In the past, numerous methods have been developed to predict MHC class II binders or T-helper epitopes for designing the epitope-based vaccines against pathogens. In contrast, limited attempts have been made to develop methods for predicting T-helper epitopes/peptides that can induce a specific type of cytokine. This paper describes a method, developed for predicting interleukin-10 (IL-10) inducing peptides, a cytokine responsible for suppressing the immune system. All models were trained and tested on experimentally validated 394 IL-10 inducing and 848 non-inducing peptides. It was observed that certain types of residues and motifs are more frequent in IL-10 inducing peptides than in non-inducing peptides. Based on this analysis, we developed composition-based models using various machine-learning techniques. Random Forest-based model achieved the maximum Matthews’s Correlation Coefficient (MCC) value of 0.59 with an accuracy of 81.24% developed using dipeptide composition. In order to facilitate the community, we developed a web server “IL-10pred”, standalone packages and a mobile app for designing IL-10 inducing peptides (http://crdd.osdd.net/raghava/IL-10pred/).

Short half-life is one of the key challenges in the field of therapeutic peptides. Various studies have reported enhancement in the stability of peptides using methods like chemical modifications, D-amino acid substitution, cyclization, replacement of labile aminos acids, etc. In order to study this scattered data, there is a pressing need for a repository dedicated to the half-life of peptides. To fill this lacuna, we have developed PEPlife (http://crdd.osdd.net/raghava/peplife), a manually curated resource of experimentally determined half-life of peptides. PEPlife contains 2229 entries covering 1193 unique peptides. Each entry provides detailed information of the peptide, like its name, sequence, half-life, modifications, the experimental assay for determining half-life, biological nature and activity of the peptide. We also maintain SMILES and structures of peptides. We have incorporated web-based modules to offer user-friendly data searching and browsing in the database. PEPlife integrates numerous tools to perform various types of analysis such as BLAST, Smith-Waterman algorithm, GGSEARCH, Jalview and MUSTANG. PEPlife would augment the understanding of different factors that affect the half-life of peptides like modifications, sequence, length, route of delivery of the peptide, etc. We anticipate that PEPlife will be useful for the researchers working in the area of peptide-based therapeutics.

Transcription factors are among the most attractive therapeutic targets but are considered largely ‘undruggable’ in part due to the intrinsically disordered nature of their activation domains. Here we show that the aromatic character of the activation domain of the androgen receptor, a therapeutic target for castration-resistant prostate cancer, is key for its activity as transcription factor, allowing it to translocate to the nucleus and partition into transcriptional condensates upon activation by androgens. On the basis of our understanding of the interactions stabilizing such condensates and of the structure that the domain adopts upon condensation, we optimized the structure of a small-molecule inhibitor previously identified by phenotypic screening. The optimized compounds had more affinity for their target, inhibited androgen-receptor-dependent transcriptional programs, and had an antitumorigenic effect in models of castration-resistant prostate cancer in cells and in vivo. These results suggest that it is possible to rationally optimize, and potentially even to design, small molecules that target the activation domains of oncogenic transcription factors.

The metastatic Skin Cutaneous Melanoma (SKCM) has been associated with diminished survival rates and high mortality rates worldwide. Thus, segregating metastatic melanoma from the primary tumors is crucial to employ an optimal therapeutic strategy for the prolonged survival of patients. The SKCM mRNA, miRNA and methylation data of TCGA is comprehensively analysed to recognize key genomic features that can segregate metastatic and primary tumors. Further, machine learning models have been developed using selected features to distinguish the same. The Support Vector Classification with Weight (SVC-W) model developed using the expression of 17 mRNAs achieved Area under the Receiver Operating Characteristic (AUROC) curve of 0.95 and an accuracy of 89.47% on an independent validation dataset. This study reveals the genes C7, MMP3, KRT14, LOC642587, CASP7, S100A7 and miRNAs hsa-mir-205 and hsa-mir-203b as the key genomic features that may substantially contribute to the oncogenesis of melanoma. Our study also proposes genes ESM1, NFATC3, C7orf4, CDK14, ZNF827, and ZSWIM7 as novel putative markers for cutaneous melanoma metastasis. The major prediction models and analysis modules to predict metastatic and primary tumor samples of SKCM are available from a webserver, CancerSPP (http://webs.iiitd.edu.in/raghava/cancerspp/).

Liver Hepatocellular Carcinoma (LIHC) is one of the major cancers worldwide, responsible for millions of premature deaths every year. Prediction of clinical staging is vital to implement optimal therapeutic strategy and prognostic prediction in cancer patients. However, to date, no method has been developed for predicting the stage of LIHC from the genomic profile of samples. The Cancer Genome Atlas (TCGA) dataset of 173 early stage (stage-I), 177 late stage (stage-II, Stage-III and stage-IV) and 50 adjacent normal tissue samples for 60,483 RNA transcripts and 485,577 methylation CpG sites, was extensively analyzed to identify the key transcriptomic expression and methylation-based features using different feature selection techniques. Further, different classification models were developed based on selected key features to categorize different classes of samples implementing different machine learning algorithms. In the current study, in silico models have been developed for classifying LIHC patients in the early vs. late stage and cancerous vs. normal samples using RNA expression and DNA methylation data. TCGA datasets were extensively analyzed to identify differentially expressed RNA transcripts and methylated CpG sites that can discriminate early vs. late stages and cancer vs. normal samples of LIHC with high precision. Naive Bayes model developed using 51 features that combine 21 CpG methylation sites and 30 RNA transcripts achieved maximum MCC (Matthew’s correlation coefficient) 0.58 with an accuracy of 78.87% on the validation dataset in discrimination of early and late stage. Additionally, the prediction models developed based on 5 RNA transcripts and 5 CpG sites classify LIHC and normal samples with an accuracy of 96–98% and AUC (Area Under the Receiver Operating Characteristic curve) 0.99. Besides, multiclass models also developed for classifying samples in the normal, early and late stage of cancer and achieved an accuracy of 76.54% and AUC of 0.86. Our study reveals stage prediction of LIHC samples with high accuracy based on the genomics and epigenomics profiling is a challenging task in comparison to the classification of cancerous and normal samples. Comprehensive analysis, differentially expressed RNA transcripts, methylated CpG sites in LIHC samples and prediction models are available from CancerLSP (http://webs.iiitd.edu.in/raghava/cancerlsp/).

Determining the aggressiveness of gliomas, termed grading, is a critical step toward treatment optimization to increase the survival rate and decrease treatment toxicity for patients. Streamlined grading using molecular information has the potential to facilitate decision making in the clinic and aid in treatment planning. In recent years, molecular markers have increasingly gained importance in the classification of tumors. In this study, we propose a novel hierarchical voting-based methodology for improving the performance results of the feature selection stage and machine learning models for glioma grading with clinical and molecular predictors. To identify the best scheme for the given soft-voting-based ensemble learning model selections, we utilized publicly available TCGA and CGGA datasets and employed four dimensionality reduction methods to carry out a voting-based ensemble feature selection and five supervised models, with a total of sixteen combination sets. We also compared our proposed feature selection method with the LASSO feature selection method in isolation. The computational results indicate that the proposed method achieves 87.606% and 79.668% accuracy rates on TCGA and CGGA datasets, respectively, outperforming the LASSO feature selection method.