Biology subjects

Mehta, N. K.

Publications and source records attributed to Mehta, N. K..

10 recordsLinked to original sources

bioRxiv2026

AntiCapt: Fine-Tuned Nucleotide Language Models for Predicting and Designing Anticancer Aptamers

Over the past decade, aptamers have emerged as promising therapeutics, with cancer therapeutics as a major research area. Existing computational methods, however, either predict aptamer-target pairs or design aptamers against a specific target. In this study, we present AntiCapt, a computational method for identifying single-stranded DNA (ssDNA) anticancer aptamers (ACAs) using sequence descriptors and fine-tuned nucleotide language models. The main dataset comprises 1,021 experimentally validated ACAs and an equal number of non-aptamer sequences. Comparative analysis revealed distinct patterns in the nucleotide, dinucleotide, and trinucleotide compositions associated with ACAs. We developed machine learning (ML) models using composition, autocorrelation, binary profiles and structural features. Among these, the best-performing composition-based model achieved an AUC of 0.93 on an independent dataset, while chemical descriptor-and structural feature-based models achieved a maximum AUC of 0.89. ML models using pretrained and fine-tuned NLM embeddings were also developed, with fine-tuned HyenaDNA embeddings achieving the highest performance, with an independent AUC of 0.94. Furthermore, we developed a model for discriminating anticancer aptamers from general aptamers. The best-performing models were integrated into AntiCapt, a web server and a standalone tool for predicting, designing, and genome-scale scanning anticancer aptamers (https://webs.iiitd.edu.in/raghava/anticapt/).

bioinformatics↗

bioRxiv2026

CPPLocPred: Subcellular Localization of Cell-Penetrating Peptides

Cell-penetrating peptides (CPPs) are widely used to deliver therapeutic cargoes into cells. Although numerous computational methods have been developed for identifying CPPs and several predictors are available for protein subcellular localization, no method has been developed to predict the subcellular localization of CPPs. Here, we present CPPLocPred, a hierarchical machine-learning (ML) framework that predicts CPPs and their subcellular localization. In the first stage, we developed ML models to identify CPPs, achieving an AUC of 0.953 with an MCC of 0.7842 on an independent set, exhibiting performance equivalent to or better than existing state-of-the-art methods. In the second stage, we developed a method for predicting the subcellular localization of CPPs. Subcellular localization methods were trained (80% data using five-fold cross-validation) and validated (20% data) on experimentally validated CPPs for 663 Cytoplasm, 287 Nucleus, 57 Mitochondria, 186 Endo_lysosome, and 328 Others. Our primary analysis revealed that Mitochondrial and Nuclear associated CPPs are abundant in positively charged arginine- and lysine-rich patterns, whereas Endo_lysosomal CPPs preferentially comprise glycine-, proline-, and cysteine-rich motifs. We used a wide range of traditional peptide features, along with the embedding of protein language models, to develop ML models. Among all evaluated models, the CatBoost-based subcellular localization models with Distance Distribution of Residues (DDR) achieved AUCs of 0.814, 0.775, 0.970, 0.782, and 0.798 for Cytoplasm, Nucleus, Mitochondria, Endo_lysosome, and Others, respectively, on validation dataset. We developed CPPLocPred, which offers a practical platform for functional annotation and rational design of localization-specific CPPs for therapeutic applications (https://webs.iiitd.edu.in/raghava/cpplocpred/). HighlightsO_LIPrediction and subcellular localization of cell-penetrating peptides. C_LIO_LILocalization of CPPs depends on their amino acid and dipeptide composition. C_LIO_LIBest feature for subcellular localization was Distance Distribution of Residues. C_LIO_LICatBoost model achieved the highest performance for subcellular localization. C_LIO_LIA web server and standalone software to facilitate its use by the scientific community. C_LI

bioinformatics↗

bioRxiv2026

PEPstrMOD2: Next-generation tertiary structure prediction of chemically modified and non-natural peptides

