Navigating the uncharted: AI-driven advances in protein structure, dynamics, interactions and ligand interactions for understudied families Authors: Shivani Singh, Ruhani Singh, Sunita Sharma Published: 2026 ABSTRACT AbstractThe structural and functional characterization of lesser-known protein families remains a major challenge in modern computational biology. Recent breakthroughs in artificial intelligence (AI) and machine learning (ML) have rapidly advanced these fields. This is particularly evident in overcoming the limitations of traditional modelling approaches, such as classical homology modelling and ab initio folding methods. This review traces the evolution of conventional methods to cutting-edge deep learning frameworks, such as AlphaFold2, RoseTTAFold, and transformer-based architectures. We explore how these AI-driven tools achieve near-experimental accuracy in structure prediction, model protein dynamics, and intrinsic disorder. We also addressed computational approaches for protein-protein interactions (PPIs), central to cellular function and interface-targeted drug design, alongside protein-ligand interactions, including novel generative methods. Two representative case studies targeting orphan G-protein-coupled receptors and intrinsically disordered regions demonstrate the transformative potential of these techniques for previously intractable systems. Despite these advances, significant challenges remain, including the need for experimental validation, effective modelling of protein flexibility, and ethical considerations surrounding AI-generated data. We also compare classical and AI-based structural biology pipelines, summarize key tools (e.g. transformers, graph neural networks, and diffusion models), and offer best practice guidelines for computational modelling and data visualization. These developments provide unprecedented insights into the dark proteome regions of the protein universe, enabling the structural illumination of previously uncharacterized and understudied proteins where structure or function was unknown. This review aims to serve as a roadmap for researchers seeking to harness AI innovations to tackle some of the most challenging aspects of proteomics.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s13040-026-00577-7.Keywords: Protein structure prediction, Protein-protein interactions. Computational biology, Machine learning, Artificial intelligence, AlphaFold2, AlphaFold3, RoseTTAFold, Intrinsically disordered proteins (IDPs), Dark proteome Introduction Proteins are fundamental biomolecules composed of linear sequences of amino acids [1]. The three-dimensional structures of proteins are essential for understanding their biological functions and guiding their applications in drug design [2]. However, determining structures using X-ray crystallography, NMR spectroscopy, or cryo-electron microscopy remains laborious, costly, and time-consuming [3]. To date, structures have been determined for only a small fraction of known protein sequences. According to the Swiss model repository, only approximately 25% of all sequences in the human proteome have been determined using structural experiments [4, 5]. In contrast, genome and proteome sequencing projects have identified millions of protein sequences, the majority of which have not yet been structurally characterized. This growing disparity between known sequences and solved structures has driven the development of computational biology techniques [3]. To benchmark the computational prediction methods developed to address this gap, the biannual Critical Assessment of Structure Prediction (CASP) competition has provided a blind, community-wide test of prediction accuracy since 1994 [6]. Computational structure prediction approaches have historically fallen into two broad categories [7]. Physics-based models attempt to simulate protein folding through molecular interactions and thermodynamic behavior. In principle, this is elegant, but in practice, these simulations are computationally expensive and remain feasible only for very small proteins. Consequently, many bioinformatics methods have been developed as knowledge-based methods that utilize evolutionary information and template structures. These include homology modelling or threading based on structurally similar solved proteins and co-evolution analysis from multiple sequence alignments (MSA) [8–11]. However, this landscape is beginning to shift with increasing supercomputer capabilities and the emergence of ML-based simulations of complex molecules. The progress in bioinformatics has been closely tied to the steady growth of structural and sequence databases worldwide. They provide both template structures for modelling and evolutionary data for comparison [12]. In parallel, machine learning approaches, particularly neural networks, have been trained on this expanding dataset to improve prediction accuracy [10, 13]. However, prior to recent developments, most of these methods produced unreliable predictions for proteins lacking close homologues (typically <~30% sequence identity) [2]. In the last decade, artificial intelligence (AI) and deep learning (DL) have transformed the field (section. “Evolution of structure prediction methods”). The first breakthrough occurred at CASP13 in 2018, when DeepMind’s AlphaFold (now referred to as AlphaFold1) demonstrated that deep learning could achieve near-experimental prediction quality for the most challenging targets [14]. AlphaFold1 uses a convolutional neural network (CNN) to predict inter-residue distance distributions from MSA-derived features. The predicted distributions were then converted into 3D structures using gradient-descent optimization. In a test at CASP13, AlphaFold1 exceeded the performance of all competing methods, achieving a Global Distance Test-Total Score (GDT-TS) of approximately 60 for the most difficult free-modeling target. GDT-TS is the principal accuracy metric used in CASP, ranging from 0 to 100, with higher scores indicating closer agreement with the experimental structure [15]. This was the first demonstration that deep learning could substantially outperform established methods in protein structure prediction [14]. Building on this, AlphaFold2 introduced a fundamentally different architecture for CASP14 in 2020. It employs transformer-based attention mechanisms to directly predict three-dimensional coordinates in an end-to-end fashion [16, 17]. AlphaFold2‘s success stems from its ability to achieve prediction accuracy comparable to or exceeding that of experimental methods. It notably outperformed all previous computational methods and enabled accurate 3D structure prediction of many globular proteins (section “Emergence of AI and deep learning in structural biology”). Shortly after, the Baker Laboratory released RoseTTAFold, a three-track network that approached AlphaFold2-level accuracy while running on a modern CUDA GPU; its architecture and accessibility are detailed in (section “Emergence of AI and deep learning in structural biology”) [18]. In 2022, the journal Nature Methods named protein structure prediction as the “Method of the Year 2021”, (“Method of the Year 2021: Protein Structure Prediction”, 2022), and in 2024, Demis Hassabis and John Jumper received the Nobel Prize in Chemistry for their work on AlphaFold. The AlphaFold Protein Structure Database, which contains nearly 214 million predicted protein structures [19], was recently released by the DeepMind group and the European Bioinformatics Institute. Additionally, the source code for both AlphaFold2 and RoseTTAFold is now publicly available. Community efforts have further broadened access: OpenFold, a fully open-source and trainable reimplementation of AlphaFold2 from the AlQuraishi laboratory, reproduced AlphaFold2-level accuracy and released its training data, and it remains actively developed [20]. The scale and inference speed of protein language models (PLMs) have rapidly increased. These transformer-based models can predict 3D structures from a single protein sequence, by passing the need for multiple sequence alignments (MSAs). For instance, ESMFold uses a 15-billion-parameter language model to generate atomic-resolution structures, enabling the prediction of nearly 618 million protein structures [21] on a metagenomic scale. This MSA-free approach is particularly valuable for orphan proteins from understudied families that lack sufficient homologues for deep alignment construction. DeepMind introduced AlphaFold3 in 2024, which employs a diffusion-based architecture to predict the joint structure of complexes. These complexes include proteins, nucleic acids, small molecules, ions, and modified residues, all within a unified framework [22]. AlphaFold3 has demonstrated better structural accuracy than specialized tools across several benchmarks [22]. Despite these advances, significant challenges remain. Key challenges include modelling proteins that lack a fixed structure, accounting for conformational flexibility, and capturing transitions between multiple conformational states. Additional difficulties arise with membrane proteins and proteins from poorly studied families, where available data are limited. This review examines recent advances in AI and machine learning for protein structure prediction, with a particular focus on their application to poorly studied and uncharacterized protein families, collectively known as the ‘dark proteome.’ The last 15 years have seen an explosion of new architectures and algorithms using deep neural networks (i.e., residual and attention-based models, graph networks, and diffusion models) and their application in predicting protein structures, simulating protein dynamics, modelling protein-protein interactions, and simulating small-molecule binding. We describe how evolutionary information and large-scale datasets were used to train these models, how large language models contribute to structure prediction and how AI - based approaches compare with traditional methodologies. We also survey the rapid development of an open-source ecosystem of prediction tools (Boltz, Chai-1, OpenFold, and D-I-TASSER), the emerging area of AI-based protein dynamics and ensemble methods, frameworks for evaluating prediction quality [23–26] and computational approaches to predicting protein-protein and protein-ligand interaction (section “Protein-protein complex prediction” and Classical Docking and Scoring Functions). Beyond single structures, predicting how proteins assemble into complexes is equally central to cellular function and to designing drugs that target interaction interfaces; we therefore give particular attention to AI-based protein-protein interaction prediction (section “Protein-protein complex prediction”). To illustrate the capabilities and limitations of these tools, we include two case studies: orphan G-protein-coupled receptors (GPCRs), a pharmacologically important but understudied family of folded proteins, and intrinsically disordered proteins (IDPs), which lack a single fixed structure. Together these cases bracket the central challenges AI methods face, modeling of under characterized ordered proteins on one hand, and capturing disorder and conformational flexibility on the other, rather than serving as an exhaustive survey of dark-proteome categories. This is a review focus on AI/ML structure-prediction methods (between 2018 and 2025), prioritizing primary methods papers, benchmark studies, and case studies relevant to understudied and dark-proteome families Evolution of structure prediction methods Traditional approaches Two principal categories underpin traditional structural prediction methodologies. First, template-based modelling (TBM) builds on experimentally solved homologous structures. Second, template-free modelling predicts folds without known structural references. TBM approaches rely on sequence alignment to identify templates, whereas co-evolutionary analysis of related sequences can provide additional structural constraints [11, 27]. As the sequence identity drops further below the homology threshold, the model quality deteriorates rapidly owing to unreliable alignments and structural divergence [27]. The threading (or fold recognition) technique improves upon TBM by using structural library alignments across a wider range of sequence identities. It also uses scoring functions that measure the structural compatibility between proteins [27]. However, both techniques are based on the assumption that proteins adopt folds previously observed in the Protein Data Bank (PDB) [12]. As the number of structures in the PDB has increased, TBM has become increasingly applicable. Empirical studies have demonstrated that each newly solved structure can be used to model tens of sequences per fold [27]. Structural modelling servers, including SWISS-MODEL [4], Phyre2 [28], and HHpred/I-TASSER [11], rely on profile-profile or hidden Markov model alignments to identify such templates. However, once the sequence identity falls into this “twilight zone,” alignment errors and structural divergence accumulate, and model accuracy plummets [2]. Template-based methods are highly effective when good templates exist, but are inherently limited by the structural coverage of known proteins. When no suitable template exists, ab initio (de novo) methods attempt to predict protein folds from first principles. These template-free approaches are particularly relevant when no detectable structural homologues are present. These methods include fragment assembly (e.g., Rosetta and QUARK [29]) and physical simulations. Rosetta’s classic approach assembles proteins by sampling short peptide fragments from unrelated structures and recombining them via Monte Carlo simulations, guided by statistical and physical energy functions [30]. Later versions of Rosetta added atomic-level refinement and the option of including experimental data as restraints in the refinement process. In CASP competitions, Rosetta was able to generate accurate models for small proteins (under 100 residues) but struggled with large and more complex structures. In CASP12, researchers added co-evolutionary contact predictions to improve the Rosetta results. This hybrid model improved the results and outperformed the best crystal structure templates for many protein families. This has helped bridge the gap for proteins that lack suitable templates [31]. Despite these advancements, ab initio prediction remains far less reliable than the template-based approaches for real-world proteins, and computational cost remains a major limitation. Energy-based folding requires extensive sampling and fine-tuning of force fields. Between 2010 and 2020, de novo methods have showed modest success on small proteins and on targets where no template was available [29, 30, 32]. Over the past two decades, the limited availability of templates and imprecision of energy functions have left considerable uncertainty in the modelling of proteins with no close relatives in structural databases. These