Biology subjects

Siu, S. W. I.

Publications and source records attributed to Siu, S. W. I..

6 recordsLinked to original sources

bioRxiv2026

DecoyOpt: A two-stage pipeline for decoy-aware enzyme discrimination and EC number prediction

Accurate and efficient EC number prediction is essential for elucidating the catalytic roles of enzymes. Although similarity-based and deep learning-based approaches have advanced this task, they often fail to distinguish active enzymes from closely related decoys inactivated by local functional-site mutations, leading to high false-positive rates (FPRs). We propose DecoyOpt, a two-stage pipeline for decoy-aware enzyme discrimination and EC number prediction. Specifically, we generated putatively inactive functional-site decoys using ESM3 and evaluated 16 protein language models across three splits. An MLP trained with classification and ranking losses distinguishes enzymes from unannotated proteins and decoys before a contrastive learning model assigns EC numbers to predicted enzymes. On the most stringent GraphPart 40% split, DecoyOpt achieves an MCC of 0.899 and a D0 decoy FPR of 6.17%, outperforming retrained baselines. Across all splits, it outperforms these baselines in EC number prediction on the full test set while maintaining performance close to CLEAN on the enzyme subset. It also correctly annotates all four wild-type enzymes and rejects all eight inactive or nearly inactive variants. We believe that DecoyOpt has potential for annotating uncharacterized proteins and their variants, with applications in genomics, synthetic biology, and biocatalysis.

bioinformatics↗

bioRxiv2026

MAERM: Predicting Enzyme-Reaction Matching Relationships with a Mixed-Attention Model

Harnessing enzyme specificity requires a thorough understanding of enzyme promiscuity, which determines enzymes catalytic scope; however, measuring this scope still relies heavily on labor-intensive analytical approaches. While data-driven approaches have emerged to predict the catalytic scope of enzymes, these methods continue to face challenges such as restricted datasets and insufficient integration of enzyme structural information and reaction transformations. Here, we introduce MAERM, an innovative mixed-attention model designed to predict enzyme-reaction matching relationships. Built on our MAERM-DB, a dataset with broad coverage of validated and chemoenzymatic catalysis data, MAERM utilizes a local-global attention module to integrate multimodal enzyme information with fine-grained reaction representations, thereby predicting enzyme-reaction matching probabilities. Results show that MAERM consistently outperforms all baselines, with an average F1-score of 0.984. Notably, on challenging test samples with less than 40% sequence identity to the training set, MAERM outperforms the second-ranked model by 5.9% in F1-score. In addition, MAERM achieves the highest top-10 success rate of 51.7% on Enzyme-405 and the highest balanced accuracy of 0.697 on BioCat-547, further supporting its generalizability in enzyme screening and chemoenzymatic catalysis. Finally, MAERM can serve as an efficient scoring module. When integrated with ProteinMPNN, MAERM has successfully guided novel enzyme design for two carbonyl reduction reactions, resulting in enhanced catalytic potential for the native substrate and demonstrating broad compatibility. Overall, MAERM has the potential to reduce the experimental cost of measuring enzymes catalytic scope, facilitate enzyme design, and ultimately accelerate the design-build-test-learn cycle in enzyme engineering.

bioinformatics↗

bioRxiv2026

Protein-peptide Interaction Representation Learning with Pretrained Language Models

Protein-peptide Interactions (PpIs) paly essential roles in diverse cellular processes, yet their systematic identification remains challenging due to the limited availability of experimentally annotated protein-peptide interaction data. To address this challenge, we present PepInter, a sequence-based Deep Learning (DL) framework that leverages large-scale pretraining on structurally derived pseudo protein-peptide pairs to learn interaction-relevant representations. Specifically, energy-dominant peptide fragments are extracted from protein complexes curated from non-redundant Protein Data Bank (PDB) structures, enabling the construction of pseudo protein-peptide interaction pairs that capture interface interaction patterns shared with canonical protein-protein interactions. This strategy allows the model to acquire interaction-aware priors in the absence of large-scale annotated protein-peptide complex datasets. Built upon the ESM-Cambrian (ESMC) architecture, PepInter adopts a two-stage pretraining strategy. In the first stage, masked language modeling is used to learn general protein sequence representations. In the second stage, the model is further trained to predict Rosetta-derived energetic scores, explicitly incorporating structural interaction signals into the learned embeddings. Following pretraining, PepInter is fine-tuned for both protein-peptide interaction classification and peptide bioactivity regression tasks. Across multiple benchmark datasets, including protein-peptide binding affinity prediction, PepInter consistently outperforms existing baseline methods and demonstrates strong generalization in identifying biologically meaningful PpIs. Case studies further highlight its ability to recover known interaction patterns and predict novel protein-peptide interactions. Together, these results establish PepInter as a scalable and effective framework for protein-peptide interaction prediction, with strong potential to accelerate peptide-based drug discovery.

