Proteins: Structure, Function, and Bioinformatics

The advent of deep learning-driven tools such as AlphaFold has revolutionized the prediction of biomolecular structures, offering unprecedented accuracy and accessibility for proteins, RNA, and their complexes. While these tools have demonstrated remarkable success in benchmarking competitions and enabled experimentalists to generate models with ease, their widespread use has also highlighted persistent challenges. These include difficulties in assessing model confidence, limitations in predicting transmembrane domains, nucleic acids, conformational diversity, and interactions with ions or ligands, as well as the tendency to misfold intrinsically disordered regions (IDRs). In this perspective, we critically evaluate the strengths and limitations of current AI-based structure prediction tools through illustrative examples with a particular emphasis on the impact of explicit ion modelling. We notably report how the explicit addition of a few potassium cations to the prediction of IDRs or G-quadruplexes can trigger massive conformational switches compared to "dry" predictions. On this basis, we suggest modelling sequences both "dry" and in the presence of explicit potassium cations as a simple, practical way to sample alternative conformations and to expose disordered regions that current predictors tend to over-fold. We discuss the importance of reporting confidence metrics in publications to avoid overinterpretation. Furthermore, we address the unique challenges of RNA structure prediction, where data scarcity and structural complexity limit the performance of both classical and deep learning methods. Our analysis underscores the need for continued methodological advancements, integration of complementary computational tools, and expansion of high-quality experimental datasets. TOC Graphic O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=111 SRC="FIGDIR/small/740587v1_ufig1.gif" ALT="Figure 1"> View larger version (16K): org.highwire.dtl.DTLVardef@fc2f42org.highwire.dtl.DTLVardef@82c1a4org.highwire.dtl.DTLVardef@7733cborg.highwire.dtl.DTLVardef@1e97db2_HPS_FORMAT_FIGEXP M_FIG C_FIG

Structure-aware proteolysis prediction often ranks candidate cleavage sites using solvent accessibility from static protein structures, yet this assumption is rarely tested against native-state cleavage data. We examined this problem using AprX, a serralysin-family protease that readily hydrolyses caseins but shows lower activity toward native whey proteins. Across six bovine milk proteins, sequence compatibility and local scissile-bond accessibility did not separate caseins from whey proteins. In contrast, accessibility fell progressively in whey proteins as the required contiguous substrate window increased, reaching 0% across the working P4-P4' footprint, while all evaluable casein windows remained accessible in isolated-chain ensembles. Local remodelling showed that whey segments could still adopt AprX-compatible backbone geometry. An external thermolysin benchmark rejected a rigid whole-protein clash metric, but did not alter the contiguous-accessibility result. These findings identify footprint-scale accessibility as a distinct structural variable and show why single-residue solvent exposure can overestimate protease accessibility in folded proteins.

Proteins naturally adopt multiple conformations in mediating cellular processes, and ensemble models are used to fit X-ray crystallography data that captures this heterogeneity. In practice, ensemble refinement produces only minor improvements in agreement with experimental data (R-free) over single-conformation models. It has recently been shown that ensemble models are universally trapped, or "tangled"; refinement algorithms strain each individual conformation in the model to fit the electron density in its immediate vicinity, missing more harmonious ways to arrange the collection of protein conformations to fit the electron density. Here, we demonstrate that this type of trap may be escaped by formulating the construction of low-energy conformations from individual conformer coordinates as an integer linear programming problem. The method successfully recovered the two original protein conformations from a previously published synthetic dataset that traps current refinement methods. Inclusion of the method in an automated refinement procedure with real data is shown to improve R-free and reduce geometric strain in a four-conformation model by comparison with controls. Applying this method in combination with human input and fitting low-occupancy waters to density features in the bulk solvent, we produce models for deposited datasets of three separate 14-19 kDa proteins with greatly improved geometry and R-factors. This includes a 0.77 [A] six-conformation model of the SARS-CoV-2 macrodomain Mac1 (PDB ID: 44PS) with an R-work of 4.7% and an R-free of 6.4%.

A universal mechanism in molecular evolution is functional and structural divergence of members of a protein family. The ability of AlphaFold to predict atomic-resolution protein structures promises to accelerate insights into this process. We study the interplay of changes in sequence, structure, and function in photoactive yellow protein (PYP), a family of bacterial blue light photoreceptors. Halorhodospira halophila contains two PYP homologs that diverged to 60% sequence identity, differ 100-fold in the lifetime ({tau}pB) of their pB signaling intermediate, and display altered peak wavelengths ({lambda}max) for color sensing. We resurrected ancestral PYPs and determined these properties along the resulting recapitulating evolutionary divergence. The resurrected ancestral PYP is functionally similar to PYP1, indicating divergence on the path to PYP2. AlphaFold predictions for PYP2 and these ancestral proteins revealed the absence of structural changes compared to the crystal structure of PYP1. To experimentally validate these predictions, we optimized second-derivative Fourier transform infrared (FTIR) spectroscopy. The FTIR spectra of PYP1 and 2 and their resurrected ancestral proteins demonstrated clear differences in their secondary structure. These results demonstrate an important limitation of AlphaFold and show how ancestral sequence reconstruction combined with spectroscopic approaches yields insights into divergence in a protein family.

