with the catalytic serine of the enzymes.

3.1.2. Reactive Poses Definition

In the final part of step 1, reactive poses are defined as those that correspond to NAC and are compatible with productive attack by the catalytic serine residue on the carbon atom of the acyl group of the ligand in question. On the protein side, key catalytic residues are defined for each biocatalyst: Ser105 of CAL B and Hic, Ser146 of TLL, Ser109 of PSST, Ser120 of SOLC, and Ser131 of THCC). On the substrate side, the key element is either the carboxylic acid, which leads to chain elongation, or the ester, which causes the reaction to reverse (see Figure 1A). In the NAC, the length of the single covalent bond between the carbon atom of the acyl group and the oxygen atom is assumed to be between 1.43 and 2.15 Å, whereas the distance (d) between the serine oxygen and the acyl carbon atom is set to a maximum distance of 3.2 Å. Concerning the oxyanionic pocket, the oxyanion intermediate must be stabilized at a maximum distance of 3 Å from the NH hydrogen bond donor atom of the residues forming the pocket. Finally, the orientation of the acyl carbons of the ligands toward the serine oxygen (ϑ) must be at an acute angle [38].

3.1.3. MD Refinement

The substrate conformation with the lowest interaction energy, computed as the sum of the short‐range electrostatic and van der Waals components, was selected as the most suitable initial structure for the MD simulations. This energy‐based criterion ensured that the selected conformation reflected a thermodynamically favorable ligand–protein complex, thus maximizing the relevance of the subsequent dynamic sampling. Following this initial selection, a series of MD simulations were performed to explore the conformational landscape of the complex. Representative conformations were extracted from the resulting trajectories and subjected to clustering analysis. The centroid of the largest cluster, which corresponded to the most frequently sampled structural motif, was selected as the representative conformation for subsequent steps in the screening pipeline.

3.1.4. Final (Induced‐fit) Docking

A second docking experiment is run over the most representative conformation from MD simulation; this represents an induced‐fit docking step, where the macromolecular 3D structure used for the target has been conditioned by the ligand. The output of this final step (Checkpoint 1) includes the number of the productive poses and their associated docking scores. For each substrate, this output represents the results of the entire workflow. Taken together, the analysis of the data from the 20 substrates and 6 catalysts can lead to the identification of (i) the best substrate(s) for a given biocatalyst and (ii) the best biocatalyst(s) for a given substrate.

3.2. Data Mining to Identify Optimum Biocatalyst(s)

Data mining over the results of the screening pipeline allows the identification of the optimum biocatalyst for each of the studied substrates, either when considering its synthesis or hydrolysis.

In both cases, the key output of the screening pipeline is not only the number of NAC productive poses for each substrate/biocatalyst pair but also the lowest mean score among these poses.

The complexity and redundancy of the data (6 biocatalysts and 20 substrates, for both synthesis and hydrolysis pathways) is observable in Table 1, which summarizes the collected data in a tabular form. The data can be analyzed by scanning across the dataset to identify substrate‐biocatalyst pairs exhibiting the highest number of productive poses and/or the lowest mean docking scores within the synthesis or hydrolysis data blocks. However, the condensed data do not allow for comparative analysis of a given biocatalyst, as we only reported the best solutions. By reading across the rows, the most effective enzyme for either synthesis or hydrolysis can be identified for each substrate. From these results, we can extrapolate that the highest number of productive poses for synthesis was obtained with the lipase from CAL B, while the highest number of productive poses for hydrolysis was obtained with the cutinase T. cellulosilytica. Conversely, the lowest number of productive poses for both synthesis and hydrolysis was associated with the F. solani cutinase, indicating that is the least efficient biocatalyst among those evaluated.

TABLE 1.

For each substrate, only the best solutions are reported: in both sets (synthetic or hydrolytic path), AutoDock scores were computed only for the poses corresponding to the induced‐fit geometry of the biocatalyst, selected by MD simulations, whereas distances and angles used to assess the poses’ “productive character” were calculated taking as a reference the C atom of the free —COOH group.

