Motivation:Protein function prediction is a challenging task and an open problem in computational biology. The Critical Assessment of protein Function Annotation (CAFA) is a triennial, community-driven initiative that provides an independent, large-scale evaluation of computational methods for protein function prediction through time-delayed benchmarking experiments. CAFA has played a key role in highlighting high-performing methodologies and fostering detailed analysis and exchange of ideas. However, outside the periodic CAFA challenges, there is no platform for the continuous evaluation of newly developed methods and tracking performance as function annotations accumulate. Results:Here we introduce the Longitudinal Assessment of Protein Function Annotation Models server (LAFA) as a persistent benchmarking system for protein function prediction methods. LAFA provides a continuous evaluation of containerized function prediction methods, enabling up-to-date and robust comparative assessment of method performance under evolving ground truth. LAFA accelerates methodological iteration, supports reproducibility, and offers a more dynamic and fine-grained view of progress in protein function prediction. Code and Data Availability:LAFA is available at https://functionbench.net/. Detailed evaluation results can be found at https://github.com/anphan0828/CAFA_forever. Contact:[ahphan|idoerg]@iastate.edu.

Single-particle cryo-electron microscopy (cryo-EM) has rapidly evolved from largely manual, stepwise workflows to AI-assisted, increasingly automated protein structure determination pipelines. This chapter presents an end-to-end view on our deep learning-based AI tools that address three major stages of cryo-EM data analysis: protein particle picking, density map enhancement, and atomic structural model building. First, we introduce two of our AI-based particle pickers: CryoTransformer, a transformer-residual architecture trained on the large-scale CryoPPP dataset, and CryoSegNet, which couples a cryo-EM-specialized attention-gated U-Net with SAM to segment particles across diverse proteins and imaging conditions. Together, they deliver high precision-recall performance and improve downstream 3D reconstructions, outperforming widely used pickers on independent benchmarks. Second, we describe CryoTEN, a 3D UNETR++-style transformer to enhance cryo-EM density maps. On a 150-map test set, CryoTEN robustly improves local interpretability and downstream protein structure modeling while running>10× faster than other deep learning methods. Third, we present our two atomic protein model building frameworks: Cryo2Struct and MICA. Cryo2Struct is a fully automated de novo modeling framework that identifies atoms and residue types from density maps alone via a 3D transformer and assembles them into chains with a Hidden Markov Model, producing more complete and accurate models than other ab initio tools. MICA, a multimodal framework that integrates cryo-EM density maps with AlphaFold3 structure predictions at both input and output levels to further refine model accuracy and completeness, achieving near-experimental TM-scores on recently released high resolution density maps. Finally, we conclude by outlining open challenges and potential future directions in cryo-EM data analysis.

Recent advances in biomolecular modeling have been catalyzed by models such as AlphaFold3 (AF3), which introduce science-informed changes to the transformer architecture. Unlike transformers, a defining characteristic of AF3-style models is their 3D attention over 2D pairwise representations which produces tensors whose computation and memory costs scale cubically with sequence length. As a result, despite moderate parameter counts, AF3-style models are far more expensive to train than size-equivalent transformers, and are severely constrained by GPU memory capacity. Our characterization shows 3D attention fundamentally changes the training workload, causing massive 3D attention maps, complex inter-operator dependencies, kernel fragmentation, and heavy host-side data pipelines which differ substantially from LLM training, leading to poor utilization on modern GPU systems. Moreover, existing GPU optimizations do not adequately address these challenges due to complex cross-layer inter-operator dependencies introduced by 3D attention. Motivated by these challenges, we introduce MegaFold, a novel cross-platform system for efficient training of next-generation 3D-attention protein models. MegaFold combines a memory-efficient 3D-attention kernel, a communication-efficient sharding strategy for quadratic representations, fused operator implementations for critical execution paths, and a determinism-aware host-device pipeline that eliminates preprocessing stalls. Evaluation on both NVIDIA H200 and AMD MI250 GPUs shows that MegaFold enables training with up to 3.36× longer sequence lengths on 32 GPUs while reducing end-to-end execution time by up to 1.73× (NVIDIA) and 1.62× (AMD).

