Supramolecular hydrogels derived from nucleosides are attracting significant attention in biomedicine due to their biocompatibility and unique properties. Applications span drug delivery, biosensors, and tissue engineering. While progress has been made in developing nucleoside-based hydrogels, predicting hydrogel formation from a nucleoside derivative remains a significant challenge. Current methods rely on trial-and-error or serendipitous discoveries, lacking a robust predictive framework. This study addresses this gap by leveraging machine learning (ML) to establish a predictive model for nucleoside hydrogel formation. ML offers a powerful approach to learn complex structure-property relationships from high-dimensional data, potentially uncovering new hydrogelators. Previous work has applied ML to predict hydrogel formation in other molecule types, such as dipeptides, but a specific model for nucleoside derivatives has been lacking due to the intricate self-assembly processes involved. This research aims to develop and validate an ML model capable of accurately predicting the hydrogel-forming ability of nucleoside derivatives, thereby accelerating the discovery of novel biomaterials.

The authors conducted a systematic literature review using MeSH terms and searching databases such as Medline, Web of Science, and SciFinder. They identified 71 nucleoside derivatives with reported hydrogel-forming ability (38 gelators and 33 non-gelators) from 18 articles, establishing a dataset for model development. The inclusion criteria for studies focused on clear definition of gelation, provision of chemical structures, use of pure water or aqueous solutions, and exclusion of base or nucleotide derivatives. The limited number of reported nucleoside hydrogels is a limitation acknowledged by the authors.

The methodology comprised several key steps: 1. **Dataset Creation:** 71 nucleoside derivatives were compiled, along with their hydrogel-forming properties, to create the initial dataset. Their chemical structures were converted into SMILES strings for computational processing. A total of 5666 molecular descriptors were initially calculated. After removing missing values, 4175 descriptors were retained. 2. **Descriptor Calculation:** A total of 4175 molecular descriptors were calculated for each nucleoside derivative using the Python package alvaDescCLIWrapper. This step involved a comprehensive characterization of various chemical features of the molecules. 3. **Feature Selection:** A three-step feature selection process was employed to reduce dimensionality and improve model accuracy. This involved: * Rank-sum test (reduced descriptors to 144) * Spearman correlation (reduced to 40, eliminating highly correlated descriptors) * Recursive feature elimination (RFE) using four machine learning algorithms (XGBoost, LR, DT, RF) to identify the most informative descriptors. The logistic regression (LR) model with 24 features after RFE emerged as the optimal feature set. 4. **Model Development:** Four machine learning algorithms (Logistic Regression (LR), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Decision Tree (DT)) were trained and compared using the selected features. The performance of each model was evaluated using metrics such as accuracy, AUC, precision, recall, and F1 score. Five-fold stratified cross-validation was employed to mitigate overfitting. A sensitivity analysis involved training the model on 80% of the dataset and testing on the remaining 20%, to further validate the model's generalizability. 5. **External Model Application:** The optimal LR model was then applied to an external dataset of 7257 nucleoside derivatives obtained from the PubChem database to identify potential new hydrogelators. This involved ranking the molecules based on their predicted probability of hydrogel formation. 6. **Experimental Validation:** Twelve high-probability and twelve low-probability nucleoside derivatives, selected from the PubChem predictions based on the cost of obtaining/synthesizing them, were experimentally tested for hydrogel formation using tube-inversion tests. This experimental validation aimed to assess the model's predictive power on unseen data. Furthermore, various characterization techniques were used to study the hydrogel properties (rheology, SEM, AFM, VT-SAXS, NMR). The self-assembly mechanism of two novel cation-independent hydrogels was investigated using techniques including NMR, fluorescence assays, CD spectroscopy, UV spectroscopy, single-crystal X-ray diffraction, and theoretical calculations. 7. **Application Exploration:** The potential applications of the novel cation-independent hydrogels were explored, particularly their use in rapid visual detection of Ag+ and cysteine.

The key findings of this study include: 1. **Development of a predictive model:** A robust machine learning model was developed to predict nucleoside hydrogel formation, with the Logistic Regression (LR) model using 24 descriptors achieving the highest test accuracy (71%, 95% CI: 0.69-0.73) and AUC (0.84). 2. **External validation:** Applying the optimal model to 7257 nucleoside derivatives from PubChem successfully identified 20 out of 24 experimentally tested molecules (83.33% accuracy) as hydrogel formers or non-formers. This demonstrated the model's effectiveness in predicting hydrogel-forming ability on unseen data. 3. **Discovery of new hydrogels:** The external validation led to the discovery of two novel cation-independent nucleoside hydrogels, 8AG-T and 8OHG-T, derived from 8-aminoguanosine and 8-hydroxyguanosine respectively. These hydrogels exhibited exceptional stability and self-healing properties. 4. **Mechanism elucidation:** The self-assembly mechanism of these cation-independent hydrogels was investigated. Dynamic borate ester bonds and G-ribbon structures, rather than G-quartets, were identified as crucial for hydrogel formation. The anti-glycosidic bond preference of the molecules plays a key role in this process. 5. **Potential application:** The 8OHG-T hydrogel demonstrated potential for use in rapid visual detection of Ag+ and cysteine, offering a simple and efficient method for detecting these important biomarkers.

The successful development and validation of an ML model for predicting nucleoside hydrogel formation significantly advances the field. The model's ability to predict hydrogel formation with high accuracy (71% internally, 83.33% externally) on both training and external datasets highlights its robustness. The discovery of two novel cation-independent hydrogels expands the library of available nucleoside-based materials, potentially opening new avenues for biomedical applications. The elucidation of the self-assembly mechanism sheds light on the key molecular features that govern hydrogel formation in this class of compounds. The model's application in the discovery of new hydrogelators significantly accelerates research in this field by reducing reliance on trial-and-error approaches. The finding that the 8OHG-T hydrogel can be utilized for rapid visual detection of Ag+ and cysteine further underscores the practical significance of this research. The development of a simple, portable detection method for these analytes offers valuable implications for environmental monitoring and clinical diagnostics.

This study demonstrates the successful application of machine learning in predicting the hydrogel-forming ability of nucleoside derivatives. The developed model exhibits high accuracy and has facilitated the discovery of two novel cation-independent hydrogels with potential biomedical applications. Future work could focus on expanding the dataset to encompass a broader range of nucleoside derivatives and exploring the applicability of the model to other supramolecular hydrogel systems. Further investigation into the mechanisms and applications of the discovered hydrogels will be crucial for their translation into practical uses.

The accuracy of the model is limited by the size and diversity of the initial dataset (71 nucleoside derivatives). The relatively small number of known nucleoside hydrogels restricts the model's ability to capture the full complexity of structure-property relationships. While external validation was performed, additional experimental validation on a larger and more diverse set of compounds would further strengthen the model's reliability. The interpretation of the 24 key descriptors identified by the model remains somewhat challenging, requiring further investigation to establish clear structure-activity relationships.