Substrates Diacid Diol		Synthesis			Hydrolysis
		Enzyme	Productive poses	E * abs, kcal mol −1	Enzyme	Productive poses	E * abs , kcal mol −1
SA	MEG	PSST	297	−2.31	SOLC	107	−2.96
(C4)	GLY	PSST	349	−2.28	PSST	111	−2.01
	BDO	PSST	271	−2.15	HIC	441	−2.79
	SRB	CAL B	61	−6.57	THCC	98	−5.77
AA	MEG	CAL B	185	−3.08	THCC	159	−1.88
(C6)	GLY	CAL B	129	−2.01	PSST	138	−2.34
	BDO	CAL B	157	−3.34	THCC	100	−2.11
	SRB	CAL B	138	−1.14	SOLC	156	−5.81
SBA	MEG	CAL B	95	−1.17	THCC	147	−1.49
(C10)	GLY	PSST	349	−2.28	PSST	111	−2.02
	BDO	PSST	81	−2.58	PSST	127	−2.77
	SRB	CAL B	93	−5.26	THCC	67	−6.05
DDA	MEG	CAL B	159	−3.29	THCC	71	−2.91
(C12)	GLY	CAL B	158	−2.85	THCC	61	−2.38
	BDO	CAL B	184	−2.66	THCC	76	−7.89
	SRB	THCC	41	−3.42	THCC	48	−1.07
ITA	MEG	CAL B	93	−2.65	THCC	145	−2.43
C=C	GLY	CAL B	288	−2.31	CAL B	138	−2.05
	BDO	CAL B	161	−3.08	SOLC	246	−3.3
	SRB	THCC	64	−1.09	SOLC	36	−2.78

A graphical example, reporting all the data, is available for butylene adipate (BDOAA, Figure 7).

FIGURE 7.

Comparison of NAC productive poses obtained for butylene adipate (BDOAA) by analyzing the attack of the two different acyl groups (carboxylic acid = blue frame; ester group = dotted frame) using AutoDock. Red squares correspond to the average of the docking free energies of the corresponding productive poses.

We observed that, in terms of synthesis, lipase CAL B was the most efficient, both because of the number of productive poses (the highest of the series) and the predicted scores (the lowest in the series).

These results were expected, as the structural differences among the substrates are minimal, with the exception of those containing succinic acid, for which the lipase from PSST yielded the highest number of productive poses. We observed a preference of CAL B for the dicarboxylic acids with the longer aliphatic chains (such as AA, SBA, DDA), in agreement with the findings of McCabe et al., who experimentally studied the kinetics of different aliphatic dicarboxylic acids (with different length: C₄–C₇) as substrates of CAL B [39]. For substrates containing adipic acid (AA), the number of poses and the energy values are very similar for synthesis; meanwhile, for sebacic acid (SBA) we observed differences for the dimers with glycerol (GLY) and 1,4‐butandiol (BDO), with the lipase from PSST that gave the highest number of productive poses. It must be mentioned that also CAL B gave productive poses, but the values were slightly lower compared to PSST. For the C12 and unsaturated diacids derivatives, CAL B yielded the highest number of poses, except for sorbitol‐based substrates (Thc_cut was the most effective hydrolase). However, for sorbitol (SRB) derivatives, the number of productive poses was lower compared to the other diols/polyols, and these results are in concordance with experimental data previously reported for enzymatic synthesis of polyols derivatives or polyesters [40, 41].

Concerning the hydrolysis, cutinases resulted the most effective enzymes for 15 out of 20 substrates, with Thc_cut as the best in class, with the highest number of productive poses and lowest docking energy values.

3.3. Data Mining to Identify Optimum Substrate(s)

The screening pipeline also allows for the identification of the optimum substrate for both synthesis and hydrolysis, over a dataset of substrates. Indeed, the key output of the screening pipeline is the number of NAC productive poses for each substrate/biocatalyst combination, as exemplified for the BDO‐based oligoesters evaluated within the catalytic site of CAL B and P. stutzeri direction (Figure 8).

