Research Projects

Peptides are short amino acid chains, play crucial biological roles as hormones, growth factors, and antimicrobials, making them promising candidates for therapeutic treatment of cancer, diabetes and cardiovascular disorders (PMID:39038391; PMID:25555724). Peptide and protein based therapeutics offer several advantages over traditional small-molecule drugs (Shown in Figure 1), especially in combating drug resistance, due to their low immunogenicity, high specificity, effectivenessand minimal drug-drug interactions (BioRxiv:2025.03.01.640936). However, their therapeutic application is often hindered by a range of toxicities such as acute and clinical toxicity, carcinogenicity, organ-specific toxicity, cellular and molecular toxicity (e.g., hematotoxicity, mitochondrial toxicity) and antimicrobial toxicity (PMID: 39899688). Despite their promise, only 31 peptide drugs have been FDA-approved, more than 200 peptides are in clinical trials, and about 600 are in preclinical stages since 2016, largely due to safety and ADMET (absorption, distribution, metabolism, excretion, and toxicity) issues. Early toxicity prediction is crucial for identifying viable drug candidates. Known experiments of toxicity assessment, including animal testing and in vitro assays like hemolytic and cytotoxicity tests which include systemic expression changes, provide reliable results but are often time-consuming and labor-intensive. Thus, in silico tools have been developed for high-throughput toxicity screening, mostly alignment-based methods which often face challenges with similarity thresholds and reducing their accuracy (PMID:40529180). Machine learning tools like ToxinPred and ToxTeller improved prediction by learning from both toxic and non-toxic sequences, while deep learning models like ToxinPred3 and NTxPred2 further enhanced prediction accuracy. However, many models ignore structural data, and lack into interpretability. To address these shortcomings, newer models such as ToxiPep, ToxIBTL, and tAMPer have been developed (Summarized in Table 1). These hybrid models combine sequence embeddings (using BiGRUs and Transformers) with atomic-level graph features derived from peptide structures using CNNs, enabling more accurate toxicity predictions, especially for short peptides (PMID:40529180; PMID:39196703). Although sequence-based features are widely used in de novo prediction, the incorporation of gene expression data, which sometimes inferred from sequences can provide insights into toxicity mechanisms, is underutilized in current models. Similarly, the rise of accurate structural predictions via AlphaFold can enhance the prediction reliability of model. This project proposes the development of an integrated model that combines sequence, structure, and gene expression data to improve the accuracy and interpretability of peptide toxicity prediction. Such a model would speed up peptide and protein based therapeutics and address the current computational gaps.