limitations motivated the adoption of deep learning to extract structural patterns directly from large sequence databases (section “Emergence of AI and deep learning in structural biology”) and model protein-protein and protein-ligand interactions (section “Protein-protein complex prediction” and Classical Docking and Scoring Functions). Emergence of AI and deep learning in structural biology Over the past decade, artificial intelligence (AI) and DL have transformed the prediction of protein structures. Early breakthroughs came from the analysis of evolutionary couplings. Among the statistical approaches employed, direct coupling analysis (DCA) has proven to be effective in identifying residue pairs that co-evolve to preserve structural contacts. These contact predictions subsequently improved fold recognition [33]. DL was then further as convolutional and residual neural networks were trained on large MSAs to predict inter-residue contacts and distance distributions. trRosetta and similar tools produce high-quality de novo models using ResNet architecture to predict not only binary contacts but also the full inter-residue distance and orientation distributions from co-evolutionary data [34]. The predicted distances and orientations are then used either as input into a differentiable folding process or as additional restraints within Rosetta to create fully atomistic 3D structures. Building upon trRosetta, several groups have combined neural network-predicted geometric potentials with physics-based energy minimization to improve the predictions for challenging targets [32, 35]. A specific example of this type of model is DeepFold [35], which combines the learned inter-residue distance and orientation potentials with a gradient-based folding process. DeepFold showed significantly improved TM scores over trRosetta (40.3% improvement) and DMPfold (44.9% improvement; Pearce et al., 2019) on the most difficult free-modelling proteins [32, 35]. Additionally, the use of learned potential greatly accelerated the folding process, with simulations running approximately 262-fold faster than the Rosetta fragment assembly on comparable targets [35]. These results demonstrate that combining deep learning restraints with efficient energy optimization can narrow the gap between de novo prediction and template-based modelling quality. The pinnacle of AI-driven prediction came in 2020–2021 with DeepMind’s AlphaFold2 [16, 17]. AlphaFold2 employs a transformer-based architecture centered on a novel module called the Evoformer, which takes the input sequence’s MSA and optional templates and uses a two-track attention mechanism (one processing pair features and one processing MSA features) to iteratively refine representations of residue pairs and single-sequence embeddings. The model is trained end-to-end to directly output three-dimensional coordinates with an invariant point attention (IPA) structure module that enforces geometric consistency. The result is a pipeline that effectively learns physics and implicitly evolutionary constraints from data. In CASP14 (2020), AlphaFold2 achieved near-experimental accuracy for the majority of targets, a result that marked a turning point for the field [16]. According to Jumper et al., AlphaFold2 became “the first method capable of predicting protein structures to near experimental accuracy in a majority of cases” [16, 17]. Quantitatively, AlphaFold2 achieved a median domain GDT-TS of 92.4 for its CASP14 predictions, far beyond any previous result [16]. AlphaFold2 can now routinely predict globular protein structures at near-atomic resolution, resolving decades-old challenges in the field. The Baker laboratory released RoseTTAFold at approximately the same time as AlphaFold [18]. RoseTTAFold uses a three-track network that processes information at three levels. These include 1D for amino acid sequences, 2D for predicted distance maps and residue orientations, and 3D for atomic coordinates [18]. This architecture outperformed the leading servers at CASP14, except for AlphaFold2 [18]. RoseTTAFold runs on a single GPU and is also available through the Robetta web server, making high-quality structure prediction accessible to laboratories without large-scale computing infrastructure. The contemporary pipeline takes an amino acid sequence as input and outputs 3D atomic coordinates. It first extracts features from co-evolutionary data (MSAs, sequence embeddings, and known templates), which are then processed by deep neural networks such as transformers or graph networks. An example of the deep-learning approaches that preceded end-to-end models, the prediction of inter-residue distance and orientation distributions from co-evolutionary features is illustrated in Supplementary Figure S1. Table 1 outlines how the field has transitioned from basic sequence-based heuristics to neural networks. This shift has produced high-quality structures whose accuracy can be meaningfully evaluated through standardized benchmarks such as GDT-TS, as used in CASP. The advent of AI has dramatically altered the landscape of protein structure prediction. After years of steady advancement, the field is now seeing rapid growth in AI-based methods capable of predicting not only monomeric structures at near-experimental accuracy [16, 17] but also complex structures that have historically been difficult to model [18, 35]. Table 1Comparison of deep-learning-based protein structure prediction models across different stages of development, summarizing their architecture, functionality, and benchmark performance. Models are grouped by category, from early deep-learning approaches to recent open-source co-folding frameworksSub-categoryModel (Year)ArchitectureFunctionality (Inputs)Performance (Benchmark)RefEarly DLtrRosetta (2020)Deep residual network (ResNet)MSA → inter-residue distance/orientation distributions → 3D structure via RosettaSubstantial improvement over contact-only methods on CASP13 FM targets[34]DeepFold (2022)Neural network potentials + gradient-based foldingMSA → learned geometric potentials → 3D structure40.3% higher TM-score than trRosetta; 262× faster than Rosetta[35]End-to-end DLAlphaFold2 (2020)Transformer (Evoformer module + IPA structure module)AA sequence + MSA → 3D protein structureCASP14: median GDT-TS 92.4 (≈1.6 Å RMSD)[16]RoseTTAFold (2021)Three-track network (MSA, sequence, and 3D coordinates)AA sequence + MSA → 3D protein structureCASP14: mean GDT-TS ≈70.1[18]MSA-free PLMsESMFold (2022)Protein language model (Transformer, no MSA)Single AA sequence → 3D protein structureCAMEO: avg. TM-score ≈0.82; 617 M metagenomic structures predicted[21]OmegaFold (2022)PLM + geometry module (no MSA)Single AA sequence → 3D protein structureCompetitive with AF2 on orphan proteins[36]Next generationAlphaFold3 (2024)Pairformer + diffusion model (unified multi-molecular)Sequence + MSA → protein + ligand + nucleic acid + ion coordinatesPoseBusters: 76% ligand success; protein-nucleic acid lDDT 0.83[22]D-I-TASSER (2025)Hybrid DL restraints + Monte Carlo physics-based refinementMSA (DeepMSA2) + templates → 3D structureCASP15: #1 server; avg TM-score 0.878; 19% higher than AF2 on FM targets[26]Open-source co-foldingBoltz-1 (2024)Pairformer + diffusion (MIT license)MSA + multi-molecular input → 3D complexMatches AF3 across lDDT, TM-score, DockQ, ligand RMSD[23]Chai-1 (2024)AF3-like + PLM embeddings + constraint featuresMSA (optional) + constraints → 3D complexPoseBusters: 77% ligand success (81% with constraints)[24]Boltz-2 (2025)Pairformer + diffusion + affinity moduleMSA + multi-molecular + affinity data → structure + binding affinityCASP16 affinity: #1 (no fine-tuning); ~1,000× faster than FEP[37] Comparison of deep-learning-based protein structure prediction models across different stages of development, summarizing their architecture, functionality, and benchmark performance. Models are grouped by category, from early deep-learning approaches to recent open-source co-folding frameworks AlphaFold3 and the emerging prediction ecosystem Architectural shift: from Evoformer to Pairformer and diffusion AlphaFold3 (AF3) is not an upgrade of AlphaFold2; the two models are built on fundamentally different architectures [22]. AF2 spent most of its computational time processing MSA information through the Evoformer module. AF3, by contrast, condenses MSA processing into four initialization blocks and passes the output to a new module called Pairformer, which consists of 48 blocks operating on pairwise and single-token representations [22]. After initialization, no further MSA processing was performed. The key difference between AF2 and AF3 is the prediction of the final 3D structure. AF2 predicts the backbone geometry and side-chain angles using specialized structural modules. AF3, however, replaces this entire module with a diffusion-based generative model that works directly on raw atomic coordinates [22, 38]. It begins with a random scatter of atoms and progressively refines their positions through learned denoising steps. This approach captures the local geometry at low noise levels and the global conformational state at high noise levels, eliminating the need for the manual programming of stereochemical terms or backbone geometry [22]. In other words, AF3 can model non-polymer entities, such as small molecules, ions, and nucleotides, along with chemically modified amino acid residues within a single unified prediction [22, 39]. Benchmark performance AF3 demonstrated impressive performance across several aspects of biomolecular interaction prediction. For protein-ligand interactions, AF3 achieved a success rate of approximately 76% with root mean square deviation (RMSD) for all AF3-predicted ligands was under 2 Å, substantially outperforming AutoDock Vina (48% success in the absence of known binding-site information) [22]. For protein-nucleic acid interactions, AF3 achieved a contact interface lDDT of 0.83, outperforming RoseTTAFold2NA (0.65) [22]. The local Distance Difference Test (lDDT) is a superposition-free measure of local model accuracy that compares inter-atomic distances between a model and a refrence structure; it ranges from 0 to 1 (higher is better) and, being local, is insensitive to domain motions [40]. Finally, AF3 reported substantial gains in protein-protein interaction prediction: in developers’ own benchmark, it produced acceptable docking predictions (DockQ ≥ 0.23) in 91% tested cases, compared with 78% for AlphaFold-Multimer v2.3 [22]. DockQ is a continuous docking-quality score on a [0, 1] scale, mapped onto the CAPRI categories (incorrect < 0.23, acceptable 0.23–0.49, medium 0.49–0.80, high > 0.80) [41]. These figures are self-reported; independent CAPRI and CASP based evaluations indicate that performance is strongly target-dependent, with overall success often in the ~40–60% range and a known skew toward easier targets such as globular homodimers [42, 43]. AF3 also demonstrated strong performance in modelling large biomolecular assemblies. For example, it accurately predicted a 4,665-residue glycosylated coronavirus spike protein complex with an lDDT score of 83.0 and a 7,663-residue ribosomal subunit with an lDDT score of 87.7 [22]. Peng et al. (2025) independently evaluated AF3 across nine benchmark datasets, providing additional validation [44]. While AF3 improved the lDDT over AF2 for protein monomers, the global fold quality (TM-score) showed only marginal improvements. For peptide-protein complexes, AF3 and AlphaFold-Multimer performed comparably, whereas for antibody-antigen complexes, AF3 was significantly superior [44]. However, RNA structure prediction remains inconsistent. AF3 performed well on secondary structural elements but showed limitations in predicting complex tertiary folds in larger RNAs [45]. Stereochemical quality also remains a concern, with a 4.4% chiral violation rate in PoseBusters and documented issues with bond lengths and angles in predicted ligand conformations [38, 46]. Limitations for understudied and orphan proteins Although AF3 substantially expands the scope of structural and interaction predictions, several fundamental limitations remain when applying this method to understudied protein families central to this review. Orphan proteins lack close sequence-based homologues, preventing AF3 from building sufficiently deep MSAs, which remain critical for prediction quality despite the reduced MSA module [22, 47]. The generation and ranking of numerous predictions can partially mitigate this limitation. A strategy that has proven effective for antibody-antigen complexes uses up to 1,000 sampling seeds, although it remains computationally intensive [22]. Single-sequence language models, such as ESMFold and OmegaFold, offer a complementary strategy by entirely bypassing MSA dependence [21, 47]. The ESM Metagenomic Atlas contains over 617 million predicted structures. This has substantially expanded the coverage of previously uncharacterized protein space [21]. In addition to the sequence alignment limitations, intrinsically disordered proteins (IDPs) present a fundamentally different challenge. AF3‘s diffusion architecture can produce hallucinations. It predicts ordered structure in regions that have no stable native conformation [22]. AF2 typically generates flexible regions as obvious extended coils, which makes these errors easy to spot. AF3‘s hallucination structures can appear convincingly ordered, making them more difficult for visual inspection. Gopalan and Narayanan (2025) quantified the issue in a systematic evaluation of 72 IDPs from the DisProt database [48], 22% of residues were classified as hallucinations. It comprising both incorrectly predicted order in experimentally disordered regions and incorrectly predicted disorder in ordered regions. and 18% of the residues annotated to biological processes were also affected, raiseing concerns for downstream drug discovery applications [48]. Notably, the same study found that hallucinated order is not reliably caught by low pLDDT. AF3 sometimes assigns high pLDDT (>70) to disordered residues, undermining the assumption that confidence scores alone can filter such errors [48]. Cross-validation with dedicated disorder predictors is therefore advisable when working with understudied proteins likely to contain disordered regions. Heteromeric protein complexes with weak or transient interfaces show limited improvement over previous prediction methods [44]. Membrane proteins also remain challenging because of the bias in training data towards soluble proteins and the absence of explicit lipid bilayer modelling [49]. Among these challenges, the inability to capture conformational dynamics is the most fundamental limitation. AF3 predicts single static structures and cannot capture the conformational ensembles underlying protein function. Zheng et al. (2025) showed that AF3 excels at predicting static protein-ligand complexes but struggles significantly with dynamic complexes undergoing conformational rearrangements upon ligand binding [50]. For example, E3 ubiquitin ligases naturally adopt an open conformation when no substrate is bound; however, AF3 predicts only the closed form, regardless of the input [22]. Running multiple prediction seeds does not resolve this issue either, as predictions typically converge on the same static fold rather than sampling different conformational states. The open-source prediction Ecosystem The open-source reproduction of these models predates AF3. As early as 2021, OpenFold [20], developed by the AlQuraishi laboratory, provided the first fully open and trainable reimplementation of Alphafold2 in PyTorch. Whereas DeepMind had released Alphafold2’s inference code and weighs but not the training pipeline, OpenFold reproduced the model from scratch and released the training code, and data-processing pipeline, matching AlphaFold2‘s accuracy [20]. This established the open-source foundation on which much of the field would later build. The subsequent release of AlphaFold3 (AF3), which was not fully open-source nor available for commercial use, then catalysed a second wave of open reproductions aimed at extending AF3‘s expanded capabilities. One of the earliest and most notable approaches of these AF3 efforts was Boltz-1 [23], developed at MIT Jameel Clinic and released under the permissive MIT license with full training and inference code and model weights. [23]. It was the first fully commercially accessible open-source model to reach AF3-level accuracy across key benchmark [23]. Its successor, Boltz-2 [37], extended the framework to binding affinity estimation. It was the first deep-learning model to approach physics-based free-energy perturbation (FEP) accuracy while running approximately 1,000-fold faster, and it outperformed all participants in the CASP16 blind affinity challenge without the need for fine-tuning. [37]. In parallel with the Boltzmann framework, other groups have developed complementary approaches. Chai-1 [24] incorporates protein language model embeddings for enhanced single-sequence prediction and trainable constraint features, and allows user-guided experimental restraints, achieving 77% ligand RMSD success on PoseBusters, comparable to AF3 [24]. Chai-1 was initially released under a restrictive license and was later made open-source under Apache 2.0. Further implementations include Protenix [51] from ByteDance and OpenFold3 (2025) [25] from the OpenFold Consortium. OpenFold3 is notable for its Federated Training Initiative, in which five pharmaceutical companies contribute proprietary protein-drug data to expand training beyond the public PDB [25, 51]. Similarly, HelixFold-3 [52] demonstrated that predicted structures can be used for FEP calculations with accuracy is comparable to that of experimental crystal structures [52, 53]. Beyond individual tools, ABCFold [54] facilitates the comparison of multiple structure-prediction tools by providing standardized input formats for running AlphaFold3, Boltz-1, and Chai-1 [54]. For understudied protein families, combining these tools in parallel to generate consensus models can improve confidence in predicted structures, particularly for difficult protein classes where individual methods show inconsistent performance. Implications for understudied protein families Advances in AF3 and its open-source alternatives have several direct implications for characterizing understudied proteins. AF3 can now model proteins with small molecules, nucleic acids, ions, and modified residues within a single prediction [22]. Researchers can now generate structural hypotheses regarding the interactions of understudied proteins with multiple molecules, even in the absence of experimental structures. Earlier, AF2 contributed to predicting a significant number of structures from understudied human proteins [55]. Several tools are now built on these predictions for further analysis, such as GalaxySagittarius-AF, which leverages AF3-derived structural libraries to predict drug-target interactions across the proteome [56]. Similarly, Foldseek-based structural searches have enabled the clustering of predicted structures into novel protein families. It has revealed hidden evolutionary relationships among previously unconnected proteins [57]. AF3‘s success in antibody-antigen prediction, where the co-evolutionary signal between binding partners is minimal, suggests that the model captures chemical and physical principles beyond sequence patterns alone. This is an encouraging indication for future predictions of understudied families [22]. However, the fundamental challenge for understudied proteins remains data availability rather than algorithmic superiority. Most current co-folding models are trained primarily on PDB structures, which are inherently biased towards well-studied, crystallizable proteins, although some efforts (e.g., the openFold3 Federated Training Initiative [25]) are beginning to expand training beyond the public PDB [44, 49, 58]. Even the most advanced architectures produce uncertain predictions when orphan proteins have fewer related sequences in MSAs, membrane proteins lack representative training examples, or IDPs defy single-structure descriptions. Whether the remaining challenges are fundamentally data-limited, algorithmically limited, or both is addressed in section “Limitations of AI/ML-based protein structure prediction”. Quality metrics for AI-Based structure prediction Determining the reliability of predicted protein structures remains a central challenge in the field of computational structural biology. There is no single correct answer to this question because many factors influence the quality of predicted protein structures. Therefore, the quality was evaluated using various metrics. These metrics assess features such as the global fold, side-chain packing, and accuracy of protein-protein interfaces. The choice of appropriate evaluation metrics becomes particularly important when little or no experimental structural data are available for direct comparison. A few commonly used quality metrics for different types of structural assessments are listed in Table 2. Table 2Quality metrics used for AI-based protein structure prediction and modelling, including measurement range, numerical range, common interpretation thresholds, superposition requirements, and key applicationsMetricWhat It MeasuresRangeKey ThresholdsRequires Experimental ReferenceSuperposition-BasedBest Used ForRefRMSDAverage atomic displacement after superposition0 → ∞ (Å)<2 Å excellent; >5 Å poorYesYesQuick comparison; sensitive to outliers[59]GDT-TS% of Cα atoms within 1/2/4/8 Å cutoffs (averaged)0–100>90 near-experimental; >80 high quality; <20 randomYesYesCASP primary metric (CASP3-14); global backbone quality[15]GDT-HAHigh-resolution GDT (0.5/1/2/4 Å cutoffs)0–100Stricter than GDT-TSYesYesNear-experimental quality assessment[60]TM-scoreLength-independent global fold similarity0–1>0.5 same fold; >0.7 high quality; ~0.17 randomYesYesCross-target comparison; CASP15+ benchmarking[61]lDDTLocal distance preservation (no superposition)0–1 (or 0–100)>0.8 high quality; <0.6 poorYesNoMultidomain proteins; individual domain evaluation[40]pLDDTPredicted local confidence (model self-assessment)0–100>90 high; 70–90 confident backbone; <50 likely disorderedNoN/AIdentifying reliable vs. unreliable regions; disorder proxy[16]pTMPredicted global fold quality0–1>0.7 high quality; >0.5 likely correct foldNoN/AOverall prediction trustworthiness[16]ipTMPredicted interface quality (multimers)0–1>0.8 confident; 0.6–0.8 uncertain; <0.6 likely failedNoN/AProtein complex interface reliability[62]PAEPredicted relative position error (pairwise)0 → ~30 (Å)<5 Å confident; >15 Å uncertainNoN/AInter-domain orientation confidence; visualized as heatmap[16]DockQProtein-protein docking quality (Fnat + iRMSD + LRMSD)0–1>0.80 high; 0.49–0.80 medium; 0.23–0.49 acceptable; <0.23 incorrectYesPartiallyComplex/docking evaluation[41]DockQ v2Extended DockQ for nucleic acids + small molecules0–1Same thresholds as DockQYesPartiallyAF3-era multi-molecular complexes[63]pDockQ/pDockQ2Predicted DockQ from AF confidence scores0–1>0.5 likely acceptable interfaceNoN/ARapid screening of predicted complexes[42, 64]Ligand RMSDPositional deviation of predicted ligand atoms0 → ∞ (Å)<2 Å successful poseYesYesProtein-ligand binding pose quality[22].lDDT-PLIlDDT for protein-ligand interface contacts0–1Higher is betterYesNoProtein-ligand interface accuracy[22].PoseBustersStereochemical validity of ligand pose (chirality, bonds, clashes)Pass/fail per criterionMust pass all checks for valid posePartiallyN/AChemical plausibility of predicted ligand conformations[65] Quality metrics used for AI-based protein structure prediction and modelling, including measurement range, numerical range, common interpretation thresholds, superposition requirements, and key applications To understand the true meaning of structure prediction, we must first know what “accuracy” means in the context of structural biology. The term “accuracy” is commonly used to describe the quality of a three-dimensional model. However, this meaning differs from its use in machine learning, where accuracy refers to the fraction of correctly classified instances. In structural modelling, it refers to the degree of structural similarity between a predicted model and an experimental reference structure. The similarity is quantified using measures such as the root-mean-square deviation (RMSD), lDDT, or GDT-TS [60, 66]. Different metrics capture different aspects of the structural agreement between predicted and experimental models. We therefore use the broader term predicted structure quality and explicitly specify the particular metric when reporting quantitative results. Table 2 indicates whether each metric requires an experimental reference structure, whether structural superposition is needed, and the primary use case for each. The metrics are organized into five groups: superposition-based global metrics (RMSD, GDT-TS, GDT-HA, TM-score); local distance-based metrics (lDDT); model-internal confidence scores (pLDDT, pTM, ipTM, PAE); and docking-quality metrics (DockQ, DockQ v2, pDockQ/pDockQ2), and ligand-specific metrics (Ligand RMSD, lDDT-PLI, PoseBusters). Model-internal confidence scores are particularly valuable for understudied proteins, where no experimental refrence is available for direct comparison. Superposition-based global metrics The simplest and most widely used metric for measuring structural similarity is the Root Mean Square Deviation (RMSD). It measures the average distance between corresponding atoms in the predicted and reference structures, typically calculated using Cα atoms. [59]. RMSD is easy to calculate but has well-known limitations. It is highly sensitive to outlier regions, for example, a single incorrectly modeled loop can substantially increase the RMSD value [15, 66]. It is also protein-length dependent which can cause larger proteins to receive higher (worse) RMSD values even when the overall fold is correct. RMSD is therefore not used as the primary metric in CASP evaluations and caution must be taken when using RMSD to compare predictions across proteins of different sizes [66]. To overcome the sensitivity of these RMSD outliers, GDT-TS was developed by Zemla (2003) [15]. Unlike the RMSD, which averages all atomic deviations, the GDT-TS evaluates the fraction of Cα atoms that fall within the predefined distance thresholds of their corresponding positions in the experimental structure. The final GDT-TS score was the average of the four distances (1, 2, 4, and 8 Å) [15]. The scores range from 0 to 100, with a score of 100 indicating a perfect structural match. GDT-TS has been used as the primary metric for evaluating prediction accuracy in the CASP competition since CASP3 (1998). In fact, at CASP14, it was used as a benchmark to measure the success of AlphaFold2 [17, 66]. The high-accuracy global distance test (GDT-HA) is a more stringent variant of this metric. It employs stricter distance thresholds of 0.5, 1, 2, and 4 Å, making it more appropriate for evaluating predictions that approach the experimental resolution. [60]. Another commonly used metric for structural prediction is the template Modelling score (TM-score). It was developed by Zhang and Skolnick in 2004 to provide a measure of structural similarity that is independent of the protein length [61]. The TM-Score ranges from 0 to 1. A score greater than 0.5 indicates that the predicted and experimental structures contain the same general topology. However, if the score is less than 0.17, it suggests that the structures do not contain the same topology. Furthermore, a score greater than 0.7 implies a high-quality structural prediction [26, 61]. At present, the TM-score is a widely used metric for quality assessment, including AlphaFold and D-I-TASSER. It is important to distinguish the TM-score from pTM, which is a model-internal confidence estimate (section “Model confidence scores: pLDDT, pTM, and PAE”). Superposition-free local metrics The metrics described above rely on the global superposition of the predicted and reference structures. This approach can be problematic for multidomain proteins, where individual domains may be accurately predicted, but their relative orientation is incorrect. In such cases, superposition-based metrics would heavily penalize the prediction, even though the individual domains are structurally correct. The local Distance Difference Test (lDDT), introduced by Mariani et al. (2013), avoids this problem entirely [40]. Instead of aligning the entire structure, lDDT evaluates whether local inter-atomic distance patterns are preserved between the predicted and reference structures. For each residue, the method compares the distances to neighboring atoms within a 15 Å radius in the prediction against the corresponding distances in the reference structure [40]. The lDDT is scored between 0 and 1 (or 0–100 as a percentage), with higher values indicating better preservation of local structural geometry. The most widely used variant is lDDT-Cα, evaluates distances using only Cα atoms. Because lDDT does not rely on global structural alignment, it can correctly evaluate individual domains within multi-domain proteins, even when their relative orientations differ [40]. This makes lDDT particularly useful for evaluating the predictions of large multi-domain proteins from understudied protein families. Since CASP15, lDDT has been adopted as one of the primary assessment metrics alongside GDT-TS and TM-score for evaluating the structure prediction performance [60]. Model confidence scores: pLDDT, pTM, and PAE A distinctive feature of modern AI-based predictors is that they provide built-in confidence estimates along with their predicted coordinates. In other words, the model reports its confidence in different regions of its prediction. These confidence scores are especially important for understudied proteins for which no experimental structure exists for direct comparison. The predicted local Distance Difference Test (pLDDT) is a per-residue confidence score produced by AlphaFold2, AlphaFold3, and related models [16]. It estimates how well the local structural environment of each residue is expected to match the experimentally determined structure. The pLDDT values range from 0 to 100 and are interpreted as follows: >90 = High confidence; backbone and side chains are likely to be well modeled accurately. >90 = High confidence; backbone and side chains are likely to be well modeled accurately. 