Zhuoyang Chen

Protein function prediction traditionally relies on structured gene ontology (GO) labels or multi-label classifiers. However, these labels or classifiers cannot flexibly describe molecular function, biological process, cellular component, and free-text functional narratives in a single output. In comparison, generation-based approaches offer an intuitive paradigm for flexible free-text protein annotation, with large language models (LLMs) as a representative method for protein-text modeling. Recent efforts on utilizing LLMs for protein semantic understanding and annotation generation have adopted sequence-only encoding or sequence-text contrastive alignment paradigms, yet without explicit consideration of three-dimensional structural information. To address these limitations in current protein function prediction methods, we present ProtBLIP2-SST, a two-stage framework built on the BLIP2 model architecture that bridges protein sequence, structure, and text for open-ended protein functional caption generation. Specifically, we first integrate sequence and structure information through SaProt, a protein language model (PLM) with a structure-aware vocabulary that fuses residue tokens with Foldseek-derived 3Di structural tokens. To empower the LLM to understand protein semantics, we employ a Q-Former (a querying transformer in BLIP2) with learnable query tokens as the cross-modal projector to align protein features from the frozen SaProt encoder and text features from a frozen BiomedBERT via protein–text contrasting, protein–text matching, and protein captioning objectives. After alignment, the protein features are linearly projected and prepended to the prompt embeddings of the LLM for protein captioning fine-tuning with LoRA. Trained on 441k protein–text pairs from Swiss-Prot with corresponding structures from the AlphaFold Database, our ProtBLIP2-SST outperforms sequence-only and sequence-text alignment baselines on protein captioning metrics, with ablation studies demonstrating the effectiveness of integrating structure with sequence information for improved protein understanding. Through a unified two-stage alignment-and-generation pipeline, ProtBLIP2-SST integrates protein sequence and structural information, overcomes the rigidity of traditional GO-centric classification, generating open-ended captions that jointly describe molecular function, subcellular location, and homology context in one single output.

Protein language models (pLMs) offer great potential for protein sequence analysis, yet the scarcity of labeled data often limits their effectiveness in fine-tuning. Data augmentation is a promising remedy, but systematic evaluation of augmentation strategies for protein sequences remains limited, and the conditions under which augmentation confers downstream benefits are not well understood. In this paper, we systematically investigate pLM-guided substitution-based augmentation across seven protein prediction tasks. We propose ProtAug, a framework that leverages encoder-based (ESM-2) and autoregressive (ProtGPT2) pLMs to generate augmented sequences with user-controlled variation levels. Our investigation focuses on four questions: (Q1) whether pLM-synthesized sequences preserve more original signals than simpler methods, (Q2) to what extent augmentation improves prediction performance, (Q3) how variation levels affect downstream accuracy across tasks and models, and (Q4) whether biological plausibility is a necessary condition for achieving improvement. Our experimental results show that: (1) ProtAug Esm generally preserves motifs and structural similarity better than simple substitution, often comparable to homology retrieval; (2) augmentation yields consistent but task-dependent improvements, with ProtAug Esm achieving the best or second-best performance in 5 out of 7 tasks at 10% variation; (3) low-to-moderate variation levels (2–30%) perform best overall, although high-variation augmentation can benefit certain structure-related tasks; (4) the necessity of biological plausibility is task- and variation-dependent—while semantic preservation correlates with performance at low-to-moderate variation levels, improved generalization at high variation levels suggests that regularization effects, rather than label preservation, can also drive performance gains.

Motivation: The three-dimensional protein tertiary structure alignment is a fundamental problem that seeks insights into functions and evolution. Previous structure alignment algorithms have adopted the sequential assumption and used dynamic programming solvers. However, many distantly related structures exhibit non-sequential similarities, and non-sequential alignment tools are less efficient and accurate than sequential ones. In this paper, we formulate the non-sequential alignment as the Entropy-regularized Partial Linear Sum Assignment Problem (epLSAP) and propose a solver based on Sinkhorn algorithms, referred to as epLSAP-Align. Results: Compared with existing non-sequential alignment solvers, our epLSAP-Align can explicitly model the gap penalty, efficiently achieve global optimality and balance coverage and fidelity. We show that epLSAP-Align can be easily integrated into the existing frameworks, such as TM-align and MICAN, resulting in the non-sequential alignment tool epLSAP-TM and epLSAP-MICAN, respectively. Both epLSAP-TM and epLSAP-MICAN achieve better performance than the existing non-sequential alignment tools in terms of biologically meaningful structure overlaps on two sequential alignment test sets MALIDUP and MALISAM, and four non-sequential alignment test sets MALIDUP-ns, MALISAM-ns, 64-difficult-case and RIPC datasets. Also, compared with the most recent non-sequential alignment tool USalign2, our epLSAP-TM is at least 22% faster under the same setting. Availability and implementation: Our source code is available at https://github.com/xzhangem/epLSAP-align.