AlphaFold3 introduces a unified framework for predicting the structures and interaction of several biomolecules including single-chain protein monomers, multi-chain protein multimers, and protein-ligand complexes. While it achieves the state-of-the-art performance in most predictions, its prediction accuracy depends on the quality of multiple sequence alignment (MSA) and structural template inputs. There are few works of using customized MSAs and templates to improve AlphaFold3. In this work, we systematically investigate how diverse and carefully engineered MSAs and templates can be leveraged to improve AlphaFold3 predictions. We evaluate our methods on protein monomers, multimers, and protein-ligand complexes, and observe consistent, sizable gains in structure prediction accuracy for monomers (TM-score 0.937 vs 0.882), for multimers (DockQ score 0.550 vs 0.525), and for protein-ligand complexes (ligand RMSD 3.258 Å vs 4 Å) compared to the default AlphaFold3. Moreover, for the first time, we demonstrate that AlphaFold3 performs significantly better than AlphaFold2 when both use the same customized MSA and template inputs. The results highlight the importance and effectiveness of using diverse MSAs and templates to improve AlphaFold3.

Motivation Deciphering gene regulatory networks (GRNs) from single-cell transcriptomics data remains a fundamental challenge in computational biology. It is hindered by data sparsity, high dimensionality, and the lack of scalable, generalizable inference models. To address this, we present GRNFormer, a generalizable graph transformer framework for accurate GRN inference from transcriptomics data across species, cell types, and platforms without requiring cell-type annotations or prior regulatory information.Results GRNFormer integrates a transformer-based gene expression encoder (Gene-Transcoder) with a variational graph autoencoder (GraViTAE) employing pairwise attention to jointly learn the representations of genes (nodes) and their co-expression relationships (edges). Leveraging TF-Walker, a transcription factor-anchored subgraph sampling strategy, it effectively captures gene regulatory interactions from either single-cell or bulk RNA-seq datasets. Benchmarking on standard datasets demonstrates that GRNFormer outperforms existing traditional and deep learning state-of-the-art methods in blind evaluations, achieving average sampled area under the receiver operating characteristic curve (Sampled_AUROC) and sampled area under the precision-recall curve (Sampled_AUPRC) values between 0.90 and 0.98 as well as 0.87-0.98 average sampled F1 score. The model robustly recovers both known and novel regulatory networks, including pluripotency circuits in human embryonic stem cells (hESCs) and immune cell modules in peripheral blood mononuclear cells (PBMCs). The architecture enables scalable, biologically interpretable GRN inference across various datasets, cell types, and species, establishing GRNFormer as a robust and transferable tool for network biology.Availability and implementation GRNFormer is available on GitHub (https://github.com/BioinfoMachineLearning/GRNformer); the version used in this work is archived on Zenodo (https://doi.org/10.5281/zenodo.18868395), with evaluation resources for reproducibility.

Protein complexes carry out many essential functions in cells, but predicting their structures requires knowing their stoichiometry, meaning how many copies of each subunit are present. This information is often unavailable, making stoichiometry prediction crucial for complexes with unknown stoichiometry. Despite its importance, few computational methods address this challenge. Here we show that combining AlphaFold3 structure prediction with information from related known protein complexes enables accurate prediction of protein complex stoichiometry. Our method generates plausible subunit combinations, builds structural models for them using AlphaFold3, ranks them using AlphaFold3 scores, and further refines predictions with template-based information when available. In the 16th community-wide Critical Assessment of Techniques for Protein Structure Prediction, our method identifies the correct stoichiometry as the top prediction for 71.4% of targets and among the top three predictions for 92.9% of targets, outperforming other methods overall. This demonstrates the complementary strengths of AlphaFold3- and template-based predictions and highlights the applicability of our approach to uncharacterized protein complexes lacking stoichiometry data.

Abstract Transcription factors (TFs) are central regulators of gene expression, and their selective recognition of genomic DNA underlies various biological processes. Experimental profiling of TF–DNA interactions using chromatin immunoprecipitation followed by sequencing (ChIP-seq) provides high-resolution maps of in vivo TF-DNA binding but remains costly, labor-intensive, and inherently low-throughput, limiting their scalability across different transcription factors, cell types, and regulatory conditions. Computational modeling therefore plays an essential role in inferring TF–DNA interactions at genome scale. However, most existing computational models rely solely on DNA sequence and chromatin features to predict TF–DNA binding, neglecting TF-specific protein information. This omission limits their ability to capture protein-dependent binding specificity. Here, we present TFBindFormer, a hybrid cross-attention transformer that explicitly integrates genomic DNA features with TF-specific representations derived from protein sequences and structures. By modeling protein-conditioned, position-specific TF–DNA interactions, TFBindFormer enables direct learning of molecular determinants underlying DNA recognition. Evaluated across hundreds of cell-type–specific TFs and hundreds of millions of genome-wide DNA bins, TFBindFormer consistently outperforms DNA-only baselines, achieving substantial gains in both area under precision-recall curve (AUPRC) and area under receiver operating characteristic curve (AUROC). Together, these results demonstrate that integrating TF and DNA features via cross-attention enables TFBindFormer to serve as an effective and scalable framework for large-scale TF–DNA binding prediction.