70-90 = Confident backbone prediction, but side-chain positions may contain errors. 70-90 = Confident backbone prediction, but side-chain positions may contain errors. 50-70 = Low confidence in local structure. 50-70 = Low confidence in local structure. <50 = region is likely to be intrinsically disordered or poorly predicted <50 = region is likely to be intrinsically disordered or poorly predicted Low pLDDT scores are particularly informative for identifying intrinsically disordered regions (IDRs), which frequently exhibit pLDDT scores below 50 [16]. Another global confidence measure is the predicted template Modelling score (pTM), which estimates the overall topological accuracy of the entire prediction [16]. The pTM ranges from 0 to 1, with values above 0.5 generally indicating that the predicted fold is likely to be correct. For multimeric protein complexes, AlphaFold also reports the interface predicted TM-score (ipTM), which specifically predicts the accuracy of inter-chain interactions [16]. In general: >0.8 indicates highly confident interface predictions. >0.8 indicates highly confident interface predictions. 0.6-0.8 suggest moderate or uncertain confidence 0.6-0.8 suggest moderate or uncertain confidence <0.6 indicates a likely incorrect interface prediction <0.6 indicates a likely incorrect interface prediction However, disordered regions with low pLDDT can artificially lower ipTM values, even when the structured portions of the complex are correctly predicted [16]. PAE is typically visualized as a two-dimensional heatmap, where each cell (i, j) represents the predicted positional error of a residue pair. Low PAE values (dark squares) indicate high confidence in relative positioning, whereas high values (light squares) indicate structural uncertainty [16, 62]. This is especially useful for multi-domain proteins. When individual domains appear as dark blocks along the diagonal (low PAE within domains) of the heatmap, but lighter regions appear between domains (high PAE between domains), this pattern indicates that the individual domains are likely predicted correctly, but their relative orientation is uncertain [62]. This pattern is commonly observed in the predictions of large multi-domain and understudied protein families. Interface and docking quality metrics Predicting protein-protein complexes requires evaluation metrics that specifically assess interface quality, as global metrics such as TM-score fail to capture errors localized at the binding interface. One widely used metric, DockQ, was introduced by Basu and Wallner in 2016 [41]. DockQ draws on the quality measures established by the Critical Assessment of Prediction of Interactions (CAPRI), the recurring community challenge that blind-tests protein-complex prediction methods, much as CASP does for single-protein structures. Specifically, it combines three such measures, the fraction of native contacts (Fnat), interface RMSD (iRMSD), and ligand RMSD (LRMSD), into a single continuous score between 0 and 1 [41]. DockQ classifies prediction quality as follows: <0.23 = incorrect prediction <0.23 = incorrect prediction 0.23-0.49 = acceptable quality 0.23-0.49 = acceptable quality 0.49-0.80 = medium quality 0.49-0.80 = medium quality >0.80 = high quality >0.80 = high quality An updated version, DockQ v2 [63], extends the metric to nucleic acid and small-molecule interfaces, making it applicable to a broader range of complexes predicted by AlphaFold3 and related models [63]. For rapid evaluation of predicted complexes without experimental references, pDockQ [67] provides an approximate DockQ estimate derived from AlphaFold’s confidence metrics; specially, it uses the interface pLDDT and the number of interface contacts, fitted through a sigmoidal function, to estimate the DockQ [67]. Although pDockQ is useful for large-scale screening, it tends to overestimate the interface quality for homomeric and large multimeric assemblies [42]. An improved version, pDockQ2, incorporates PAE information for more reliable interface-quality estimates [42]. Ligand-specific metrics The prediction of protein-ligand complexes introduces its own evaluation challenges. In such cases, ligand poses must be evaluated not only for positional accuracy, but also for chemical and stereochemical plausibility. The most commonly used metric is the Ligand RMSD, which measures the positional deviation of ligand atoms from the experimentally determined binding pose after aligning the protein backbone. A threshold of 2 Å is typically used to define successful pose prediction [22]. Another metric, lDDT-PLI (lDDT for Protein-Ligand Interactions), extends the lDDT framework to specifically assess protein-ligand contact geometry. Because it does not require structural superposition, the lDDT-PLI provides a robust measure of the protein-ligand interface quality [22]. Beyond positional metrics, PoseBusters evaluates the chemical validity of the predicted ligand conformation [65]. The methods check for stereochemical violations, including incorrect bond lengths, bond angles, chirality, and steric clashes, problems that RMSD-based evaluation alone cannot detect. This validation step has become particularly important for diffusion-based models, such as AlphaFold3, which can sometimes produce poses that appear correct by RMSD but contain physically implausible geometries. For example, AF3 shows a 4.4% chiral violation rate in the PoseBusters benchmark dataset [22, 65]. Practical interpretation for understudied proteins For understudied protein families that lack experimentally determined structures, model-internal confidence scores serve as the primary tools for assessing prediction reliability. These scores include pLDDT, pTM, ipTM, and PAE scores. A few practical guidelines can help researchers effectively navigate these confidence measures. The first consideration is the reliable prediction of the global fold. When a predicted structure shows consistently high pLDDT values above 70 across most residues and a pTM above 0.7, the overall fold is considered reliable. These models can serve as a reasonable starting point for generating hypotheses regarding protein function, binding sites, and molecular interactions [16]. The second consideration addresses the uncertainty of domain orientation. In cases where individual domains show high pLDDT values but the PAE between domains remains high, the structure of each domain may be considered reliable. However, their relative orientations should be interpreted with caution [16, 22, 62]. The third consideration relates to the cross-model consensus. For orphan proteins, predictions should ideally be generated using multiple methods, such as AF2, AF3, ESMFold, and D-I-TASSER. Regions where independent methods converge on the same conformation carry greater weight than the confidence score of any single model [50]. While these guidelines support the interpretation of predictions, current computational tools have limitations that may mislead researchers, particularly when working with understudied proteins that lack experimental validation. Two key drawbacks warrant attention. The first is the hallucinated order in disordered regions discussed in section “Limitations for understudied and orphan proteins”, where diffusion-based models can generate visually convincing but spurious structure that is not reliably flagged by low pLDDT [48]. The second drawback is the confusion of confidence with structural correctness. A high pLDDT score reflects the model confidence in a given prediction but does not confirm that the prediction is structurally correct. This distinction is particularly important for proteins with no close homologues in the training data, where the model may produce confident but inaccurate structural outputs. Illuminating the dark proteome: challenges, opportunities, and AI-Driven insights The dark proteome refers to proteins or regions lacking experimentally determined structures or functional annotations. These uncharted proteins constitute a large fraction of genomes, often > 40% in eukaryotes, including many regulatory factors. For example, intrinsically disordered proteins (IDPs), which lack a correct 3D fold, are major components of the dark proteome. They play a role in signaling, regulation, and phase separation in membrane-less organelles [68]. In other words, “dark” proteins are not intrinsically unimportant; rather, their structures and functions remain unexplored. Recent estimates suggest that even after the AlphaFold2 revolution, ∼10% of human proteins remain structurally dark [69], underscoring the challenge of this “sequence-structure gap.” These technical limitations make it difficult to study dark proteins. Because they lack closely related analogues with respect to structures already established by X-ray crystallography or nuclear Magnetic Resonance (NMR), dark proteins cannot be modeled using template-based models [68]. Most dark proteins exhibit high flexibility and/or contain long IDRs, causing difficulties in both experimental structure determination (NMR or crystallography) and computational prediction. In addition, a number of dark proteins form a stable tertiary structure only upon binding to a partner [68, 70], making them even less amenable to single-structure-based analysis [70]. Membrane proteins are another class of prominent in the dark proteome. experimental structures of transmembrane (TM) domains are scarce and most AI models trained predominantly on soluble proteins perform poorly on transmembrane helices [58]. Even where structure prediction has advanced dramatically, functional annotation of the dark proteome remains a challenging task. Tools such as DeepFRI [71] address this by employing graph convolutional networks (GCNs) to predict Gene Ontology (GO) functions, integrating features from sequence embeddings and 3D structures. In addition to these approaches, transformer-based models such as ProtTrans [72], ESMFold, and ESM-2 [21] (which are large protein language models) are capable of capturing long-range sequence dependencies that further enhance functional inference. Most recently, the structure prediction tools discussed in section “AlphaFold3 and the emerging prediction ecosystem” and “Evolution of structure prediction methods”0.4, including AlphaFold3, RoseTTAFold All-Atom, and protein language model-based predictors, have begun to address these challenges. They are providing structural models and functional hypotheses for previously uncharacterized proteins, as detailed in the following sections. To systematically prioritize proteins for computational characterization, the NIH Illuminating the Druggable Genome Initiative classifies human proteins into four Target Development Levels (TDLs):, Tclin, validated drug targets; Tchem, proteins with known small-molecule modulators; Tbio, proteins with established biological information but limited chemical probes; and Tdark, understudied proteins that lack functional annotation and known ligands. This framework highlights the large number of poorly characterized proteins and emphasizes the importance of computational approaches for exploring these understudied regions of the human proteome [73]. AlphaFold2 predictions have dramatically expanded the structural coverage of the dark proteome. For example, Binder et al. (2022) reported that over half of human “Tdark” proteins now have confident or very high confidence predicted structures (pLDDT > 70) [55]. In Fig. 1, we conclude that most (~54%) of the uncharacterized proteins are classified as high confidence, while ~46% are predicted with low confidence (pLDDT < 70) [55]. This may indicate true intrinsic disorder, limited evolutionary data, or unique structural characteristics that challenge existing prediction techniques. These two groups require distinct computational approaches. The high confidence group can be directly used for functional annotation and drug discovery. The low-confidence group requires tailored strategies, as discussed in section “AI-Driven dynamics and ensemble methods” and Conclusion and future perspectives (see Supplementary Table S1). Distribution of human Tdark proteins across AlphaFold2 prediction-confidence categories. Bars indicate the number of Tdark proteins (left axis) in each pLDDT confidence bin; the line shows the percentage of total Tdark proteins per category (right axis). Analysis is based on AF2 models for 5,592 human Tdark proteins, proteins classified as understudied by the NIH Illuminating the druggable genome initiative, lacking functional annotation and known ligands. AF2 produced structural models for nearly 100% of Tdark proteins, with approximately two-thirds of residues across all models achieving pLDDT ≥ 80. Data from Binder et al. (2022) [55]; confidence categories follow the standard AlphaFold pLDDT classification scheme This review highlights two representative case studies that demonstrate the use of AI-based techniques. GPCRs and IDPs represent the challenges of membrane proteins and disordered proteins. In both cases, no traditional approach yielded satisfactory results. However, AI methods also have their own limitations and should be used strategically in such scenarios [74]. Selecting the most appropriate tool is crucial when working with the dark proteome and this choice depends on the specific requirements of each target. By using a case-specific and integrative approach, researchers can uncover the structure and function of proteins that are still hidden within the dark proteome. Protein-protein complex prediction Proteins generally function within a vast network of protein-protein interactions (PPIs), which mediate the formation of larger molecular assemblies and underlie signal transduction. Understanding cellular organization and function therefore requires determining not only the three-dimensional (3D) structure of individual proteins but also the 3D architecture of their interaction networks. Knowledge of the 3D structure of PPI complexes also enables the rational design of drugs that act on specific binding interfaces. This section review focuses on computational methods for predicting the 3D structure of protein-protein complexes, including traditional docking and newer co-folding approaches, with emphasis on additional challenges encountered when one or both proteins are understudied or orphan proteins. Classical docking approaches Most classical protein-protein docking methods rest on a few basic assumptions. First, the proteins are treated as a rigid or semi-rigid