In this study, we investigate the performance of nine protein structure alignment tools by analyzing the influence of alignment results in accomplishing three downstream biological tasks: homology detection, phylogeny reconstruction, and function inference. These tools include (1) traditional sequential methods using both 3D and 2D structure representations, (2) non-sequential methods, (3) flexible methods, and (4) deep-learning methods. Canonical sequence alignment methods Needleman-Wunsch algorithm and BLASTp are used as baseline methods. We show that accuracies in downstream tasks can be uncorrelated with alignment quality metrics such as TM-score and RMSD, highlighting the discrepancy between the alignment results and the purposes of using them. We identify scenarios where structure alignment results outperform sequence alignment results. In homology detection, structure-based methods are much better than sequence alignment. In phylogeny reconstruction, structure-based methods generally outperform sequence-based methods on the filtered dataset with proteins sharing low sequence similarity. Moreover, we show that structure information improves the overall performance of these tools when used together with sequence information in phylogeny reconstruction and function inference. We also test the running time and CPU/GPU memory consumption of these tools for a large number of queries. Our study suggests that biological problems that were previously addressed with sequence-based methods using only sequence information could be further improved by using structure information alone or using both sequence and structure information. The trade-off between task accuracy and speed is the major consideration in developing new alignment tools for downstream tasks. We recommend both TMalign and KPAX for these tasks because of their good balance between running time and memory consumption, and relatively good and stable accuracy performance in downstream tasks. In tasks that require a large number of pairwise comparisons, such as homology detection and function inference, traditional methods outperform DL methods at the cost of long running time, and Foldseek is the best choice to achieve relatively high accuracy in a reasonable time. 1 Graphical Fig. 1: Graphical

Protein-protein Interaction (PPI) networks are crucial for automatically annotating protein functions. As multiple PPI networks exist for the same set of proteins that capture properties from different aspects, it is a challenging task to effectively utilize these heterogeneous networks. Recently, several deep learning models have combined PPI networks from all evidence, or concatenated all graph embeddings for protein function prediction. However, the lack of a judicious selection procedure prevents the effective harness of information from different PPI networks, as these networks vary in densities, structures, and noise levels. Consequently, combining protein features indiscriminately could increase the noise level, leading to decreased model performance. We develop DualNetGO, a dual network model comprised of a classifier and a selector, to predict protein functions by effectively selecting features from different sources including graph embeddings of PPI networks, protein domain and subcellular location information. Evaluation of DualNetGO on human and mouse datasets in comparison with other network-based models show at least 4.5%, 6.2% and 14.2% improvement on Fmax in BP, MF and CC Gene Ontology categories respectively for human, and 3.3%, 10.6% and 7.7% improvement on Fmax for mouse. We demonstrate the generalization capability of our model by training and testing on the CAFA3 data, and show its versatility by incorporating Esm2 embeddings. We further show that our model is insensitive to the choice of graph embedding method and is time- and memory-saving. These results demonstrate that combining a subset of features including PPI networks and protein attributes selected by our model is more effective in utilizing PPI network information than only using one kind of or concatenating graph embeddings from all kinds of PPI networks. The source code of DualNetGO and some of the experiment data are available at: https://github.com/georgedashen/DualNetGO. Supplementary data are available at Bioinformatics online.

Motivation: Protein-protein Interaction (PPI) networks are crucial for automatically annotating protein functions. As multiple PPI networks exist for the same set of proteins that capture properties from different aspects, it is a challenging task to effectively utilize these heterogeneous networks. Recently, several deep learning models have combined PPI networks from all evidence, or concatenated all graph embeddings for protein function prediction. However, the lack of a judicious selection procedure prevents the effective harness of information from different PPI networks, as these networks vary in densities, structures, and noise levels. Consequently, combining protein features indiscriminately could increase the noise level, leading to decreased model performance. Results: We develop DualNetGO, a dual network model comprised of a classifier and a selector, to predict protein functions by effectively selecting features from different sources including graph embeddings of PPI networks, protein domain and subcellular location information. Evaluation of DualNetGO on human and mouse datasets in comparison with other network-based models show at least 4.5%, 6.2% and 14.2% improvement on Fmax in BP, MF and CC Gene Ontology categories respectively for human, and 3.3%, 10.6% and 7.7% improvement on Fmax for mouse. We demonstrate the generalization capability of our model by training and testing on the CAFA3 data, and show its versatility by incorporating Esm2 embeddings. We further show that our model is insensitive to the choice of graph embedding method and is time- and memory-saving. These results demonstrate that combining a subset of features including PPI networks and protein attributes selected by our model is more effective in utilizing PPI network information than only using one kind of or concatenating graph embeddings from all kinds of PPI networks. Availability and implementation: The source code of DualNetGO and some of the experiment data are available at: https://github.com/georgedashen/DualNetGO . Contact: luo@ust.hk , zchen190@connect.hkust-gz.edu.cn Supplementary Information: Supplementary data are available at Bioinformatics online.