Background:The Critical Assessment of Functional Annotation (CAFA) is a community effort held to understand the field of computational protein function prediction. Every three years, since 2010, the organizers initiate an experiment to collect function predictions on a large set of proteins and then evaluate the performance of predicting methods on a subset of proteins that have accumulated experimental annotations between the submission deadline and the evaluation time. CAFA provides an independent and rigorous assessment of the current state of the art, thus leveling the playing field, highlighting successes, revealing bottlenecks, and offering a forum for the exchange of ideas in protein science. Here, we report the results of the fourth CAFA experiment (CAFA4). Results:CAFA4 featured the participation of 148 methods from 70 research groups on a total of 46,205 unique proteins over a 5-year annotation accumulation phase, the longest in any CAFA. In a comparison across CAFA2-CAFA4 methods, the prediction of Gene Ontology (GO) terms has clearly improved across all three GO aspects and traditional evaluation settings. While not achieving the first rank, several CAFA2 and CAFA3 methods featured in the top ten methods in many evaluations, suggesting that earlier methods still hold relevance. The performance is weaker in the newly introduced "partial knowledge" evaluation category (proteins with experimental annotations before submission deadline that gained additional annotations in the same GO aspect during the annotation accumulation phase), highlighting the need for a new class of methods. The rankings of the methods were stable over the years in traditional evaluation settings, but less so in the new partial knowledge evaluation. Overall, the field continues to progress with some influx of new participants. Sustained efforts will be necessary to substantially advance it.

Alzheimer's Disease (AD) remains a leading cause of cognitive decline with no known cure, motivating the development of therapies that slow neurodegeneration. Rapamycin, an FDA-approved inhibitor of the mammalian target of rapamycin (mTOR) pathway, has demonstrated promising anti-aging and neuroprotective effects. However, characterizing its treatment effects and identifying the biological factors that contribute to treatment response remain challenging because of complex interactions across multiple biological systems and the limited availability of patient data. In this work, we propose a three-stage multimodal deep learning framework called TreatmentFormer for predicting rapamycin treatment status from heterogeneous biomedical data including both brain imaging data and tabular data (e.g., microbiome profiles, blood-based biomarkers, cerebral blood flow measurements, and clinical variables (e.g., gender, age, and body mass index)). First, a Random Forest-based feature selection module reduces noise in high-dimensional tabular data while preserving representation across modalities. Second, modality-specific encoders map imaging and tabular inputs into a shared latent space via self-supervised contrastive learning, enabling alignment across modalities. Finally, a transformer-based architecture integrates these representations to capture cross-modal interactions and perform treatment classification. Evaluated on a cohort of 23 participants with baseline and post-treatment timepoints, TreatmentFormer achieves an average prediction accuracy of 71.25\% across 10 independent test runs. Despite the challenges of small sample size and heterogeneous data, the model demonstrates stable and consistent performance. Post hoc SHAP-based feature analysis further identifies key biomarkers associated with treatment response, particularly within blood-based and inflammatory modalities. These findings demonstrate that combining feature selection with multimodal representation learning provides a promising and robust approach for modeling treatment effects in small-sample biomedical studies. Importantly, this framework may have significant implications for clinical research and medical applications by identifying the biological features and quantitative measurements that drive individual responses to rapamycin. Such insights could facilitate the development of predictive biomarkers, improve patient stratification, and ultimately inform future approaches to AD diagnosis and therapeutic development.