bodies. Second, complexes are sampled across relative positions and orientations and scored using physico-chemically based scoring functions. Third, fast Fourier transform (FFT) based docking algorithms enable the evaluation of an enormous number of candidate conformations [75]. While this approach is computationally efficient, the rigid-body approximation limits docking accuracy for proteins that undergo significant conformational changes upon binding. Because this is a sampling limitation, it cannot be overcome by improved scoring, even when decoys are evaluated with hundreds of scoring functions [76]. To address these limitations, information-driven and flexible docking approaches have been developed. These methods are designed to utilize experimental constraints in addition to being flexible enough to allow for backbone and side chain movements around the area where docking occurs. For example, HADDOCK [77] integrates experimental restraints from NMR, mutagenesis, or cross-linking mass spectrometry with semi-flexible docking to allow for the movement of proteins around the docking site. Another docking methodology that includes experimental constraints is SwarmDock [78]. SwarmDock utilizes particle swarm optimization in conjunction with local energy minimization to search for possible docked conformations while allowing the movement of the docking proteins. LightDock [79] employs the Glowworm Swarm Optimization algorithm to combine integrative data from various experimental sources. In general, these three approaches illustrated methodological improvements and increased accuracy achievable for systems exhibiting induced-fit. Each requires more computation than rigid-body docking; in practice, however, all three are available as free web servers, which lowers the barrier to use for the community. Deep learning has also been applied to docking. One example is EquiDock [80] which uses an SE(3)-equivariant graph neural network to predict the rigid body transformation between two proteins without iterative sampling or scoring, allowing substantial speedups over classical methods. As a rigid-body method, however, it does not model induced fit and has seen limited uptake in community benchmarks such as CAPRI. All classical and DL-based docking approaches share a fundamental limitation. They require pre-existing structures of the individual proteins as input. For understudied proteins with no experimental structure, these methods can only be applied to AI-predicted models. Introducing an additional layer of uncertainty that can degrade interface-prediction quality [43]. Co-folding with AlphaFold-Multimer and AlphaFold3 To address the structural reliance present in docking techniques, co-folding methods have emerged as a transformative approach for predicting protein-protein interactions. They directly assume the configuration of multichain structures from sequence data alone. AlphaFold-Multimer [62], an extension of AlphaFold2 trained specifically on multimeric inputs, achieved successful interface predictions (DockQ ≥ 0.23, as defined in section “Interface and docking quality metrics”) in 67–70% of heteromeric cases on a benchmark of 4,433 recently deposited complexes [62]. High-accuracy predictions (DockQ ≥ 0.8) were achieved in 23–26% of the cases. Performance was substantially better for homomeric complexes than for heteromers and notably poor for antibody-antigen complexes [42, 62]. Subsequent developments have further expanded this framework. AlphaFold3 (AF3) has shown significant advancements in predicting various protein complexes and improved efficacy for antibody-antigen interactions (section “Benchmark performance”). Comprehensive benchmarking by Peng et al. (2025) confirmed that for peptide-protein complexes, AF3 and AF-Multimer perform comparably, whereas for antibody-antigen complexes, AF3 also offers faster inference, though it requires more recent GPU hardware than AlphaFold-Multimer, which may limit accessibility for some groups [44]. The CASP15-CAPRI experiment further demonstrated that AF-based methods fundamentally changed the landscape of complex structure prediction. Currently, the majority of top-performing groups use some form of AlphaFold-derived strategy [43]. Several limitations remain important for researchers working with understudied proteins. First, co-folding approaches require known stoichiometry as input. They cannot predict whether two proteins interact or determine the number of copies of each chain in the assembly. The ipTM score (see section “Model confidence scores: pLDDT, pTM, and PAE”) provides some guidance, with values above 0.8 indicating confident interface predictions. Although distinguishing true interactors from non-interactors remains imperfect [42, 64]. Second, constructing paired MSAs can substantially degrade prediction quality [62, 67]. Third, the prediction of interactions mediated by short linear motifs (SLiMs), common among disordered proteins. They require protein fragmentation strategies, as AF-Multimer largely fails to identify these interfaces when full-length sequences are used [81]. Omidi et al. (2024) demonstrated that AF-Multimer can predict structures adopted by intrinsically disordered regions (IDRs) upon binding, although accuracy decreases for IDRs that interact in heterogeneous binding modes [82]. Open-source Co-folding alternatives Building on advances pioneered by AlphaFold3, several open-source co-folding frameworks have recently been released, including, Boltz-1, Chai-1, and Protenix. These tools have broadened the prediction landscape by enabling joint modelling of protein-protein complexes alongside other biomolecular interactions. In head-to-head comparisons, the open-source models converge on similar performance for non-antibody complexes, although AF3 retains an advantage for antibody-antigen prediction owing to its more extensive sampling strategy [23, 54]. To enable consistent benchmarking, the ABCFold pipeline offers a unified framework for evaluating AF3, Boltz-1, and Chai-1 on the same inputs [54]. Such cross-model benchmarking is especially valuable for difficult, lesser-studied targets where predictive accuracy is substantially between methods. Recent applications of co-evolutionary analysis at proteome scale have advanced PPI discovery substantially. Zhang et al. (2025) developed a deep-learning model based on RoseTTAFold2-PPI optimized for both prediction accuracy and computational speed [83]. They screened 200 million human protein pairs using deep MSAs derived from 30 petabytes of genomic data, identifying17,849 high-confidence interactions at an estimated 90% precision, of which 3,631 were previously uncharacterized and were subsequently validated through computational and experimental follow up [83]. This study demonstrates that AI-driven PPI prediction has reached the scale required to systematically illuminate unknown interaction networks involving understudied human proteins. Docking and Scoring Functions Classical docking and scoring functions Physics-based molecular docking continues to play a major role in structure-based drug discovery. Widely used docking tools include AutoDock Vina [84] and GLIDE [85], both of which generate candidate ligand conformations within the protein-binding pocket. The molecules are then ranked using empirical or physics-based scoring functions that approximate the interaction energies. Because docking software efficiently samples large numbers of candidates, it is well-suited for large-scale virtual screening campaigns involving hundreds of thousands to tens of millions of compounds. However, docking software has several significant drawbacks. Most docking engines treat the protein receptor as rigid or allow only limited side chain flexibility. This fails to capture the induced-fit conformational changes that frequently occur upon ligand binding. The success rates for identifying the correct binding pose (ligand RMSD < 2 Å, as defined in section “Ligand-specific metrics”) typically range from 50 to 70% when the experimental binding site is known. The performance of docking algorithms drops substantially when the binding site must be predicted or when the receptor undergoes conformational rearrangement [86]. These limitations are particularly problematic for poorly studied proteins that lack experimental structures. In such cases, docking accuracy is compromised by structural inaccuracies in the binding pocket, including side-chain misplacements and the predominance of apo-like conformations. Díaz-Rovira et al. (2023) showed that virtual screening against AlphaFold-predicted structures yields lower hit rates than screening against experimental holo structures [87]. This is primarily because the predicted structures often exhibit side-chain misplacements in the binding pocket and represent the apo rather than the ligand-bound conformation. Machine learning scoring and affinity prediction The limitations of empirical scoring functions in classical docking have motivated the development of machine learning-based methods. They model ligand-protein interaction energies directly from structural data. One of the earliest methods was RF-Score [88], which used Random Forest Regression to predict the ligand-protein binding affinity using atom pair counts. RF-Score achieved a Pearson correlation of approximately 0.78 on the PDBbind benchmark while outperforming multiple empirical scoring functions [89]. Subsequently, deeper architectures were developed. DeepAtom [90] utilized 3D convolutional neural networks on voxelized representations of protein-ligand structures, achieving a Pearson correlation of 0.83 on the PDBbind 2016 core set [90]. While these ML scoring functions marked important milestones in data-driven molecular recognition, they have been substantially outperformed by co-folding and generative techniques. A significant advancement in binding affinity prediction has been achieved by Boltz-2, a model that simultaneously predicts protein-ligand complex structure and binding affinity [37]. Boltz-2 was among the first deep-learning model to approach the accuracy of free energy perturbation (FEP). The gold-standard physics-based method for computational binding-affinity prediction, while running approximately 1,000 times faster [37]. Boltz-2 won the blind CASP16 affinity challenge, outperforming all top-ranking participants without any task-specific fine-tuning [37]. This represents a fundamental shift in the field, as binding affinity prediction transitions from a post-docking rescoring step to an integral component of structural prediction itself. However, Boltz-2 affinity predictions show reduced effectiveness for certain protein classes, particularly GPCRs, which are well known to challenge even physics-based FEP methods and to require careful, customized system preparation for reliable results [91]. This indicates that further development is required before these models can replace conventional computational chemistry workflows. Scoring functions and affinity prediction models are typically evaluated using the Comparative Assessment of Scoring Functions (CASF) framework. CASF utilizes the PDBbind database to assess four capability functions: scoring, ranking, docking, and screening powers. It also covers classical, machine learning, and deep-learning scoring approaches [92]. Deep learning-based docking and Co-folding Recent advancements in machine learning-driven scoring have further improved ligand docking using generative modelling techniques. DiffDock [93] reframed ligand docking as a generative diffusion process, learning a probability distribution over ligand poses in translational, rotational, and torsional spaces. On the PDBbind benchmark, DiffDock achieved a 38% top success rate (ligand RMSD < 2 Å). It substantially outperformed both physics-based docking (20–23%) and earlier deep learning methods [93]. By generating multiple pose candidates with associated confidence scores, it provides an intrinsic mechanism for uncertainty estimation. The emergence of co-folding frameworks has led to a paradigm shift. Rather than requiring a predefined protein structure as input, AF3 and its open-source counterparts, including Boltz-1 and Chai-1, directly predict protein-ligand complex structures from the sequence and ligand chemical information in a single forward pass [22]. On the PoseBusters benchmark, AF3 markedly outperforms classical docking, though its residual stereochemical errors (chiral violations and distorted ligand geometries) remain a concern (section “Benchmark performance”; see also section “Ligand-specific metrics”) [22, 38]. SiteAF3 (2025) refines this approach by fixing the receptor structure and incorporating the binding pocket information as a conditional input to the diffusion process. It consistently outperformed the standard AF3 for site-specific predictions and showed particular advantages for orphan proteins and allosteric binding sites [94]. The RoseTTAFold All-Atom [95] provides an alternative co-folding architecture based on a three-track network with atomic graph representations and outperforms automated baseline pipelines in the Continuous Automated Model Evaluation (CAMEO) blind ligand-docking challenge [95]. This conceptual shift is significant for several reasons. The field is transitioning from a two-step approach of sequential structure prediction followed by docking to the integrated prediction of the complex in a single step. Co-folding methods prevent the error propagation that is inherent to sequential pipelines. Nonetheless, these methods currently predict the most thermodynamically favorable binding mode. This may not correspond to a therapeutically relevant outcome. Furthermore, they remain unsuitable for large-scale compound library screening, a capability that traditional docking methods retain [22]. Binding site prediction and virtual screening For both docking and co-folding methods, the absence of a known ligand-binding site poses a significant challenge, particularly for understudied proteins. In such cases, computational pocket identification is an essential prerequisite for subsequent modelling and virtual screening. Deep learning has substantially improved the accuracy of binding site prediction. DeepSite [96] applies 3D CNNs to protein structure grids to identify probable pocket locations. Kalasanty [97] used 3D U-Net segmentation to delineate pocket boundaries, achieving approximately 66% success in placing the predicted pocket center within 4 Å of the true binding site. PocketMiner [98] adopts a different approach. It employs graph neural networks trained on molecular dynamics simulation data to predict the locations of cryptic pockets. These are transient binding sites that are absent in the ground-state structures but become accessible through conformational fluctuations. This capability is particularly relevant for understudied proteins, where functionally important binding sites may not be apparent from a static AI-predicted structure alone. When the ligand