Kuan-Yuan Chang

Accurate protein subcellular localization is essential for biological function, and mislocalization is linked to numerous diseases. While current methods like DeepLoc 2.0 employ lightweight fine-tuning of protein language models (PLMs), their ability to predict multi-compartment localization remains limited. To address this, we introduce DualLoc, a multi-label localization predictor for ten compartments. DualLoc leverages full-parameter fine-tuning of a cascaded dual-transformer architecture, built upon foundational PLMs and augmented with attention and dropout layers. We evaluated this framework using three foundational PLMs—ProtBERT, ESM-2, and ProtT5—as backbones. Cross-validation on Swiss-Prot and independent validation on the Human Protein Atlas demonstrate consistent superiority over state-of-the-art baselines. The best-performing variant, DualLoc-ProtT5, achieves 0.5872 accuracy, 0.8271 micro-F1, and 0.7811 macro-F1, with substantial gains in the Matthews correlation coefficient for the nucleus (+0.13), cell membrane (+0.13), and extracellular space (+0.07). Pointwise mutual information analysis of model outputs reveals biologically relevant compartment couplings, notably between the Golgi apparatus and endoplasmic reticulum (PMI = 0.25, P <10-6), accurately reflecting secretory pathway coordination. DualLoc provides both a highly accurate predictive tool and a robust framework for investigating protein multi-localization mechanisms.

Background: Current research on severe acute respiratory syndrome coronavirus 2 (SARS CoV-2) has predominantly focused on identifying post-infection protein biomarkers to predict disease severity. However, protein signatures of pre-infection susceptibility remain poorly understood. No previous study has leveraged baseline serum protein expression profiles to predict an individual’s risk of SARS-CoV-2 infection prior to viral exposure. Methods: A machine-learning framework, PreRisk-CoV2, was developed to predict SARS-CoV-2 infection risk by integrating a genetic algorithm for feature selection with a k-nearest neighbors classifier. The model was trained and internally validated using baseline serum protein expression data from the COVID-19 Health Action Response for Marines cohort (CHARM; n = 77) via leave-one-out cross-validation. Independent external validation was performed using the COVID-19 Early Immune Markers Study dataset (CEIM; n = 18). Findings: Using the PreRisk-CoV2 framework, we identified a parsimonious seven protein panel—monocyte chemoattractant protein-3 (MCP-3), leukaemia inhibitory factor receptor (LIF-R), TNF-related activation-induced cytokine (TRANCE), fibroblast growth factor 23 (FGF-23), neurotrophin-3 (NT-3), C-X-C motif chemokine ligand 1 (CXCL1), and C-X-C motif chemokine ligand 6 (CXCL6)—that predicts SARS-CoV-2 susceptibility with high accuracy. The model demonstrated robust performance in the discovery cohort (accuracy: 90·1%; sensitivity: 91·7%; specificity: 85·8%) and maintained predictive utility in an independent external validation cohort (accuracy: 73·2%; sensitivity: 73·3%; specificity: 73·0%). These findings establish that baseline immune-related protein profiles are significantly associated with individual infection risk and can be leveraged to predict pre-exposure risk. Interpretation: Pre-infection serum protein expression profiling can identify individuals at heightened risk of SARS-CoV-2 infection within three months of exposure. The validated PreRisk-CoV2 framework provides a biological basis for precision prevention strategies, such as prioritized vaccination and proteomic-based early-warning systems, to mitigate community transmission. These findings advance our understanding of the immune determinants that govern viral susceptibility.

Papain-like protease (PL^pro) is critical to COVID-19 infection. Therefore, it is a significant target protein for drug development. We virtually screened a 26,193 compound library against the PL^pro of SARS-CoV-2 and identified several drug candidates with convincing binding affinities. The three best compounds all had better estimated binding energy than those of the drug candidates proposed in previous studies. By analyzing the docking results for the drug candidates identified in this and previous studies, we demonstrate that the critical interactions between the compounds and PL^pro proposed by the computational approaches are consistent with those proposed by the biological experiments. In addition, the predicted binding energies of the compounds in the dataset showed a similar trend as their IC₅₀ values. The predicted ADME and drug-likeness properties also suggested that these identified compounds can be used for COVID-19 treatment.

An increasing number of studies have demonstrated the antiviral nature of polyphenols, and many polyphenols have been proposed to inhibit SARS-CoV or SARS-CoV-2. Our previous study revealed the inhibitory mechanisms of polyphenols against DNA polymerase α and HIV reverse transcriptase to show that polyphenols can block DNA elongation by competing with the incoming NTPs. Here we applied computational approaches to examine if some polyphenols can also inhibit RNA polymerase (RdRp) in SARS-CoV-2, and we identified some better candidates than remdesivir, the FDA-approved drug against RdRp, in terms of estimated binding affinities. The proposed compounds will be further examined to develop new treatments for COVID-19.