Accurate identification of protein particles in cryo-electron microscopy (cryo-EM) micrographs is crucial for high-resolution structure determination, but remains challenging due to the heavy reliance on extensive annotated datasets and the difficulty of ensuring robustness under low signal-to-noise ratio (SNR) conditions. Current approaches require large annotations and exhibit poor generalization to new protein targets. We present CryoFSL (Cryo-EM Few Shot-Learning), a novel few-shot learning framework built on Segment Anything Model 2 with lightweight adapters, enabling robust particle picking with as few as five labeled micrographs and significantly reducing the annotation burden. The framework's hierarchical adapter design supports dynamic feature modulation for low-SNR and heterogeneous conditions, resolving the trade-off between annotation burden and performance. CryoFSL surpasses both traditional template-based methods and state-of-the-art deep learning models across diverse proteins in the few-shot learning setting, achieving superior recall, precision, and 3D reconstruction resolution with minimal supervision. It maintains stability across heterogeneous micrographs and consistently detects high-quality particles with fewer false-positives. Notably, CryoFSL achieves competitive resolution in density map reconstruction with just a fraction of the particles picked by other methods, redefining efficiency and quality in cryo-EM analysis. This work paves the way for scalable, generalizable, and annotation-efficient particle-picking pipelines. The code is available at https://github.com/biplabpoudel25/CryoFSL.

The apolipoprotein E ε 4 (APOE4) allele is the strongest genetic risk factor for late-onset Alzheimer's disease (AD), the most common form of dementia. APOE4 carriers exhibit cerebrovascular and metabolic dysfunction, structural brain alterations, and gut microbiome changes decades before the onset of clinical symptoms. Better understanding of the early manifestion of these physiological changes is critical for development of timely AD interventions and risk reduction protocols. Multi-modal datasets encompassing a wide range of APOE ε 4 and AD associated biomarkers provide a valuable opportunity to gain insight into the APOE4 phenotype; however, these datasets often present analytical challenges due to small sample sizes and high heterogeneity. Here, we propose a two-stage multimodal AI model (APOEFormer) that integrates blood metabolites, brain vascular and structural MRI, microbiome profiles, and other clinical and demographic data to predict APOE4 allele status. In the first stage, modality-specific encoders are used to generate initial representa-tions of input data modalities, which are aligned in a shared latent space via self-supervised contrastive learning during pretraining. The contrastive learning objective encourages learning of informative and consistent representations across modalities through leveraging cross-modality relationships. In the second stage, the pretrained representations are used as inputs to a multimodal transformer that integrates information across modalities to predict a key AD-risk genetic variant (APOE4). Across 10 independent experimental runs with different train-validation-test splits, APOEFormer predicts whether an individual carries an APOE4 allele with an average prediction accuracy of 75%, demonstrating robust performance under limited sample sizes. Post hoc perturbation analysis of the predictive model revealed valuable insights into the driving components of the APOE4 phenotype- including key blood biomarkers and brain regions strongly associated with APOE4.

Transcription factors (TFs) regulate gene expression by binding to specific DNA sites on genome, making accurate TF binding site prediction critical for understanding gene regulation and downstream phenotypes. Almost all current deep learning methods use only DNA-related information to predict TF binding sites, ignoring the fact that different TF protein sequences and structures recognize distinct DNA patterns. Not leveraging TF information not only limits prediction accuracy but also makes the methods not generalizable to predicting binding sites of new TFs that do not exist in the training data. Here, we present TransBind, a protein-aware deep learning architecture that integrates DNA sequence information with protein embeddings containing both sequence and structural information derived from a protein language model pretrained on DNA-binding proteins, to improve TF binding site prediction. Through the cross-attention, a TF embedding selectively attends to genomic regions according to its unique binding properties. Evaluated on the data of 690 ChIP-seq experiments spanning 161 TFs across 91 human cell types, TransBind achieves an AUROC of 0.9508 and AUPR of 0.3741-representing a [Formula: see text]11.8% relative AUPR improvement over state-of-the-art methods including TBiNet, EPBDXDNABERT-2, DanQ, and DeepSEA. The model outperformed existing methods in [Formula: see text]98% of TF-cell type combinations. It also recovered 160 known TF binding motifs in the JASPAR database, providing the biological interpretability of the model. Moreover, the approach enables label-zero-shot prediction for unseen TFs, demonstrating its potential of generalizing to new, poorly characterized TFs. The source code of TransBind is available at https://github.com/jianlin-cheng/TransBind. The version used in this work is archived at https://doi.org/10.5281/zenodo.19462292.

Recent advances in artificial intelligence have introduced novel methods for high-accuracy prediction of protein tertiary structures, protein complex structures, and interactions between proteins and other biomolecules, such as small molecules and nucleic acids. Such advancements are accelerating biomedical research and the development of new protein design and bioengineering methods among many other important biotechnology applications. In this review, we outline the recent advances in protein-centric biomolecular structure and interaction prediction, highlight some major challenges in the field, and discuss potential directions to address them.