identity is unknown and co-folding tools cannot be applied, AlphaFill [99] offers an alternative strategy by transplanting probable ligands and cofactors onto AlphaFold-predicted structures based on homologous complexes in the PDB. Although AlphaFill cannot handle novel ligand chemistries, it provides rapid initial hypotheses about the likely binding partners for proteins with detectable structural homologues. The complete AI-driven drug discovery workflow, from structure prediction to experimental validation, is illustrated in Fig. 2. Practical workflow for AI-driven drug discovery targeting understudied proteins. The pipeline progresses from structure prediction through binding site identification, virtual screening, ML-based affinity rescoring, and experimental validation. Tool names indicate representative methods available for each step; timelines are approximate. Dashed arrows indicate iterative refinement, in which experimental results constrain subsequent rounds of prediction. This workflow was validated in the TAAR1 case study discussed (section “Case study 1: orphan GPCRs: from prediction to ligand discovery”), in which the screening of approximately 16 million compounds against an AlphaFold2-predicted orphan GPCR yielded a 60% experimental hit rate, demonstrating the viability of the approach [100] Challenges and limitations for understudied targets A major challenge in ligand discovery for understudied proteins is the scarcity of solved holo-structures. Docking to apo structures or AI-predicted conformations typically yields lower hit rates than docking to crystallographic complexes, where the binding pocket is already organized for ligand binding [87]. The induced-fit challenge is further amplified for poorly studied proteins, such as GPCRs and kinases, which undergo large-scale structural rearrangements upon ligand binding. Hybrid AI and molecular dynamics workflows (section “AI-Driven dynamics and ensemble methods”), which generate binding-competent conformations from AI-predicted structures, represent a promising strategy for identifying druggable pockets [101]. The chemical diversity of important biological ligands presents a further challenge. The training data for ML-based docking methods predominantly consist of conventional small-molecule drug-like compounds. Consequently, the performance of these methods on less conventional ligand types, including peptides, glycans, nucleotides, metal complexes, and PROTACs, remains substantially inferior [22, 38]. The stereochemistry of the ligand pose is an additional concern. Both AF3 and Boltz-1 generated poses with notable errors in the bond lengths, angles, and chirality. This may lead to incorrect conclusions in downstream medicinal chemistry studies [38, 46]. Therefore, PoseBuster validation (section “Ligand-specific metrics”) should be performed on all computationally generated protein-ligand complexes prior to their use in experimental study design. Orphan GPCRs represent one of the most challenging classes of understudied drug targets [100, 102]. Several factors contribute to this difficulty. First, predicted structures frequently exhibit inaccuracies in side-chain placement within the orthosteric site [87]. Second, static models cannot account for the dynamic equilibrium between active and inactive receptor states, which governs the pharmacological response [103]. Third, predicted models show persistent distortions in ligand-binding pockets and allosteric sites, driven by training-data biases and a lack of integrated physical and biological principles [104, 105]. Ligand prediction in such system is therefore particularly challenging, because neither AF2 nor AF3 yet achieves the high-resolution side-chain precision needed for reliable drug discovery. It underscores the need to combine complementary computational methods with targeted experimental validation, as outlined in the roadmap presented in the discussion(Section “Conclusion and future perspectives”, figure S2). Case study 1: Orphan GPCRs: from prediction to ligand discovery Orphan G protein-coupled receptors (GPCRs), defined as receptors with unknown endogenous ligands, are rich but challenging drug targets. Over the past decade, AI and machine learning have revolutionized GPCR structural modelling, enabling the prediction of receptor folds directly from sequences [106]. Tools such as AlphaFold2 and RoseTTAFold generate highly accurate 3D models of GPCRs using MSAs [106]. Protein language models, including ESMFold and OmegaFold, do not require MSAs, enabling even faster structure prediction [102, 107]. These tools enable the generation of plausible structural models for completely uncharacterized GPCRs, providing a computational starting point for orphan receptor studies. For example, refinements to AlphaFold2, such as providing receptor templates or biasing the MSA, can capture both active and inactive GPCR conformations [103]. In practice, a given orphan GPCR, such as TAAR1, can be modeled using multiple algorithms, including AF2, RoseTTAFold, ESMFold, and OmegaFold, to build consensus structures and reveal potential ligand-binding pockets [36, 107]. With a predicted receptor structure in hand, computational ligand discovery proceeds via virtual screening. The receptor binding site is typically defined by alignment to known GPCR templates or homology models. Large chemical libraries are then docked into this pocket using tools such as AutoDock Vina or GLIDE or using ML-based docking methods. Docking and ligand-based screening can yield candidate modulators in silico. However, these predictions carry inherent uncertainty, as docking frequently overpredicts binders and misplaces side chains [106]. Nonetheless, in-silico screening can enrich true actives. A recent study on TAAR1 illustrates this workflow effectively. Díaz-Holguín et al. (2024) used an AlphaFold2-derived model of TAAR1, a human trace amine-associated orphan GPCR, to screen approximately 16 million compounds. Of the 30 top-ranked hits tested, 18 were confirmed agonists (0.03–12 μM potency), giving a 60% hit rate, more than double that obtained with a conventional homology model [100]. The best agonist exhibited selective TAAR1 activity in vitro and produced antipsychotic-like effects in wild-type but not TAAR1-knockout mice. This study demonstrates that high-quality AI-predicted structures can meaningfully accelerate ligand discovery for orphan GPCRs [100]. This complete sequence-to-drug discovery pipeline for orphan GPCRs is outlined in Fig. 3. Conceptual workflow of the AI-driven sequence-to-drug discovery pipeline for orphan GPCRs and other understudied protein targets. Starting from a protein sequence (UniProt ID), multiple sequence alignments (MSAs) and template searches are performed using tools such as HH-suite [108] and PSI-BLAST [109]. The resulting features are used by multiple AI-based structure prediction methods, including AlphaFold2, AlphaFold3, RoseTTAFold, OmegaFold, ESMFold, and open-source co-folding models such as Boltz-1 and Chai-1 (section “The open-source prediction Ecosystem”), to generate candidate three-dimensional structures. Predicted models may be refined through molecular dynamics (MD) simulation to explore conformational flexibility and validate stability. Binding pocket prediction is then performed using structural alignment or machine learning-based methods such as PocketMiner (Sect. 6.4). Candidate ligands are identified through virtual screening and docking using engines such as AutoDock Vina [84] or diffusion-based approaches like DiffDock (Sect. 6.3), followed by filtering for drug-likeness and PAINS removal. Top-ranked compounds are selected for experimental validation through in vitro assays. This pipeline has been successfully demonstrated in practice: díaz-Holguín et al. (2024) used an AlphaFold2-derived TAAR1 model to screen approximately 16 million compounds, achieving a 60% hit rate among the top 30 candidates, with one agonist showing antipsychotic-like effects in vivo (section “Case study 1: orphan GPCRs: from prediction to ligand discovery”). Each step is iterative, experimental results can feed back to refine predicted structures, binding pockets, and scoring models Limitations and validation AI models are powerful but imperfect. AlphaFold2 and related tools predict the overall GPCR topology well but often misplace side chains and loop conformations, especially in the dynamic binding pocket [106]. Docking algorithms also tend to generate many false positives [106, 110]. Therefore, every in silico hit must be experimentally vetted using radioligand assays, reporter cell systems, or biophysical methods. Mutagenesis should be used to confirm the predicted receptor-ligand contacts. Likewise, functional selectivity and polypharmacology must be assessed, as AI models cannot fully capture receptor dynamics or cellular context. In summary, AI-based tools, including AlphaFold2, RoseTTAFold, ESMFold, and OmegaFold, enable hypothesis generation for orphan GPCRs but complement rather than replace classical screening and validation techniques. When applied in an integrated manner, this AI-driven pipeline can substantially reduce the time and resources required to progress from sequence to lead compounds for orphan GPCRs [100, 110]. Case study 2: Intrinsically disordered protein regions As introduced in section “Limitations for understudied and orphan proteins”, IDPs lack a stable fold, which makes them an important test case for evaluating prediction methods. Analyses of the AlphaFold Protein Structure Database (AlphaFoldDB) indicate that a considerable fraction of the human proteome has low-confidence AlphaFold predictions corresponding to putative disordered regions [111]. This observation is expected because IDPs do not adopt a single folded conformation by definition, and AlphaFold appropriately assigns low confidence scores to such regions. Research evaluating AlphaFold2‘s performance on IDPs has found that its built-in confidence measures can serve as effective disorder predictors [112, 113]. Specifically, a residue’s pLDDT score has been shown to be a more rigorous indicator of disorder than simple rules based on secondary structure content. When a naïve criterion, such as classifying all non-helix or non-strand residues as disordered, was applied, the disorder was substantially overestimated. In contrast, pLDDT provides a considerably stricter and more accurate measure [113]. Nevertheless, specialized disorder prediction algorithms have been reported to outperform pLDDT in certain cases [113], highlighting that AlphaFold’s outputs, while useful, are not infallible for identifying all disordered residues. Complementing these findings, a large-scale community benchmark, the Critical Assessment of Protein Intrinsic Disorder (CAID), showed that AlphaFold’s predictions are competitive with state-of-the-art tools for disorder detection [111]. In particular, AlphaFold not only detects many intrinsically disordered regions (IDRs) correctly but also identifies certain conditionally folded regions, segments that are disordered in isolation but fold upon binding to a partner [111]. Overall, these studies confirm that AlphaFold2 is generally effective at identifying intrinsically disordered regions while highlighting the value of combining its metrics with established disorder predictors to improve accuracy. Beyond static structure prediction, AlphaFold2 has also shown the ability to capture aspects of protein dynamics relevant to IDPs. Its outputs include per-residue error estimates (pLDDT scores) and inter-residue distance uncertainty (predicted aligned error, PAE maps), which contain information about the flexibility. For many folded proteins, high pLDDT and low PAE values correlate with low experimental B-factors and reduced RMSF in simulations, indicating structural rigidity. Lower confidence regions tend to exhibit greater mobility [114]. Thus, AlphaFold encodes not only structural predictions but also residue-wise flexibility information within these confidence metrics [114]. However, this approach has limitations in the context of fully disordered proteins. For example, Guo et al. (2022) found that when an IDP lacks detectable sequence homologues, AlphaFold’s confidence scores may not reliably reflect its flexibility [114]. In such cases, pLDDT-based scores do not correlate with experimental dynamics. This suggests that the reliability of AlphaFold diminishes for evolutionarily isolated IDPs lacking homologous sequence data [114]. This indicates that AlphaFold2, while powerful, still partly depends on evolutionary information. Truly orphan IDPs may therefore fall outside the reliable range of disorder predictions. Recent advances have begun to leverage AlphaFold2 to model IDP ensembles rather than single static conformations. Although AlphaFold was trained primarily on well-structured proteins, subsequent studies have demonstrated that its internal predictions retain meaningful signals for disordered regions [115]. Brotzakis et al. (2025) harnessed these signals using AlphaFold-derived distance predictions as constraints to generate structural ensembles of IDPs [115]. Although AlphaFold2 returns a single conformation for an IDP, the patterns encoded in its predicted inter-residue distances can be used to sample the broad conformational space occupied by IDPs. This approach, which combines AlphaFold outputs with ensemble modelling and statistical reweighting, provides a means of extracting dynamic information from an originally static prediction framework [115]. This represents a meaningful advance in reconciling AlphaFold’s deterministic predictions with the intrinsic flexibility of IDPs. The resulting conformational ensembles have been shown to align well with the experimental observables. These examples are summarized in Table 3. Table 3Experimental and AlphaFold2-based structural insights into intrinsically disordered proteins (IDPs). The table summarizes UniProt identifiers, experimentally observed disorder-to-order transitions upon ligand binding, AlphaFold2 structural predictions with pLDDT confidence scores, and biological interpretations validating these predictions. References cited correspond to the original experimental and computational findingsProteinUniProt IDExperimental ContextAlphaFold2 PredictionBiological InterpretationKey References4E-BP2Q13542Disordered when unbound; folds into β-strands upon binding eIF4EStructured elements (α-helix + β-strand) in bound-state motif with high pLDDTAF2 correctly flags folded binding region, matching its functionally induced structure[50]ACTR (NCOA3)Q9Y6Q9Transcriptional coactivator; disordered alone, forms helices upon binding CBPFully α-helical prediction with very high confidenceAF2 reveals foldable potential, identifying regions that become ordered upon binding[50]Ephrin-B2P52799Contains structured domains and a disordered C-terminal regionLow pLDDT in C-terminal regionAF2 confidence score correctly maps known IDR, validating the IDR prediction