Hydrogels, soft materials mimicking natural systems, have garnered significant interest due to their biocompatibility and applications in diverse fields. Peptidic hydrogels, in particular, offer high biocompatibility, low immunogenicity, and similarity to the extracellular matrix, making them ideal for biomedical and materials science applications. However, current design strategies rely on amino acid sequences from natural proteins, experience, or serendipity, limiting the development of novel hydrogels. The inefficient nature of these methods highlights the critical need for accurate prediction of hydrogel formation and de novo design capabilities to broaden the available hydrogel-forming peptide library. Coarse-grained molecular dynamics (CGMD) simulations have been used to model peptide self-assembly and provide design rules, but the computational cost limits their application to longer peptides. Existing computational methods, while helpful, often predict self-assembly but not necessarily hydrogel formation. A systematic study integrating computational, experimental, and machine learning (ML) approaches to predict and design peptidic hydrogels is lacking. This work aims to address this gap by developing a robust and efficient framework for discovering and designing tetrapeptides with high hydrogelation potential.

Previous research has explored the use of CGMD simulations to understand peptide self-assembly and predict the formation of aggregates. Ulijn and Tuttle's groups have developed approaches to provide design rules for dipeptide and tripeptide systems, overcoming the limitations of serendipity in discovering self-assembly. However, MD simulations are computationally expensive, especially for longer peptides, and do not directly predict hydrogel formation. Machine learning approaches have been applied to predict peptide self-assembly and hydrogelation based on chemical features, offering a promising alternative to computationally intensive simulations, but these studies have not fully integrated experimental validation and iterative improvement in a human-in-the-loop fashion. This study builds upon existing work by incorporating experimental feedback to iteratively refine a machine learning model, thereby improving the accuracy and reliability of hydrogel formation predictions.

This study employed a three-loop iterative approach combining CGMD simulations, ML regression and classification models, and experimental validation. **1. CGMD Simulations and ML for Aggregation Propensity (AP) Prediction:** A library of 10,000 tetrapeptide sequences was generated using hypercubic sampling. CGMD simulations using the Martini2 force field were performed to determine the aggregation propensity (AP) for each peptide. A machine learning model (Support Vector Machine, SVM) was trained to predict AP values based on the CGMD simulation data. Various training conditions, including different algorithms and feature representation approaches, were tested to optimize the model's performance. The optimized SVM model was then used to predict the AP for a larger library of 160,000 tetrapeptide sequences. **2. Experiment and ML for Score Function Refinement:** An initial score function (APH) was defined as AP^2 * logP^0.5, where AP is the predicted aggregation propensity and logP represents hydrophobicity. Based on APH, 55 peptides were selected for chemical synthesis and experimental validation of their hydrogel-forming ability. The experimental gelation results (gelation: yes/no) were used to train a classification model (also SVM) to generate a gelation corrector (Cg). This corrector was then incorporated into the score function to create an updated score function (APHC). This iterative process (experiment-ML loop) was repeated three times to refine the gelation corrector and improve prediction accuracy. The process was repeated using additional synthesized peptide data to further improve the accuracy of the classification model and the APHC function. **3. Peptide Synthesis, Purification, and Characterization:** Selected tetrapeptides were synthesized using solid-phase peptide synthesis (SPPS). The purity of the peptides was verified by HPLC and MS, and their structures were confirmed using NMR. Hydrogel formation was assessed using the vial-inverting method. The morphology of the hydrogels and non-hydrogels were characterized using Transmission Electron Microscopy (TEM), and FTIR spectroscopy was used to analyze the secondary structures of the hydrogels. **4. Biological Application (Vaccine Adjuvant):** A de novo designed peptide hydrogel (YAWF) was selected and used as an immune adjuvant in a mouse model. The hydrogel containing the receptor-binding domain (RBD) of SARS-CoV-2 was used as vaccine in mice and compared to RBD alone and RBD with aluminum adjuvant. Antibody responses (IgG, IgG1, IgG2b, IgG2c), cytokine secretion (IL-5, IFN-γ), and dendritic cell activation (CD83, CD80, CD86, IL-6, TNF-α) were assessed to evaluate the adjuvant's effectiveness.

The optimized SVM model demonstrated reliable performance in predicting AP values, with a mean absolute difference (MAE) of 0.095 and a coefficient of determination (R²) of 0.928 in training data, and MAE of 0.092 and R² of 0.933 in testing data. The iterative refinement of the score function (APH to APHC) significantly improved the prediction accuracy. The final APHC score function yielded a gelation hit rate of 87.1% within the top 8,000 sequences. Experimental characterization of synthesized tetrapeptides confirmed the predicted hydrogel formation for many peptides, demonstrating a strong correlation between APHC score and hydrogel formation. TEM and FTIR analyses revealed the formation of nanofibrous and nanosheet structures in the hydrogels, consistent with MD simulations. The analysis of the 165 synthesized peptides (100 hydrogels and 65 non-hydrogels) revealed specific patterns regarding the contribution of amino acids at different positions in the tetrapeptide sequence to hydrogel formation. Specifically, aromatic amino acids (F, Y, and to a lesser extent W), and hydrophobic amino acids (I, L, V, M) at positions 3 and 4 (C-terminus) were found to be important for hydrogel formation, and the presence of certain polar amino acids (S, T) was also found to promote hydrogel formation. Conversely, highly charged amino acids and glycine (G) were detrimental to hydrogelation. The de novo-designed tetrapeptide hydrogel (YAWF) significantly boosted the immune response to the RBD protein in mice, exceeding the effectiveness of the aluminum adjuvant control. The hydrogel-based vaccination resulted in a substantial increase in RBD-specific IgG antibodies, IgG subtypes, and the secretion of cytokines, including IL-5 and IFN-γ, indicating enhanced humoral and cellular immune responses. Furthermore, the hydrogel enhanced dendritic cell activation, suggesting an important mechanism of action in driving the observed immune response.

This study successfully demonstrated a human-in-the-loop framework integrating CGMD, machine learning, and experiments for efficient prediction and discovery of peptide hydrogels. The iterative refinement of the prediction model through experimental validation resulted in a highly accurate score function (APHC), significantly improving the success rate of identifying hydrogel-forming peptides compared to previous methods. The identification of key amino acid sequence features contributing to hydrogel formation provides valuable insights for rational design of new peptide hydrogels. The successful application of a de novo-designed hydrogel as a vaccine adjuvant further validates the framework’s potential for discovering functional peptide materials. The framework developed here could facilitate the discovery and design of other functional materials based on short peptide building blocks.

This research presents a powerful “human-in-the-loop” framework for predicting and discovering peptide hydrogels, combining CGMD simulations, machine learning, and experimental validation. The resulting APHC score function achieved a remarkable 87.1% success rate in predicting hydrogel formation. Furthermore, a de novo-designed peptide hydrogel effectively boosted the immune response in a mouse model, showcasing the framework's potential for biomedical applications. Future work could focus on automating the framework using robotic platforms for peptide synthesis and further expanding its application to other functional materials.

The study focused on tetrapeptides, and the generalizability of the findings to longer peptides requires further investigation. The experimental validation was conducted using a limited number of peptides, although this was sufficient to validate the prediction model. The mouse model used for the vaccine adjuvant study may not fully represent the human immune response. The in vivo studies are limited to evaluating one specific tetrapeptide hydrogel as an adjuvant.