Although scientists around the world have put lots of effort into the development of new treatments for COVID-19 since the outbreak, no drugs except Veklury (remdesivir) have been approved by FDA. There is an urgent need to discover some alternative antiviral treatment for COVID-19. Because polyphenols have been shown to possess antiviral activities, here we conducted a large-scale virtual screening for more than 400 polyphenols. Several lead compounds such as Petunidin 3-O-(6″-p-coumaroyl-glucoside) were identified to have promising binding affinities and convincing binding mechanisms. Analyzing the docking results and ADME properties sheds light on the potential efficacy of the top-ranked drug candidates and pinpoints the key residues on the target proteins for the future of drug development.

It is essential for future research to develop a new, reliable prediction method of DNA binding sites because DNA binding sites on DNA-binding proteins provide critical clues about protein function and drug discovery. However, the current prediction methods of DNA binding sites have relatively poor accuracy. Using 3D coordinates and the atom-type of surface protein atom as the input, we trained and tested a deep learning model to predict how likely a voxel on the protein surface is to be a DNA-binding site. Based on three different evaluation datasets, the results show that our model not only outperforms several previous methods on two commonly used datasets, but also demonstrates its robust performance to be consistent among the three datasets. The visualized prediction outcomes show that the binding sites are also mostly located in correct regions. We successfully built a deep learning model to predict the DNA binding sites on target proteins. It demonstrates that 3D protein structures plus atom-type information on protein surfaces can be used to predict the potential binding sites on a protein. This approach should be further extended to develop the binding sites of other important biological molecules.

Not everyone gets sick after an exposure to influenza A viruses (IAV). It was unclear whether forecasting symptomatic flu infection based on pre-exposure host gene expression might be possible. We utilized the human gene expression data infected with IAV subtype H1N1 or H3N2 viruses to examine this problem using the state-of-the-art deep learning approach. The results indicate that such forecast is possible and, in other words, gene expression could reflect the strength of host immune systems. Here we have developed DeepFlu—based on deep neural networks—to predict who would catch the flu prior to an exposure to IAV. We validated DeepFlu both on a leave-one-person-out internal cross-validation and an independent external cohort. DeepFlu performed better than the models using convolutional neural network, random forest, or support vector machine, achieving 70.0% accuracy, 0.787 AUROC, and 0.758 AUPR for H1N1 and 73.8% accuracy, 0.847 AUROC, and 0.901 AUPR for H3N2. Besides, DeepFlu which was trained only by pre-exposure data worked the best than by other time spans and mixed training data of H1N1 and H3N2 did not necessarily enhance prediction. These findings suggest that deep learning models could moderately recognize individuals susceptible to flu to prevent the spread of IAV. Author summary Influenza A, one of the leading infectious diseases, causes major health and socioeconomic burdens in the world. Hence, it is important to develop methods to accurately forecast who will succumb to influenza A virus (IAV) in advance. To demonstrate that forecasting influenza susceptibility is possible, we develop DeepFlu, based on the powerful deep learning method, to predict who will catch the flu based on ‘healthy’ pre-viral-exposure host gene expression. DeepFlu, which outperforms other machine learning methods, can be used to forecast susceptibility for two IAV subtypes H1N1 and H3N2. The results suggest that we could identify people at risk for the flu and alarm them in advance to lessen the spread of IAV and reduce the death tolls caused by the flu and its complications.

Whether there is any inclination between structures and functions of antimicrobial peptides (AMPs) is a mystery yet to be unraveled. AMPs have various structures associated with many different antimicrobial functions, including antibacterial, anticancer, antifungal, antiparasitic and antiviral activities. However, none has yet reported any antimicrobial functional tendency within a specific category of protein/peptide structures nor any structural tendency of a specific antimicrobial function with respect to AMPs. Here, we examine the relationships between structures categorized by three structural classification methods (CATH, SCOP, and TM) and seven antimicrobial functions with respect to AMPs using an enrichment analysis. The results show that antifungal activities of AMPs were tightly related to the two-layer sandwich structure of CATH, the knottin fold of SCOP, and the first structural cluster of TM. The associations with knottin and TM Cluster 1 even sustained through the AMPs with a low sequence identity. Moreover, another significant mutual enrichment was observed between the third cluster of TM and anti-Gram-positive-bacterial/anti-Gram-negative-bacterial activities. The findings of the structure–function inclination further our understanding of AMPs and could help us design or discover new therapeutic potential AMPs.