While most existing methods are limited to predicting the tertiary structures of proteins containing only canonical residues, the PEPstrMOD server (developed in 2015) pioneered structure prediction for chemically modified and non-natural peptides. Despite its widespread use, the original framework was restricted to peptides of 7 to 25 residues and relied on older backbone-prediction algorithms. To address these limitations, we present PEPstrMOD2, which introduces three major advancements over its predecessor. First, it replaces the original in-house coordinate generation with state-of-the-art deep learning (DL) algorithms, leveraging AlphaFold2 and ESMFold for highly accurate initial structure prediction. Secondly, it greatly expands the accessible chemical space through incorporation of new, AMBER force-field compatible library of 257 post-translational modifications (PTMs), 428 non-canonical amino acids (NCAAs), and 243 terminal modifications. Lastly, through the application of native scalability of AlphaFold2 (AF2) and ESMFold (EF), PEPstrMOD2 eliminates the original restrictions of the length, enabling the structural modeling of longer, complex therapeutic peptides and small proteins. We evaluated the performance of PEPstrMOD2 against state-of-the-art methods across three distinct peptide datasets. For the AfCyc dataset consisting of 80 cyclic peptides, PEPstrMOD2 obtained a competitive average atom-level Root Mean Square Deviation (RMSD) of 2.05 [A], compared to 1.13 [A] by AlphaFold3 (AF3) and 1.82 [A] by AfCycDesign. Remarkably, for the modified peptide ModPep433 dataset, PEPstrMOD2 outperformed AF3, achieving the lower average RMSD score of 4.49 [A] against 4.67 [A] of AF3. Furthermore, in the case of the ModPep16 benchmark, PEPstrMOD2 achieved 2.50 [A] average RMSD value, which is two times more accurate than that of the original PEPstrMOD (5.84 [A]). In summary, PEPstrMOD2 provides a powerful, high-throughput, and highly accurate platform to facilitate peptide-based drug development and structural biology research. While the original PEPstrMOD was restricted to a web server interface, PEPstrMOD2 is available as both an intuitive webserver and a standalone command-line tool via GitHub, featuring Docker support for easy deployment and reproducible, large-scale modeling pipelines (https://webs.iiitd.edu.in/raghava/pepstrmod/). HighlightsO_LIPEPstrMOD2 predicts structures of chemically modified peptides using deep learning and molecular dynamics refinement. C_LIO_LISupports 928 chemical modifications through custom AMBER based force-field libraries including PTMs, NCAAs, terminal, D- as well as cyclic modifications. C_LIO_LIRemoves peptide length restrictions through AlphaFold2 and ESMFold integration. C_LIO_LIutperforms PEPstrMOD and is competitive with AF3 on different benchmark datasets. C_LIO_LIAvailable as both a web server and a standalone platform via GitHub and Docker. C_LI Authors BiographyO_LISaloni Jain is currently working as Ph.D. in Computational biology from Department of Computational Biology, Indraprastha Institute of Information Technology, New Delhi, India C_LIO_LINaman Kumar Mehta is currently working as Ph.D. in Computational biology from Department of Computational Biology, Indraprastha Institute of Information Technology, New Delhi, India C_LIO_LISahil Raina is currently working as Ph.D. in Computational biology from Department of Computational Biology, Indraprastha Institute of Information Technology, New Delhi, India C_LIO_LIPankaj Kumar is currently working as Ph.D. in Computational biology from Department of Computational Biology, Indraprastha Institute of Information Technology, New Delhi, India C_LIO_LIVarun is currently an integrated BS-MS student at Indian Institute of Science Education and Research (IISER) Pune, India. He is currently working as an intern on a project position at Department of Computational Biology, Indraprastha Institute of Information Technology (IIIT), New Delhi, India. C_LIO_LIGajendra P. S. Raghava is currently working as a Professor in the Department of Computational Biology, Indraprastha Institute of Information Technology, New Delhi, India. C_LI

bioinformatics↗

bioRxiv2026

HAIRpred2: Human Host-Specific Prediction of Antibody-Interacting Residues Using Hybrid Physicochemical and Structural Features