The integration of artificial intelligence, protein engineering, and sustainable nanomedicine is driving a paradigm shift in theranostics by enabling highly precise disease diagnosis and targeted therapy. AI-driven methodologies, including machine learning and deep learning, facilitate the rapid analysis of complex biological and chemical datasets, accelerating protein structure prediction, molecular docking, and structure-activity relationship modeling. These capabilities support the rational design of proteins and peptides with enhanced specificity, therapeutic efficacy, and safety, while enabling personalized treatment strategies tailored to individual molecular profiles. In parallel, sustainable nanomedicine focuses on the development of biodegradable, biocompatible, and environmentally benign nanomaterials to improve drug bioavailability, stability, and controlled release. AI-assisted optimization further refines nanocarrier design by balancing therapeutic performance with safety and environmental impact. Advanced intelligent nanocarriers capable of real-time monitoring, adaptive drug release, and degradation into non-toxic by-products represent a significant advancement over conventional static systems. The theranostic paradigm has become central to precision medicine, particularly in oncology, especially where AI-designed nanoplatforms enable targeted delivery of imaging agents and therapeutics to tumors, while allowing continuous treatment monitoring and minimizing off-target effects. Emerging applications in neurological, infectious, and cardiovascular diseases further highlight the broad clinical potential of this approach. Accordingly, this review summarizes AI-driven protein design strategies, sustainable nanocarrier engineering, and their convergence in next-generation theranostic systems, critically discussing mechanistic insights, translational challenges, and design principles required for developing safe, scalable, and clinically adaptable intelligent nanomedicines.

Motivation: Protein folding is a dynamic process during which a protein's amino acid sequence undergoes a series of 3-dimensional (3D) conformational changes en route to reaching a native 3D structure; the resulting 3D structural conformations are called folding intermediates. While data on native 3D structures are abundant, data on 3D structures of non-native intermediates remain sparse, due to limitations of current technologies for experimental determination of 3D structures. Yet, analyzing folding intermediates is crucial for understanding folding dynamics and misfolding-related diseases. Hence, we search the literature for available (experimentally and computationally obtained) 3D structural data on folding intermediates, organizing the data in a centralized resource. Additionally, we assess whether existing methods, designed for predicting native structures, can also be utilized to predict structures of non-native intermediates. Results: Our literature search reveals six studies that provide 3D structural data on folding intermediates (two for post-translational and four for co-translational folding), each focused on a single protein, with 2-4 intermediates. Our assessment shows that an established method for predicting native structures, AlphaFold2, does not perform well for non-native intermediates in the context of co-translational folding; a recent study on post-translational folding concluded the same for even more existing methods. Yet, we identify in the literature recent pioneering methods designed explicitly to predict 3D structures of folding intermediates by incorporating intrinsic biophysical characteristics of folding dynamics, which show promise. This study assesses the current landscape and future directions of the field of 3D structural analysis of protein folding dynamics.

Cryo-electron tomography (cryo-ET) enables in situ three-dimensional visualization of many protein complexes and other macromolecular assemblies such as ribosomes in cells, yet automated macromolecule particle identification in 3D cryo-ET tomograms remains a major bottleneck due to dose-limited low signal-to-noise ratios, missing-wedge artifacts, and densely crowded cellular backgrounds. We present TomoSwin3D, an end-to-end three-dimensional (3D) macromolecule particle identification and classification pipeline centered on a Swin Transformer-based U-Net that performs particle identification and classification and outputs particle centroid coordinates. TomoSwin3D leverages a multi-channel input representation that augments raw tomogram densities with complementary 3D feature maps capturing edge strength (Sobel gradients), local contrast enhancement (morphological top-hat), and multiscale blob responses (Difference-of-Gaussians), improving detectability of small and low-contrast targets. To better preserve particle geometry and avoid hand-crafted shape assumptions, it adopts occupancy-preserving supervision that directly uses available 3D instance masks rather than heuristic Gaussian/spherical labels and applies scalable patch-wise inference followed by lightweight post-processing (connected-component analysis, size filtering, centroid extraction) for robust centroid coordinate extraction. Across diverse simulated and experimental cryo-ET tomogram benchmarks including SHREC 2021 and 2020 test datasets, EMPIAR dataset, and Cryo-ET data portal dataset, TomoSwin3D achieves strong and consistent performance in detecting proteins and other particles, outperforming existing methods, with a pronounced advantage in picking hard, small protein particles. These results establish TomoSwin3D as a scalable and accurate solution for high-throughput cryo-ET macromolecule particle picking and downstream subtomogram averaging.