FIGURE 8.

Comparison of NAC productive poses obtained for different dimers containing BDO molecules using different enzymes by analyzing the attack of the two different acyl groups using AutoDock. Red squares correspond to the average of the docking free energies of the corresponding productive poses.

The results unveil distinctive substrate preferences among the evaluated lipases. Specifically, the lipase derived from PSST demonstrates optimal catalytic efficacy when coupled with butylene succinate. In parallel, the lipase sourced from CAL B exhibits pronounced productive poses with butylene dodecanoate, signifying a heightened affinity for this substrate.

These nuanced substrate‐specific interactions are integral to deciphering the catalytic dynamics of these enzymes and offer valuable insights into their biocatalytic potential. The identification of preferential substrates not only enhances our comprehension of the enzymatic mechanisms involved but also holds implications for the rational design and optimization of polyesterification reaction catalyzed by lipases.

3.4. Data Mining for a Wider Perspective

In a screening pipeline aimed to identify the best substrates for a given catalyst, the inclusion of quick preliminary steps allows to consider a larger set of molecules. The procedure presented here includes heavy calculations, especially the MD, which should be run only for those substrates with higher probability to achieve good final scores (either for synthesis or hydrolysis). With the aim of investigating whether there is any molecular feature of substrates able to induce in each biocatalyst a significant discrimination in terms of binding efficiency (either for hydrolysis or synthesis), we decided to integrate the full collected data with some molecular descriptors, commonly used in the field of chemoinformatics. Results are graphically reported as heatmaps (Figures 9 and 10), given the complexity of data.

FIGURE 9.

Heatmap of enzyme–substrate interactions relevant to aliphatic polyester synthesis. The upper panel shows the number of favorable docking poses between enzymes (rows) and substrates (columns), with color intensity indicating the number of poses, with minimum set to 0 (white) and maximum set to 300 (dark blue). The enzymes are categorized as either cutinases (THCC, Hic, and SOLC) or lipases (CAL B, PSST, and TLL). The ligands along the X‐axis are ranked by increasing calculated LogP, with values shown in the lower panel. Color intensity reflects the number of productive NAC poses, with darker colors indicating a greater number of poses. Clustering reveals groups of enzymes with similar interaction profiles, which suggests that they have comparable substrate recognition and specificity.

FIGURE 10.

Heatmap of enzyme–substrate interactions for polyester hydrolysis, illustrating docking results in terms of number of poses. Darker purple cells indicate higher pose counts, revealing promising enzyme–substrate combinations. Hierarchical clustering groups enzymes with similar binding behavior. This reveals distinct preferences between lipases and cutinases toward ester and carboxylic functional groups.

Preliminary analysis (reported as supplementary material) highlighted that both lipophilicity and hydrogen bonding are related to the difference between the substrates with different predisposition to the diverse catalysts (Figure S1).

To evaluate the potential of enzyme–ligand interactions in the context of the enzymatic synthesis of aliphatic polyesters, the output data from the automated workflow were analyzed and summarized in the heatmap shown in Figure 9. This heatmap integrates docking results ((reported as number of poses) with ligand hydrophobicity (expressed as LogP). Hierarchical clustering, conducted across each enzyme class (lipases or cutinases), reveals patterns of similarity among enzyme–ligand pairs. Figure 9 shows that several enzyme–substrate combinations, involving ligands such as BDODDA, BDOSA, MEGSA, GLYSA, GLYITA, GLYSBA, BDOITA, and BDOAA, exhibit a high number of docking poses for more than one enzyme (blue/deep purple cells in the heatmap), indicating frequent favorable geometry. These interactions suggest that these combinations are promising candidates for further experimental validation, particularly when using enzymes such as CAL B and PSST. Instead, we observed distinct binding profiles for certain ligands (e.g. MEGSBA, BDOSBA, and SRBSBA) which exhibited highly variable docking behavior across different enzymes, especially among the two classes (lipases and cutinases).