bioinformatics↗

bioRxiv2025

ClsDiff-AMP30: Generating Antimicrobial Peptides by a Classifier Guidance Noise Predictor

Antimicrobial peptides (AMPs) represent a promising therapeutic strategy to combat the increasing challenge of multidrug-resistant pathogens, a crisis intensified by the overuse of conventional antibiotics. In addition to their broad-spectrum antimicrobial activity, low toxicity, and reduced propensity for resistance development, AMPs offer significant advantages over traditional antibiotic therapies. However, the discovery of novel AMPs through biological experiments remains constrained by high costs, labor-intensive workflows, and time-consuming procedures, underscoring the urgent need for in silico computational methods to design AMP sequences. Notably, shorter AMPs ([≤] 30 residues) demonstrate superior antimicrobial efficacy, improved structural stability, and minimal cytotoxicity toward human cells. To address these challenges, we present a classifier-guided diffusion framework specialized for generating AMPs shorter than 30 residues (ClsDiff-AMP30). The architecture integrates two interdependent submodels, including a noisy AMP classifier that evaluates AMP likelihood at intermediate denoising steps and a noise predictor guided by classifier-derived probability scores, dynamically adjusted via a self-optimized coefficient to modulate guidance strength. ClsDiff-AMP30 achieves a validation accuracy of 66% across 10,000 synthesized sequences by a self-developed AMP classifier. Furthermore, wet lab experiments demonstrated that all 11 selected sequences exhibited high antimicrobial activity against at least one of the three tested bacterial strains and low hemolytic activity.

bioinformatics↗

bioRxiv2025

BERT-T6: Towards High-accuracy T6SS Bacterial Toxin Identification Using Protein Language Model

Type VI secretion system effectors target the cell wall, membranes and nucleic acids, leading to the killing of bacteria or impairment of host cell defense mechanisms. Accurate identification of T6SEs will be beneficial to understand the virulence of these bacteria via type VI secretion systems as well as bacterial pathogenesis. Although some traditional machine learning-based and deep learning-based tools have been developed to distinguish T6SEs from non-T6SEs, we believe there is still room for further improvement. To obtain the robust feature for model construction, we successively investigate various classic sequence-based features and embeddings from pre-trained transformer-based protein language models. Building upon the model incorporating ProtBert embeddings, we employed a transfer learning approach to fine-tune the ProtBert protein language model with a downstream T6SE classification task. The resulting BERT-T6 model demonstrates performance significantly superior to baseline models. More importantly, with an accuracy of 0.959, a sensitivity of 0.909, a specificity of 0.973, a precision of 0.905, a F1-score of 0.907, MCC of 0.881, our model achieves performance competitive with state-of-the-art binary and multi-class predictors. This work highlights the effectiveness of utilizing BERT with transfer learning for T6SE prediction. BERT-T6 provides a robust and precise approach for identifying T6SEs, offering promise for enhancing studies of bacterial virulence mechanisms.

bioinformatics↗

bioRxiv2021

Multi-Branch-CNN: classification of ion channel interacting peptides using parallel convolutional neural networks

Ligand peptides that have high affinity for ion channels are critical for regulating ion flux across the plasma membrane. These peptides are now being considered as potential drug candidates for many diseases, such as cardiovascular disease and cancers. There are several studies to identify ion channel interacting peptides computationally, but, to the best of our knowledge, none of them published available tools for prediction. To provide a solution, we present Multi-branch-CNN, a parallel convolutional neural networks (CNNs) method for identifying three types of ion channel peptide binders (sodium, potassium, and calcium). Our experiment shows that the Multi-Branch-CNN method performs comparably to thirteen traditional ML algorithms (TML13) on the test sets of three ion channels. To evaluate the predictive power of our method with respect to novel sequences, as is the case in real-world applications, we created an additional test set for each ion channel, called the novel-test set, which has little or no similarities to the sequences in either the sequences of the train set or the test set. In the novel-test experiment, Multi-Branch-CNN performs significantly better than TML13, showing an improvement in accuracy of 6%, 14%, and 15% for sodium, potassium, and calcium channels, respectively. We confirmed the effectiveness of Multi-Branch-CNN by comparing it to the standard CNN method with one input branch (Single-Branch-CNN) and an ensemble method (TML13-Stack). To facilitate applications, the data sets, script files to reproduce the experiments, and the final predictive models are freely available at https://github.com/jieluyan/Multi-Branch-CNN.

bioinformatics↗