1.Prediction of conformational B-cell epitopes is critical for vaccine design, immunotherapy, and antibody engineering. To date, several host-independent computational methods have been developed for predicting antibody-interacting residues in antigen structures. However, it is well established that antigen-antibody (Ag-Ab) interactions vary depending on the host immune system indicating the importance of developing host-specific prediction models. In this study, we present, for the first time, a human host-specific method, HAIRpred2, that predicts antibody-interacting residues in an antigen from its tertiary structure. The dataset was derived from HAIRpred and comprises 277 human Ag-Ab complexes, with 221 structures used for training and 56 for independent testing. Preliminary analysis revealed that residues with a relative surface accessibility (RSA) below 0.05, corresponding to buried regions, are highly likely to be non-interacting, underscoring the importance of structural accessibility in antibody recognition. To identify the most informative features, we evaluated multiple feature representations, including RSA, large language model (LLM)-based embeddings, distance-based features, and physicochemical properties. A model trained on single-residue RSA features achieved an AUC of 0.72. Incorporating a sliding window of 15 residues to capture local structural context improved performance to an AUC of 0.75. The best performance (AUC = 0.78 on the independent test set) was achieved by integrating RSA with physicochemical descriptors. Benchmarking against existing antibody-interaction prediction methods on the same independent dataset demonstrated that HAIRpred2 outperforms current tools, further highlighting the advantage of host-specific modeling. HAIRpred2 is freely available as a web server at https://webs.iiitd.edu.in/raghava/hairpred2/. HighlightsO_LIDevelopment of HAIRpred2, the first human host-specific method for predicting antibody-interacting residues. C_LIO_LIAnalysis of 277 human antigen-antibody complexes to capture host-dependent interaction patterns. C_LIO_LIRelative surface accessibility (RSA) identified as a key determinant, with buried residues rarely participating in interactions. C_LIO_LIIntegration of RSA with physicochemical features achieved the best performance (AUC = 0.78) on an independent dataset. C_LIO_LIHAIRpred2 outperforms existing methods and is available as a web server for epitope prediction. C_LI

bioinformatics↗

bioRxiv2025

A PLM-Based Method for Predicting Protein Ion Channel Modulators for Drug Discovery and Safety Evaluation

Ion channels are central to regulating neuronal communication, cardiac rhythm, and muscle contraction. Their modulation can induce therapeutic benefits but may also lead to adverse or toxic effects. This study presents IonNTXpred, a protein language model (PLM)-based method for predicting protein ion channel modulators, including channel-specific (Na{square}, K{square}, Ca{superscript 2}{square}, and others) and moonlighting proteins capable of modulating multiple ion channels. We train, test, and evaluate our models on the largest dataset of non-redundant ion channel-modulating proteins, where no two proteins have more than 40% sequence identity. Composition analysis revealed that residues Cys, Gly, and Trp are highly prevalent, whereas Ala, Glu, Leu, Gln, and Val are scarce in ion channel-modulating proteins, with Cys identified as a key discriminative residue. We explored both alignment-based (BLAST, MERCI) and alignment-free (machine learning, deep learning, and PLM-based) approaches. Among these, our evolutionary information-based PLM (ESM2-t33) achieved the best performance, with an AUROC of 0.97 across ion channels, which further improved to 0.98 when integrated with BLAST output. The proposed method outperformed existing approaches on independent datasets. We used IonNTXpred to screen FDA-approved and organismal proteins to identify candidates with ion channel-modulating potential, supporting drug repurposing, discovery of new therapeutic proteins, and safety assessment of existing biologics. We implemented these models in a user-friendly web server, IonNTXpred, which facilitates the design and discovery of ion channel-modulating protein-based drugs and supports the biosafety evaluation of therapeutic proteins through neurotoxin screening (https://webs.iiitd.edu.in/raghava/ionntxpred/). HighlightsO_LIPrediction, design, and large-scale screening of ion channel-modulating proteins. C_LIO_LIIdentifies neurotoxins to support safety evaluation of therapeutic proteins. C_LIO_LIEvolutionary information-based protein language model ESM2-t33 for prediction. C_LIO_LIPredicts moonlighting proteins capable of modulating multiple ion channels. C_LIO_LIProvides standalone software, a web server, and a PyPI package for user accessibility. C_LI

bioinformatics↗

bioRxiv2025

Benchmarking of Quantum SVM and Classical ML Algorithms for Prediction of Therapeutic Proteins