Abstract Motivation Protein function prediction is a challenging task and an open problem in computational biology. The Critical Assessment of protein Function Annotation (CAFA) is a triennial, community-driven initiative that provides an independent, large-scale evaluation of computational methods for protein function prediction through time-delayed benchmarking experiments. CAFA has played a key role in highlighting high-performing methodologies and fostering detailed analysis and exchange of ideas. However, outside the periodic CAFA challenges, there is no platform for the continuous evaluation of newly developed methods and tracking performance as function annotations accumulate. Results Here we introduce the Longitudinal Assessment of Protein Function Annotation Models server (LAFA) as a persistent benchmarking system for protein function prediction methods. LAFA provides a continuous evaluation of containerized function prediction methods, enabling up-to-date and robust comparative assessment of method performance under evolving ground truth. LAFA accelerates methodological iteration, supports reproducibility, and offers a more dynamic and fine-grained view of progress in protein function prediction. Code and Data Availability LAFA is available at https://functionbench.net/. Detailed evaluation results can be found at https://github.com/anphan0828/CAFA_forever.

Cryogenic Electron Tomography (cryo-ET) is an emerging experimental technique to visualize cell structures and macromolecules in their native cellular environment. Accurate segmentation of cell structures in cryo-ET tomograms, such as cell membranes, is crucial to advance our understanding of cellular organization and function. However, several inherent limitations in cryo-ET tomograms, including the very low signal-to-noise ratio, missing wedge artifacts from limited tilt angles, and other noise artifacts, collectively hinder the reliable identification and delineation of these structures. In this study, we introduce ETSAM - a two-stage Segment Anything Model 2 (SAM2)-based fine-tuned AI method that effectively segments cell membranes in cryo-ET tomograms. It is trained on a diverse dataset comprising 83 experimental tomograms from the CryoET Data Portal (CDP) database and 28 simulated tomograms generated using PolNet. ETSAM achieves state-of-the-art performance on an independent test set comprising 10 experimental tomograms for which ground-truth annotations are available. It robustly segments cell membranes with high sensitivity and precision, significantly outperforming existing deep learning methods.

With the advancements in instrumentation, image processing algorithms, and computational capabilities, single-particle cryo-electron microscopy (cryo-EM) has achieved atomic resolution in determining the 3D structures of viruses. The virus structures play a crucial role in studying their biological function and advancing the development of antiviral vaccines and treatments. Despite the effectiveness of artificial intelligence (AI) in general image processing, its development for identifying and extracting virus particles from cryo-EM micrographs has been hindered by the lack of manually labeled high-quality datasets. To fill the gap, we introduce CryoVirusDB, a labeled dataset containing the coordinates of expert-picked virus particles in cryo-EM micrographs. CryoVirusDB comprises 9941 micrographs from nine datasets representing seven distinct non-enveloped viruses exhibiting icosahedral or pseudo-icosahedral symmetry, along with coordinates of 339,398 labeled virus particles. It can be used to train and test AI and machine learning (e.g., deep learning) methods to accurately identify virus particles in cryo-EM micrographs for building atomic 3D structural models for viruses.

Motivation Predicting protein function is a fundamental and challenging task that requires integrating diverse biological data modalities to capture complex functional relationships. Traditional machine learning methods often rely on single modalities or combine only a limited number (typically two), without aligning them in a unified representation, thereby constraining predictive accuracy. Moreover, most existing machine learning approaches are limited to preselected subsets of Gene Ontology (GO) function terms with sufficient annotations, making the prediction of novel function terms a persistent challenge.Results Here, we present FunBind, a multimodal AI model that jointly learns from five modalities, i.e., protein sequences, textual descriptions, domain annotations, structures, and GO terms, to enhance prediction accuracy and infer previously unseen functions. FunBind operates in two modes: (1) self-supervised pretraining using contrastive learning to align the sequence modality with other heterogeneous modalities in a unified latent space, enabling unsupervised zero-shot function prediction, and (2) supervised fine-tuning of the pretrained model to leverage all non-function modalities for comprehensive and accurate function classification. Our results show that FunBind's zero-shot capabilities allow it to generalize effectively to novel function terms never encountered before, while its joint multimodal fine-tuning strategy outperforms single-modality models and current state-of-the-art deep learning methods in typical function prediction settings.Availability https://github.com/jianlin-cheng/FunBind