Clustering analysis over the enzymes reveals peculiar behaviors of TLL among the lipases and of Hic among the cutinases. From the analysis we might speculate that interactions are context‐dependent and influenced by variations in enzyme structure. Conversely, ligands such as BDOSA, MEGSA, and SRBITA demonstrate more consistent binding profiles across various enzyme contexts, highlighting their potential as versatile substrates for polyester synthesis.

Additionally, LogP analysis closely aligns with observed docking behavior, with more hydrophobic ligands generally on the right correlating with high docking poses. This suggests that hydrophobic interactions play a significant role in stabilizing ligand–enzyme complexes.

Overall, Figures 9 and 2S effectively consolidates the docking data, emphasizing the robust binding interactions and highlighting BDOSA, MEGSA, and GLYSA as the top substrate candidates due to their favorable docking properties and potential for efficient polyester synthesis.

Clear patterns emerged across the hydrolysis heatmap (Figure 10) regarding the binding preferences of lipases and cutinases towards specific substrates. Ligands such as GLYDDA, BDODDA, and BDOITA tend to have higher pose numbers, as indicated by the deeper purple coloring in the docking poses. This suggests enhanced docking stability and an increased capacity of enzymes to accommodate these ligands. Such ligands are promising candidates for biocatalytic applications, due to their consistent performance across multiple enzymes.

In contrast, ligands such as BDOSBA, SRBSBA, and SRBDDA demonstrate less favorable binding interactions, as indicated by lighter colors and fewer docking poses. This suggests structural incompatibility or lower number of interactions within the active sites of enzymes. These ligands may therefore be less suitable for immediate downstream experimental validation without further optimization.

Substrate‐specific interactions were also evident, with enzymes such as CAL B, THCC, and SOLC frequently showed robust interactions, particularly with ligands such as GLYDDA and BDOITA. These findings highlight the importance of enzyme–substrate specificity for enzyme engineering efforts focused on targeted substrate interactions.

Enzyme clustering reveals groups that exhibit similar docking behavior, which may indicate conserved binding pocket features or catalytic residues that facilitate ligand accommodation. Similarly, clustering among ligands suggests common structural or physicochemical features that influence their compatibility and binding behavior with enzymes.

Additionally, the LogP panel accompanying the docking data reveals a relationship between ligand hydrophobicity and docking interactions. Ligands with higher LogP values generally exhibit increased docking poses, indicating a slight correlation between the hypothetical abundance of hydrophobic interactions and enhanced enzyme–ligand affinity.

Overall, the hydrolysis heatmap clearly distinguishes between strong‐binding ligands (GLYDDA, BDODDA, and BDOITA) and less suitable candidates (BDOSBA, SRBSBA, and SRBDDA), thereby guiding the selection of compounds for further experimental investigations and enzyme optimization efforts.

3.5. Experimental Validation for the Enzymatic Synthesis

Selected results produced by the described pipeline were validated by comparison with previous experimental data. For the experimental validation of the theoretical data, one diacid (adipic acid) and 1,4 butandiol and glycerol were selected for the synthesis of the short oligoesters catalyzed by two lipases (from CAL B and PSSTc) and two cutinases (Hic and from SOLC). The schematic representation of the reaction is presented in Figure 11. The reactions were performed in a solvent‐less system, at temperature up to 70°C; the conversion of the raw materials was followed by ¹H‐NMR analysis and the medium molecular weights were determined based on ESI‐MS analysis. The results, presented in Table 2, indicate that the highest conversion of the adipic acid was obtained for lipase from CAL B. Moreover, the conversion values obtained in the presence of the other three hydrolases are in accordance with the ranking obtained from the developed automatic pipeline.

FIGURE 11.

Enzymatic route for the synthesis of the adipic acid‐based oligoesters.