Over the past decade, quantum machine learning, particularly quantum support vector machines (QSVMs), has emerged as an optimistic alternative to classical machine learning (CML) techniques. This study rigorously benchmarks the performance of QSVM and CML-based models across four diverse datasets relevant to therapeutic proteins and peptides. Specifically, we evaluated these approaches for the prediction of B-cell epitopes (CLBtope), exosomal proteins (ExoPropred), hemolytic peptides (HemoPI), and toxic peptides (Toxinpred3). The maximum area under the receiver operating characteristic curve (AUC) for the CLBtope dataset achieved was 0.68 for QSVM and 0.82 for CML models. For the ExoPropred dataset, the maximum AUCs were 0.66 (QSVM) and 0.72 (CML). In contrast, both QSVM and CML models demonstrated high performance on the HemoPI dataset, yielding maximum AUCs of 0.95 and 0.98, respectively. Similarly, for the Toxinpred3 dataset, the maximum AUCs were 0.84 (QSVM) and 0.94 (CML). All models were evaluated using independent validation datasets not used during training. These results suggest that although CML currently demonstrates superior predictive capability for these tasks, the similar progression in performance indicates potential for future advancements in QSVM. HighlightsO_LIComparative study of QSVM and CML models on four bioinformatics datasets C_LIO_LIQSVM performance tries to approach CML in tasks involving hemolytic and toxic peptide prediction C_LIO_LIIndependent validation confirms robustness of performance metrics C_LIO_LIResults highlight the potential of QSVMs as real-world quantum hardware continues to matures C_LI

bioinformatics↗

bioRxiv2025

IL4Pred2: Prediction of Interleukin-4 Inducing Peptides in Human and Mouse

In 2013, our group developed IL4pred, a host-independent method for predicting interleukin-4 (IL-4) inducing peptides, which has been widely used by the scientific community. In this study, we present a second-generation method, IL4Pred2, which is a host-specific approach designed to predict IL-4 inducing peptides separately for human and mouse hosts. All models were trained, tested, and benchmarked on experimentally validated data obtained from the IEDB. We employed a wide range of state-of-the-art techniques for prediction, including similarity-based approaches, machine learning, deep learning methods, and large language models. Our best model achieved highest AUC 0.80 with MCC 0.45 for human and AUC 0.82 with MCC 0.50 for mouse on independent set of main datasets. All models were trained, test and optimized on training dataset. We validate our final model on an independent dataset which is not used in training or hyperparameter optimization of models. In this study, we developed models on three types of datasets called Main, Alternate1 and Alternate2 for predicting IL-4 inducing peptides. This abstract show performance of our models on Main dataset, performance of models on other datasets have been discussed in manuscript. One of the major objectives of this study is to facilitate research community in the area of immunotherapy and vaccine development. Thus, we developed, a web server and standalone software IL4pred2 for predicting, designing and scanning IL-4 inducing peptides in proteins (https://webs.iiitd.edu.in/raghava/il4pred2/ and https://github.com/raghavagps/il4pred2).

bioinformatics↗

bioRxiv2025

A large language model for predicting pancreatic ductal adenocarcinoma patients from blood-derived exosomal transcriptomics data

Traditional machine learning approaches for text or sequence classification rely on converting textual data into numerical representations. In this study, we investigate a reverse strategy in which numerical features are transformed into sequence representations and classified using large language models (LLMs). We applied this methodology to predict pancreatic ductal adenocarcinoma (PDAC) using the expression profiles of 50 genes from 284 PDAC and 217 non-PDAC patients. Gene expression values were converted into sequence data, with each gene represented as a residue in a 50-residue protein sequence. Major LLMs like PeptideBERT, ProtBERT, and ESM2 were fine-tuned on a protein training dataset and evaluated on an independent dataset. The best-performing model, ProtBERT, achieved an AUC of 0.962 on an independent dataset. Additionally, an alignment-based approach employing BLAST and MERCI motifs was explored, and an ensemble model combining the LLM-based and alignment-based methods was developed. Our LLM-based model outperformed traditional machine learning models. To the best of our knowledge, this is the first study demonstrating the application of LLMs for mining transcriptomic profiles of cancer patients. HIGHLIGHTSO_LIIdentification of over and under-expressed genes in PDAC patients C_LIO_LIConvert numeric gene expression data to peptide sequence C_LIO_LILLM based models for predicting PDAC patients using peptide sequences C_LIO_LIMining of transcriptomics data using ProtBert and ESM2 C_LIO_LIGene expression profile for diagnostic of PDAC patients C_LI

bioinformatics↗

bioRxiv2024

CytoLNCpred - A computational method for predicting cytoplasm associated long-coding RNAs in 15 cell-lines