approach[111] Experimental and AlphaFold2-based structural insights into intrinsically disordered proteins (IDPs). The table summarizes UniProt identifiers, experimentally observed disorder-to-order transitions upon ligand binding, AlphaFold2 structural predictions with pLDDT confidence scores, and biological interpretations validating these predictions. References cited correspond to the original experimental and computational findings Collectively, Table 3 compares three cases of IDPs, focusing on the promise and the current limitations of applying AlphaFold2 to intrinsically disordered proteins. On one hand, the AI predictions revealed latent structural propensities in 4E-BP2, ACTR, and Ephrin-B2. It aligned with known conditionally folded states and offered valuable insight into potential functional conformations. On the other hand, these proteins highlight the persistent challenges of dynamic, context-dependent proteins. A single high-confidence model cannot capture the full ensemble of fluctuating structures or the requirement of specific partners and cellular conditions for folding. Looking forward, addressing these challenges will require integrating ensemble modelling to represent multiple conformational states, and context-aware AI frameworks capable of accounting for environmental factors and binding interactions. Together these strategies should markedly improve prediction accuracy for IDPs and other flexible proteins, bridging the gap between static computational models and the dynamic reality of protein function. This integrated perspective provides a natural transition to the broader discussions on how next-generation structural-biology tools can harness AI while respecting the complex behavior of proteins in their native context. AI-Driven dynamics and ensemble methods Proteins are inherently dynamic molecules that exist as ensembles of interconverting conformations rather than fixed rigid structures. These conformational dynamics encompass domain rearrangements, loop fluctuations, allosteric transitions, cryptic pocket openings, and disorder-to-order transitions upon binding. Such conformational states underline most biological functions [103, 116]. Consequently, a single predicted structure, regardless of its accuracy, captures only one snapshot within a broader conformational landscape. This limitation is particularly relevant for understudied protein families. Orphan GPCRs undergo large conformational shifts between active and inactive states, while intrinsically disordered proteins (IDPs) sample vast conformational spaces. Furthermore, many dark proteins exhibit fold-switching behavior that cannot be adequately described by a single structural model [116, 117]. This section reviews AI-driven approaches for modelling conformational dynamics and generating structural ensembles, with particular emphasis on their application to understudied proteins. Generative ensemble models One of the most transformative recent developments is the emergence of generative deep learning models that directly produce structural ensembles, bypassing the need for explicit MD simulations. BioEmu [117, 118], reported in Science, is a diffusion-based generative framework that emulates protein equilibrium ensembles by generating thousands of statistically independent structures per hour using a single GPU. This represents a speedup of four to five orders of magnitude compared with conventional MD. BioEmu integrates three types of training data: over 200 ms of MD simulation trajectories, static structures from the AlphaFold Database, and more than 500,000 experimental protein stability measurements [117]. By combining these heterogeneous datasets, BioEmu captures a wide range of functional motions, including cryptic pocket formation, local unfolding events and domain rearrangements. It also predicts relative free energies with an accuracy of 1 kcal/mol when benchmarked against millisecond-scale MD and experimental data. [117, 119]. However, BioEmu is currently restricted to single-chain protein applications. It cannot model complexes involving protein-ligand interactions, protein-protein interfaces, or membrane-embedded systems. A complementary generative approach, AlphaFlow [119], addresses structural diversity using a flow-matching framework to repurpose the pre-trained AlphaFold2 network. Trained on both PDB structures and short MD trajectories from the ATLAS dataset, AlphaFlow captures local backbone fluctuations on the nanosecond timescale. Although it does not capture the slow, functionally relevant motions that BioEmu addresses, it is computationally more accessible. A related variant, ESMFlow, incorporates embeddings from the ESM-2 language model and eliminates the need for MSA computations. This is particularly advantageous for orphan proteins that lack homologous sequences [119]. Additional generative frameworks include the Distributional Graphormer (DiG), which uses score matching to produce conformational distributions comparable to millisecond-long simulations [120]. Str2Str is a score-based framework that enables zero-shot conformational sampling from crystal structures alone [121]. The conceptual foundations of these approaches are derived from Boltzmann generators [122]. This demonstrated that normalizing flows could learn to sample directly from the equilibrium distribution without performing MD simulations. Although earlier implementations were limited to small proteins of approximately 58 residues, they established the theoretical foundation that BioEmu and related models have since scaled to larger and more complex systems through diffusion-based architectures [123]. In comparative benchmarks, Boltz-2 has shown that training on MD ensembles alongside static structures can also improve dynamics prediction within a co-folding framework, matching or outperforming AlphaFlow and BioEmu on RMSF correlation metrics. [37]. Machine-learned inter-atomic Potentials In addition to generative ensemble models, another strategy for accelerating molecular dynamics involves replacing the empirical force field with a machine-learned model that achieves quantum-mechanical accuracy at a fraction of the computational cost. Machine-learned interatomic potentials (MLIPs) are neural network models trained on quantum mechanical calculations, typically density functional theory, to predict energies and forces on atoms [124]. Unlike generative models, MLIPs do not bypass MD simulations. Instead, they enable MD with improved accuracy and efficiency. The MACE framework [125] employs a higher-order E(3)-equivariant graph neural network with equivariant message passing, achieving high accuracy with small training datasets [126]. Another model, ANI [127], was one of the first transferable neural network potentials for organic molecules. The recent MACE-OFF23 model extends the MACE framework to a transferable force field covering the ten most prevalent elements in organic chemistry [128]. A major emerging trend is the development of foundational MLIPs. Models pre-trained on large and diverse datasets that generalize across the chemical space without system-specific retraining. One example is a transferable coarse-grained MLIP that captures the conformational dynamics across multiple protein systems [129]. MLIPs hold particular value for understudied proteins because they improve the accuracy of MD simulations launched from AI-predicted structures. This is especially relevant for systems involving unusual chemistries, metal-binding sites, post-translational modifications, and non-canonical amino acids. All of these are poorly described by the classical force fields. Hybrid AI and MD strategies: combining prediction with simulation The direct integration of AI-predicted structures with molecular dynamics simulations represents one of the most practical strategies for studying understudied proteins. AlphaFold-seeded molecular dynamics have emerged as a powerful paradigm for exploring conformational landscapes when experimental structures are not available [101]. Meller et al. demonstrated this approach for cryptic pocket discovery by identifying transient binding sites that are absent in ground-state structures but accessible through conformational fluctuations [101]. This was achieved by stochastically subsampling the MSA input of AlphaFold to generate diverse starting conformations, from which MD simulations were subsequently launched [101]. In six of the ten benchmark cases, AlphaFold directly sampled the open pocket conformation, whereas MD simulations completed partial openings in the remaining cases. The resulting Markov state models provided free energy landscapes consistent with well-tempered metadynamics simulations [101]. A similar strategy was applied to the cancer drug target PPM1D phosphatase, where MD simulations from an AlphaFold-predicted structure revealed a cryptic pocket at the flap-hinge interface that was absent in static predictions [130]. Virtual screening against this dynamically revealed pocket correctly reproduced the relative potency of known allosteric inhibitors [130]. These studies provide important insights. AI prediction alone may capture the ground state, but integrating prediction with dynamics can reveal excited states that are often critical for protein function and druggability of the target. Another hybrid framework, AF2-RAVE [131], integrates AlphaFold2 predictions with enhanced-sampling MD and machine-learned collective variables. It computes Boltzmann-weighted rankings of alternative conformational states from protein sequence information. This enables the estimation of the relative populations of conformational states, a capability that static prediction tools fundamentally lack. Ensemble methods for intrinsically disordered proteins IDPs and IDRs populate broad conformational ensembles (section “Case study 2: intrinsically disordered protein regions”). Standard AI predictors assign them low confidence and often produce unrealistic conformations. To address this limitation, several specialized generative models for IDP ensembles have recently been developed [132]. IdpGAN [133] uses generative adversarial networks to produce coarse-grained conformational ensembles, whereas the improved idpSAM employs a latent diffusion architecture with transformers for enhanced transferability across different disordered sequences [132, 133]. Another promising hybrid approach, AF-Meta inference, combines AlphaFold-derived distance restraints with meta-inference-based molecular dynamics to generate solution-state ensembles consistent with NMR and SAXS experimental measurements [115]. This hybrid approach is particularly useful for the disordered regions of the dark proteome, where neither AI prediction nor MD simulation alone can produce satisfactory results (see section “Illuminating the dark proteome: challenges, opportunities, and AI-Driven insights”). Implications and current limitations The dynamics and ensemble methods discussed in this section have significant implications for the study of understudied proteins. Generative models, such as BioEmu, enable the rapid generation of conformational ensembles at scales that make proteome-wide dynamics analysis increasingly feasible. Hybrid AI-MD workflows provide a practical strategy for drug discovery by enabling the identification of cryptic binding sites and transient functional states (see Sect. 6.4, Fig. 3). ML force fields are also beginning to address the limitations of traditional force fields when modelling unusual protein chemistries frequently encountered in poorly characterized and understudied proteins. However, significant limitations remain. Current generative models, including BioEmu and AlphaFlow, are limited to single-chain proteins, and cannot model multi-protein complexes or membrane-embedded systems. Although increasingly transferable, ML force fields still require careful validation in new chemical environments. The quality of any dynamic simulation depends on the accuracy of the initial structure. If the AI predictions of novel folds are unreliable, the resulting simulations may compound these structural errors. Addressing these limitations represents the next frontier in computational structural biology, as discussed in the following sections of this review. Limitations of AI/ML-based protein structure prediction In biological systems, proteins exist as dynamic ensembles of conformations rather than as fixed structures. Several studies have noted that AlphaFold’s outputs typically represent a single optimal fold and therefore lack information about conformational distributions across multiple states [110, 111]. Consequently, static models often fail to capture functional mechanisms, such as allostery or conformational switching, which arise from population shifts among multiple structural states [110, 111]. Experimental methods such as nuclear magnetic resonance (NMR) and computational approaches such as molecular dynamics simulations generate structural ensembles and time-resolved trajectories that better represent this conformational variability. For example, Fowler and Williamson found that while AlphaFold2 predictions frequently match or exceed the structural accuracy of NMR-derived models, they lack the ensemble variability that allows NMR to capture flexible and dynamic regions, particularly when the pLDDT is low [111, 114]. Thus, although modern ML predictors can provide high-resolution structural snapshots, they do not yet represent the dynamic structural landscapes underlying protein function. Ligand binding and Allostery Another important limitation of current AI-based structure predictors is their inability to explicitly model ligand binding and ligand-induced conformational change. Most predictors are trained to fold protein sequences in isolation and, therefore, do not account for bound ligands, cofactors, or interacting compounds. For example, AlphaFold2 models typically exclude metal ions, small-molecule cofactors and substrates. The AlphaFold database contains only polypeptide coordinates, and specialized tools such as AlphaFill are required to incorporate likely ligands into the predicted structures [99]. However, these post-processing approaches have critical limitations. AlphaFill and similar approaches do not support polymeric ligands, such as nucleic acids, peptides, or post-translational modifications, including glycosylation or phosphorylation [99]. Most importantly, ligand binding frequently induces conformational changes either at the active site (orthosteric) or at distant regulatory sites (allosteric effects). Because AlphaFold2 does not simulate ligand-binding events, it cannot predict the structural rearrangements associated with these processes [111]. Consequently, key dynamic regulatory mechanisms, such as allostery, are absent from the predicted structures. As Nussinov et al. noted, “AlphaFold does not capture conformational mechanisms like frustration and allostery, which are rooted in ensembles” [110]. This limitation can be significant in