(A-40) Machine learning models for sequence-based prediction of the aggregation and solubility of therapeutic proteins and peptides

Introduction:: The design of proteins and peptides for therapeutic or industrial use has become increasingly important. Nevertheless, understanding and predicting protein/peptide solubility is a challenging and important problem in biology and engineering, due to its importance in prognosing disease and the development of new protein/peptide-based products. Additionally, there is interest in being able to design proteins with specified solubility for applications such as biomaterials. In wet labs, identifying amyloidogenic sequences is of significant concern for peptide solubility due to their role in formation of fibrous aggregates. Therefore, accurate computational methods for the prediction of protein/peptide solubility based on amino acid sequences are highly desirable prior to synthesis. In this study, we have developed data-driven models to classify soluble and non-soluble sequences aimed at identifying factors that contribute to the solubility of proteins and peptides. Past work identified oscillation in the physiochemical properties of protein sequences could be used to categorize protein structure and function and we have extended upon their approach to predict protein solubility.


Materials and Methods:: Developing machine-learned classifications require datasets with known input (e.g., amino acid sequences) and output (e.g., solubility class labels). In this study, we have used two datasets namely; eSOL, and TANGO containing 3157 and 250 samples of aggregating and non-aggregating protein/peptides, respectively. First, we used the AAindex dataset containing 544 amino acid properties to perform hierarchical clustering based on Pearson correlation coefficient to identify representative properties that could be used to encode amino acid sequences (Figure1). We also used the BLOSUM100 matrix as another approach to encoding protein/peptide sequences. After encoding protein/peptide sequences, Fourier transforms were applied to each of the encoded sequence to generate Fourier coefficients measuring the oscillation of each physiochemical/BLOSUM descriptor. Ultimately, the Fourier coefficient served as input features to train our SVM models. Monte Carlo cross-validation has been applied to train and tune classification models based on five train/test sets of each dataset. Feature selection was performed based on non-linear SVMs using a criterion which was previously, derived based on sensitivity analysis of the dual formulation of SVM models. Performance of model was compared with the recently developed models based on various performance metrics.


Results, Conclusions, and Discussions:: The hierarchical clustering of the amino acid properties from AAindex dataset was based on a data matrix consisting of 20 instances (amino acids) of 544 variables (properties). The clustering analysis resulted in 83 clusters of amino acid properties visualized as a heatmap in Figure 2. Clustering of the rows (20 naturally occurring amino acids) shows some patterns of similarities and shows that the amino acids are roughly organized into clusters based on physicochemical properties. The 83 selected amino acid properties were used to generate 83 property vectors for each of the protein sequences. After Fourier transforms, this resulted in 332 features based on 4 frequency components from 83 property vectors for each sequence. For the BLOSUM-based encoding 200 and 660 features for eSOL and TANGO datasets, respectively, were generated. Based on the generated features, we performed feature selection to rank the periodicity features. The feature ranking procedure we applied to each of the five training sets for each dataset and a consensus ranking was generated based on the average rank of each feature across the five training sets. For our best models (with top-ranked features), we further analyzed the performance by generating receiver operator curves (ROC) for all train/testing sets and computed the average area under the curve (AUC). The results obtained are better or on par with other recently developed predictors. The feature selection results provide us with some insight into potential design considerations for increasing the solubility of de novo designed peptide/protein. However, future work will be aimed at further analyzing the relationship between amino acid property periodicities and solubility with the ultimate goal of improving solubility in therapeutic peptides.


Acknowledgements (Optional): :

References (Optional): :