The function of long non-coding RNA (lncRNA) is largely determined by its specific location within a cell. Previous methods have used noisy datasets, including mRNA transcripts in tools intended for lncRNAs, and excluded lncRNAs lacking significant differential localization between the cytoplasm and nucleus. In order to overcome these shortcomings, a method has been developed for predicting cytoplasm-associated lncRNAs in 15 human cell-lines, identifying which lncRNAs are more abundant in the cytoplasm compared to the nucleus. All models in this study were trained using five-fold cross validation and tested on an independent dataset. Initially, we developed machine and deep learning based models using traditional features like composition and correlation. Using composition and correlation based features, machine learning algorithms achieved an average AUC of 0.7049 and 0.7089, respectively for 15 cell-lines. Secondly, we developed machine based models developed using embedding features obtained from the large language model DNABERT-2. The average AUC for all the cell-lines achieved by this approach was 0.6604. Subsequently, we also fine-tuned DNABERT-2 on our training dataset and evaluated the fine-tuned DNABERT-2 model on the independent dataset. The fine-tuned DNABERT-2 model achieved an average AUC of 0.6336. Correlation-based features combined with ML algorithms outperform LLM-based models, in the case of predicting differential lncRNA localization. These cell-line specific models as well as web-based service are available to the public from our web server (https://webs.iiitd.edu.in/raghava/cytolncpred/) . HIGHLIGHTSO_LIPrediction of cytoplasm-associated lncRNAs in 15 human cell lines C_LIO_LIMachine learning using composition and correlation features C_LIO_LIDNABERT-2 embeddings for lncRNA localization prediction C_LIO_LICorrelation-based models outperform LLM-based models C_LIO_LIWeb server and models available for public use C_LI AUTHORS BIOGRAPHYO_LIShubham Choudhury is currently working as Ph.D. in Computational Biology from Department of Computational Biology, Indraprastha Institute of Information Technology, New Delhi, India. C_LIO_LINaman Kumar Mehta is currently working as Ph.D. in Computational Biology from Department of Computational Biology, Indraprastha Institute of Information Technology, New Delhi, India. C_LIO_LIGajendra P. S. Raghava is currently working as Professor and Head of Department of Computational Biology, Indraprastha Institute of Information Technology, New Delhi, India C_LI

bioinformatics↗

bioRxiv2024

Prediction of Hemolytic Peptides and their Hemolytic Concentration (HC50)

Several peptide-based drugs fail in clinical trials due to their toxicity or hemolytic activity against red blood cells (RBCs). Existing methods predict hemolytic peptides but not the concentration (HC50) required to lyse 50% of RBCs. In this study, we developed a classification model and regression model to identify and quantify the hemolytic activity of peptides. Our models were trained and validated on 1924 peptides with experimentally determined HC50 against mammalian RBCs. Analysis indicates that hydrophobic and positively charged residues were associated with higher hemolytic activity. Our classification models achieved a maximum AUC of 0.909 using a hybrid model of ESM-2 and a motif-based approach. Regression models using compositional features achieved R of 0.739 with R{superscript 2} of 0.543. Our models outperform existing methods and are implemented in the web-based platform HemoPI2 and standalone software for designing hemolytic peptides with desired HC50 values (http://webs.iiitd.edu.in/raghava/hemopi2/). HighlightsO_LIDeveloped classification and regression models to predict hemolytic activity and HC50 values of peptides. C_LIO_LIA hybrid model combining machine learning and motif prediction excels in accuracy. C_LIO_LIBenchmarking of the existing classification methods on independent datasets. C_LIO_LIWeb server, standalone software, and pip package for hemolytic activity prediction of peptides/proteins. C_LI

bioinformatics↗