practical applications, such as drug discovery and virtual screening. Docking or virtual screening against static models may fail to capture important induced-fit effects or mis-evaluate binding affinity if the protein undergoes significant conformational changes upon ligand binding. Training biases and environmental context The structural training dataset used for modern AI-based predictors is primarily derived from the Protein Data Bank (PDB). More than 90% of the structures in the PDB are determined using X-ray crystal structures or cryo-electron microscopy (cryo-EM) [111]. Consequently, the training data are inherently biased towards proteins that are experimentally tractable and adopt well-defined conformations under laboratory conditions. These experimentally determined structures typically represent single thermodynamically stable conformers, often captured under non-physiological experimental conditions. Consequently, AF2 predictions tend to reflect “the structured state under experimental conditions where a protein is likely to crystallize,” rather than the full ensemble of conformationally accessible in vivo structures [111]. This bias can lead to the misrepresentation of dynamic regions. For example, flexible loops or alternative domain orientations stabilized by crystal contacts may appear ordered in crystal structures and, therefore, in AI prediction models [111]. Similarly, environmental factors such as membrane embedding, pH variation, molecular crowding, or protein-protein interactions are not explicitly modeled in current prediction frameworks. Consequently, context-dependent states may be underrepresented or absent from prediction models. Underrepresented protein classes Certain protein classes remain challenging for AI-based structural prediction methods. IDRs remain among the most difficult cases (discussed in detail in section “Case study 2: intrinsically disordered protein regions”); because they populate dynamic ensembles rather than a single fold, AI-based methods typically assign low confidence (low pLDDT) to such segments and often render them as unstructured loops [111, 114]. Membrane proteins present another major challenge in proteomics. Many membrane proteins fold within a lipid bilayer environment, creating amphipathic constraints that differ substantially from those encountered by soluble proteins. Membrane proteins are underrepresented in structural databases, making it difficult for models to accurately predict the membrane-embedded topology [49, 58]. Recent benchmarking studies from the TmAlphaFold project show that α-helical transmembrane proteins often violate the expected membrane topology (e.g., incorrect helix placements) in AF2 models, reflecting the limited representation of membrane proteins in the training data [49, 58]. In practice, users should be cautious about trusting AI models for disordered regions, membrane proteins, unusual folds, or engineered variants without additional computational or experimental evidence. Need for experimental validation and hybrid approaches Given these limitations, experimental characterization is essential for interpreting AI-predicted structures. In practice, AlphaFold models are best viewed as hypotheses or starting points for structural solutions. Many researchers have integrated AF2 outputs into workflows, for example, using predicted models to aid cryo-EM fitting or X-ray phasing before further refinement [74, 114]. NMR spectroscopy complements AF2 by revealing dynamics, validating folded regions, and characterizing intrinsically disordered segments [114]. Similarly, molecular dynamics simulations can use AI-generated structures as the starting conformation to explore the structural flexibility and evaluate the stability of the predicted interactions under realistic solvated and dynamic conditions. To ensure the responsible use of predicted structures, researchers should follow the community guidelines. Predicted models should be annotated with confidence scores and provenance and, where possible, cross-validated using independent data. As emphasized by Laurents (2022) [114] and Terwilliger et al. (2024) [134], predicted structures should not be treated as definitive representations of protein architecture but rather as testable hypotheses. Responsible AI use in structural biology The increasing use of AI-based tools in structural biology raises important considerations regarding transparency, data stewardship, and reproducibility. Researchers should clearly document the source of training data for any model use (e.g., PDB or metagenomic sequences) and the known limitations of the prediction models used in any analysis. The structural biology community has begun to establish best practices for responsible AI use. Many deep learning predictors (such as AlphaFold) have provided openly accessible codes, databases, and prediction pipelines. Journals and funding agencies increasingly expect the deposition of predicted structures and associated data to ensure reproducibility. The FAIR Principles (Findable, Accessible, Interoperable, Reusable) are critical for maintaining transparency and enabling reproducible computational research. By combining AI-based predictions with careful experimental validation and transparent reporting of resources, researchers can harness the strength of machine learning while minimizing the risks associated with misinterpretations [135]. Conclusion and future perspectives Artificial intelligence (AI) is transforming structural biology. However, significant challenges persist, particularly for dark proteins, such as orphan GPCRs. The absence of close homologues, combined with significant conformational changes between the active and inactive receptor states, renders conventional modelling approaches largely inadequate for these targets. Intrinsically disordered proteins (IDPs) present a different but equally challenging set of problems in structural biology. They inherently lack stable 3D structures and cannot be accommodated by traditional single structure representations. Similarly, novel folds and membrane-bound proteins pose additional challenges because of their underrepresentation in structural training datasets, whereas large assemblies introduce further size- and environment-related constraints. Furthermore, diffusion-based models such as AlphaFold3 can generate hallucinated ordered structures in regions that are experimentally verified to be disordered, making them particularly risky to apply to understudied targets that lack experimental reference structures [22, 48]. In summary, each of these categories requires specialized treatment. Although AI models can predict overall folds, loops, cofactors, membranes, or multiple subunits can confound their accuracy [50]. To address these challenges systematically, this review introduces a dedicated quality-metrics framework (section “Quality metrics for AI-Based structure prediction”) and separates protein-protein complex prediction (section “Protein-protein complex prediction”) from small-molecule docking (Section “Architectural shift: from Evoformer to Pairformer and diffusion”). We also survey AI-driven dynamics and ensemble methods (section “AI-Driven dynamics and ensemble methods”) and proposed a tiered roadmap for characterizing understudied protein families (Supplementary figure S2). Despite these challenges, recent AI technologies have made tremendous advancements in overcoming them. Recent methods have shown significant improvements in globular protein folding (AlphaFold2, RoseTTAFold), in identifying potential binding sites, and through language model-based approaches (ESMFold and OmegaFold), that operates on a single amino-acid sequence rather than requiring a deep multiple sequence alignment. These approaches can also be used to predict difficult sequences with limited homology [136]. AlphaFold3 [22] has expanded the scope of prediction by jointly modelling proteins, nucleic acids, small molecules, and ions within a unified diffusion-based framework, achieving leading performance on the PoseBusters ligand benchmark (section “Benchmark performance”) [22]. The rapid emergence of open-source alternatives, including Boltz-1, Chai-1, OpenFold3, and Protenix, has democratized access to AF3-level capabilities and collectively formed a competitive ecosystem for biomolecular structure prediction (section “The open-source prediction Ecosystem”) [23, 51]. Hybrid approaches such as D-I-TASSER, which couples deep learning restraints with Monte Carlo physics-based refinement, demonstrated 19% higher TM-scores than AlphaFold2 on the most challenging free-modelling targets, showing that end-to-end deep learning is not the only path to high accuracy [26]. Generative models, such as RFdiffusion and RoseTTAFold All-Atom, demonstrate how AI can design new proteins and binders from scratch, with many de novo designs validated experimentally by crystallography and binding measurements [95, 137]. In practice, researchers often run multiple predictors (AF2, AF3, Boltz-1, Chai-1, ESMFold, D-I-TASSER) in tandem to build consensus structures, a strategy that is particularly valuable for understudied targets for which no single predictor is reliably the best (section “The open-source prediction Ecosystem”). These high-quality in silico models have already enabled notable successes. For example, a study using an AlphaFold2-generated TAAR1 structure to screen approximately 16 million compounds yielded a 60% hit rate of active agonists, more than double the rate achieved with a conventional homology model (section “Case study 1: orphan GPCRs: from prediction to ligand discovery”) [100]. Such integrations of AI and experiments can greatly accelerate the path from gene to lead compounds. The field of structural biology is undergoing a transition from a capturing static snapshot of proteins towards dynamic structure and function. Although current predictive models provide a single “optimal” structure for proteins, in reality proteins exist as conformational ensembles. This shift in focus is currently taking place. For example, generative ensemble models like BioEmu [117] are able to simulate protein equilibrium ensembles at speeds 10,000–100,000 times faster than conventional molecular dynamics. Additionally, AlphaFlow refines AlphaFold2 within a flow-matching framework to produce conformational diversity [119]. Rapid progress has also been made in machine learning-based inter-atomic potentials that enable MD simulations that are nearly quantum mechanical in terms of accuracy at a fraction of the cost of traditional methods [126–128]. Combining AI predictions with dynamics has emerged as a particularly powerful approach. For instance, the AlphaFold-seeded MD pipeline introduced by Meller et al. (2023) enables experimentally validated drug discovery for targets lacking experimental structures: AI prediction alone capture the ground-state conformation. While coupling prediction with MD revealed excited states critical for function and druggability (section “AI-Driven dynamics and ensemble methods”) [98, 101, 130]. Binding affinity prediction, long a bottleneck in computational drug discovery, has seen a paradigm shift with Boltz-2‘s demonstration that deep learning can approach FEP-level accuracy in the CASP16 blind affinity challenge (Section Classical Docking and Scoring Functions) [37]. Beyond dynamics there is also growing emphasis on deriving mechanistic and functional insight from structures. Graph-based networks (e.g. DeepFRI) and new transformer-based models aim to predict functions, ligand interactions, or GO annotations directly from sequence and structure [71, 138]. For IDPs specifically, specialized generators such as idpSAM and the AF-Metainference framework are bridging the gap between static prediction and the conformational heterogeneity that underlies IDP function (section “AI-Driven dynamics and ensemble methods”) [47, 115]. Nevertheless, many understudied proteins remains challenging because of limited training data, algorithmic limitations, or the need for new physics-based approaches. There is currently no definitive answer to which these is the dominant barriers; it is reasonable to assume that overcoming all three will be required for the most difficult proteins targets (section “Limitations of AI/ML-based protein structure prediction”; see also [26, 47, 117]). As the number of AI tools increases, ensuring their responsible and reproducible use is becoming increasingly important. For the fair usage of AI tools, reproducibility requires clear versioning of models and sharing of source codes and data. Ethical considerations, from bias in training data to equitable access, are important in any AI-driven science [139]. The growing open-source ecosystem (Boltz, OpenFold, Chai-1, ColabFold) is encouraging in this regard, ensuring that state-of-the-art prediction capabilities are accessible to researchers regardless of institutional resources [23, 140]. In the spirit of open science, the community should standardize benchmarks for challenging targets, such as novel folds, membrane proteins, and disordered regions, while actively seeking experimental feedback to ground-truth AI predictions. To support this effort, Supplementary Figure S2 illustrates a five-tiered roadmap, from sequence-based characterization through multi-tool structure prediction, functional annotation, dynamics and ensemble modelling, to experimental validation, as a practical framework for the systematic de-orphanization of understudied protein families. In summary, the combined power of AI prediction and experiments promises to illuminate the dark proteome. Realizing this promise will depend on continued innovation in algorithms, close collaboration between computational and experimental researchers, and adherence to Findable, Accessible, Interoperable, Reusable (FAIR) principles for responsible reporting of AI-derived structural data [135, 138]. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Author contributions S.S. conceptualized the study, designed the outline, and wrote the conclusion, future perspectives, and tables. R.S. conducted the case study on IDPs, prepared the tables, and designed the figures. Su.S. supervised the study, reviewed the outline, and performed critical revision and correction of the manuscript. All authors read and approved of the final manuscript. Funding Open access funding provided by Sharda University. Data availability No datasets were generated or analysed during the current study. Declarations Ethics, consent to participate and consent for publication Not Applicable. AI-use transparency Author used AI-based tools were used to improve language clarity and grammar, suggest alternative phrasings, check spelling, create icons and consistency. AI tools were not used to generate scientific content, interpret results, write reference lists, or formulate conclusions. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement No datasets were generated or analysed during the current study.