TABLE 2.

The experimentally determined values of conversions of adipic acid and average molecular weight of the oligoesters obtained starting from adipic acid and 1,4 butandiol.

Sample		Lipase CAL B (BDOAAGLYAA		Lipase PSST BDOAAGLYAA		Cutinase Hic BDOAAGLYAA		Cutinase SOLC BDOAAGLYAA
		(BDOAA)n [37] (GLYAA)n		(BDOAA)n (GLYAA)n		(BDOAA)n (GLYAA)n		(BDOAA)n (GLYAA)n
Conversion AA, %		94	96	28	34	37	41	˜5	˜7
Mn**, g mol−1		683	912	298	514	220	620	n.d	n.d
Mw**, g mol−1		729	940	320	530	240	645	n.d	n.d
Đ**		1.06	1.03	1.07	1.03	1.09	1.04	n.d	n.d

Note: ** determined on the basis of ESI‐MS analysis.

3.6. Correlation between Docking Predictions and Experimental Oligoester Enzymatic Synthesis

The analysis of the results from the workflow, reported in the heatmaps (Figures 9 and 10), is strongly supported by the experimental conversion and molecular weight data, presented in Table 2. The lipase CAL B, which showed favorable docking profile with BDOAA and GLYAA ligands, also exhibited the highest adipic acid conversions (94%–96%), and produced oligomers with the highest molecular weight (M _w: 729–940 g/mol).

However, while the CAL B efficiency is in agreement with the computed data, predictions related to BDOAA and GLYAA highlighted THCC to be relevant for both BDOAA and GLYAA and for the cutinase from SOLC a variegate docking performance profile (low pose count for GLYAA but higher for BDOAA); for the former, no experimental data were collected, whereas for the latter we yielded very low conversions (∼5%–7%) and undetectable molecular weight (n.d.). Intermediate cases such as lipase from P. stutzeri and cutinase Hic—which had moderate docking profile—correspond to moderate experimental results, both in terms of conversions (28%–41%) and molecular weight (M _w: 240–645 g/mol). Overall, the general agreement between computational predictions and experimental synthesis results underlines the usefulness of the computational pipeline for the preliminary screening of enzyme candidates in polycondensation reactions involving diacids and diols. In the context of hydrolysis, the theoretical data remained unverified experimentally, due to the unavailability of commercial THCC. However, previous studies have demonstrated the effectiveness of this cutinase in hydrolyzing various polyester substrates [6]. These results support the plausibility of the computational predictions obtained in the present study. Consequently, conclusions regarding hydrolytic efficiency primarily rely on in silico results, supplemented by evidence from the literature which represents an inherent limitation of this study. Nevertheless, the consistency observed across docking scores, productive pose analysis, MD refinement, and clustering trends provides confidence in the proposed workflow's ability to identify and rank promising enzymes for polyester hydrolysis. Moreover, the automated pipeline is designed to be readily extended to experimental validation as soon as suitable enzyme preparations become available, thereby enabling future confirmation and refinement of the predicted hydrolytic activities.

This work uses the concept of rational eco‐design, referring to a data‐driven, computationally guided approach for prioritizing enzyme–substrate combinations relevant to the synthesis and degradation of bio‐based polyesters. Rather than providing a complete design solution, the proposed workflow enables for the systematic screening and ranking of biocatalysts and substrates, thereby supporting informed decision‐making in the early stages of sustainable polymer process development, to guide synthetic strategies. Overall, it represents a concrete and actionable contribution to rational eco‐design within the bioeconomy framework and can be readily extended to a broader substrate space and additional experimental validation campaigns.

The trends emerging from the automated screening pipeline strongly agree with previously reported experimental data on the enzymatic synthesis and degradation of aliphatic and bio‐based polyesters. In terms of synthesis, CAL B, a lipase from Candida antarctica, has been shown to consistently exhibit the highest number of productive NAC poses and favorable docking scores. It is well known as one of the most efficient biocatalysts for solvent‐free and bulk polycondensation of diacids and diols, including adipate‐, succinate‐, and sebacate‐based systems [6, 7, 8, 32, 39]. Experimental studies have repeatedly demonstrated CAL B's preference for longer aliphatic diacids, consistent with the increased number of productive poses predicted for adipic, sebacic, and dodecanedioic acid derivatives.

Similarly, the moderate performance predicted for P. stutzeri lipase is consistent with experimental observations reporting good activity toward short‐chain aliphatic substrates but lower efficiency than CAL B for bulk polyester synthesis [25, 37]. The reduced number of productive poses predicted for sorbitol‐based substrates mirrors the difficulties reported in experimental enzymatic polycondensation of polyol‐rich systems, where steric hindrance and excessive hydrogen bonding limit chain growth [40, 41].

In the hydrolysis direction, the dominance of cutinases, particularly Thermobifida cutinase (THCC), is consistent with literature describing their superior ability to cleave and accommodate ester bonds in polymeric substrates, including aliphatic and aromatic polyesters [1, 9]. Although direct experimental validation of THCC was not possible in this study due to limited enzyme availability, several independent studies report the high hydrolytic efficiency of Thermobifida cutinases toward polyester materials. This supports the plausibility of our computational predictions.

4. Conclusions

Considering the urgent need to develop and implement efficient biocatalytic processes in a timely and cost‐effective manner, we have devised a computational pipeline to identify optimal substrate–catalyst combinations. This pipeline relies on well‐established computational techniques, including docking and MD simulations to address three key challenges: (i) the identification of the optimal biocatalyst for polymerizing or hydrolyzing a given substrate; (ii) the identification of optimal substrates for a given biocatalyst; and (iii) the comprehensive characterization of a given biocatalyst in terms of its predicted reactivity toward a pool of substrates.

The process was implemented in modeFRONTIER, a package equipped with a graphical interface, in which each call to an external program is conveniently represented by a node.

In the present work, protein structures and protein–ligand interactions were investigated using MD and docking methods. Specific software packages (such as Gromacs and AutoDock) were selected and integrated into the automated workflow. The results indicate that the ability of enzymes to hydrolyze or synthesize polyesters can be rationally predicted through the integration of different computational tools.

Conclusions regarding polyester hydrolysis are currently limited by the lack of direct experimental validation due to the unavailability of cutinase THCC in either a commercial or a recombinant form. Consequently, trends in hydrolytic performance are primarily based on computational predictions and supported by previously published experimental evidence. Future experimental studies are required to validate these predictions, and to further extend the applicability of the proposed workflow to guide enzyme selection for polyester degradation processes.

Overall, this study demonstrates that automated computational workflows can effectively be employed to rapidly select enzymes capable of efficiently catalyzing the hydrolysis and/or synthesis of bio‐based polyesters. These results represent a significant contribution to the rational eco‐design of a new generation of tailored sustainable polymers and constitute a first step toward the in silico generation and screening of new biocatalysts [9] for the polymer sector.

Supporting Information

Additional supporting information can be found online in the Supporting Information section. Supporting Fig. S1: Principal Component Analysis (PCA) and Consensus PCA (CPCA) of the dataset. Green: CalB (lipase from Candida antarctica B), yellow: PsSt (lipase from Psudomonas stuzeri), Red: TLL (lipase from Thermomyces lanuginosus), Blue: HIC (cutinase from Humicola insolens), Dark blue: ThcC (cutinase from Thermobifida cellulosilytica). Supporting Fig. S2: Heatmap of enzyme‐substrate interactions for aliphatic polyester synthesis. The colour intensity of each cell represents the docking score, with darker shades indicating stronger binding affinities. Hierarchical clustering reveals clusters of enzymes and ligands with similar binding profiles.

Funding

This work was supported by Horizon 2020 Framework Programme (Grant 101029444).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supplementary Material

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.