Designed binders are routinely filtered on predicted interface confidence, which reports how self-consistently a model places an interface, not where it lies. We designed 200 nanobody loops against the survivin BIR-domain groove, co-folded 196, and checked epitope contact in every predicted complex. Confidence showed no detectable association with epitope engagement ({rho} = +0.100, 95% CI -0.04 to +0.24, n = 196); its highest and lowest quartiles failed alike (odds ratio 0.97, 95% CI 0.41 to 2.26). It tracked instead the run's other confidence estimates ({rho} = +0.61). Twelve-seed replication separated two candidates the criteria had not: one reached the complete epitope in 10 of 12 runs, the other never. Across 155 predictions the top-ranked model was no better than one drawn at random from its five (P = 0.610), though only 3 of 31 failed on every model. Nothing was assayed; nb_cand_057 is an in silico candidate only.

RNA function is dictated by 3D architecture. Although 2D structural models based on canonical Watson-Crick base pairs are widely used, they often fail to capture the non-canonical interactions, tertiary contacts, and coaxial stacking important for biological activity. We have developed RNAbridge, a comprehensive database and web application that systematically identifies, quantifies, and visualizes extended non-canonical helices and multi-way junctions. Using a geometry- and stacking-based pipeline, we analyzed the Protein Data Bank and compiled a catalog of 135,541 structural motifs. RNAbridge includes a user-friendly interface with interactive filters, synchronized 2D and 3D visualizations, and detailed structural data, including helical bend angles and stacking paths. By connecting simplified 2D topologies with complex 3D structures, RNAbridge serves as a valuable resource for structural biologists and lays the foundation for future machine learning applications in RNA structure prediction. The database, available at https://rnabridge.cs.put.poznan.pl/, is automatically updated once a week.

Improvements in computational protein structure prediction have enabled searches for remote homologs of proteins whose molecular function remain unknown. However, the reliability of functional annotations from such searches has not been systematically quantified. In this study, we assembled a time-split benchmark from Pfam families annotated as domains of unknown function (DUFs) in Pfam 28.0 and were subsequently assigned a function in the current release (Pfam 38.0). These retrospective-DUFs provide ground truth for assessing functional annotation from remote homology searches. In our benchmark analysis, we paired retrospective-DUFs with a difficulty-matched arm from known domains to distinguish query difficulty from the method's performance. Across four model proteomes (yeast, C. elegans, Drosophila, and mice), Foldseek searches against AlphaFold/Swiss-Prot, PDB100, and CATH50 recovered the later-assigned function for 15.9% of retrospective DUF queries, compared with 30.1% of matched known-domain queries, after masking for self-family and self-clan level hits to remove circularity. At a fixed confidence cut-off (qTM [≥] 0.5), 55.4% of informative calls on retrospective DUFs were incorrect, establishing an error model for prospective use. Furthermore, in our comparison of tools for homology searches, Foldseek outperformed MMseq2-based sequence search but was not statistically separable from the ESM-2 protein language model embedding baseline. Applying the calibrated structural homology search pipeline to 296 currently unannotated DUF queries in Pfam 38.0 yielded 50 confident, specific functional assignments. We provide the benchmark and ranked candidates as a resource to facilitate functional studies.

MAP4K4 is a serine/threonine kinase member of the mitogen activated protein kinase family. It acts through the JNK, p38 MAPK, and ERK1/2 pathways and is implicated in cancer proliferation and invasion, TNF--driven insulin resistance, macrophage-mediated inflammation, and cardiomyocyte apoptosis in heart failure. No selective small-molecule inhibitor of MAP4K4 has reached clinical use, and knowledge of the specific epitopes on MAP4K4 is lacking for some existing antibodies, leaving a role for small protein binders directed at specific surface sites. We used a computational pipeline that combines RFdiffusion, ProteinMPNN, and AlphaFold2, as well as BindCraft, an integrated "one-shot" tool, for de novo design of small protein binders (~50-130 residues) to specific MAP4K4 surface hotspots. We also created an automated hotspot determination algorithm that weighs geometry, chemistry, rigidity, and AlphaFold pLDDT. From thousands of candidate sequences, we evaluated 20 of the most promising binders with AlphaFold3 (AF3). The interface predicted template modeling (ipTM) scores ranged from 0.16 to 0.90, with nine candidates having ipTM [≥] 0.80, and five scoring [≥] 0.87. Two binders engage non-overlapping hotspots on opposite faces of MAP4K4, making them a candidate pair for a sandwich assay. BLASTp searches of all designed protein sequences returned only low-significance matches to half of them, indicating that they represent truly novel binding solutions and previously unexplored regions of protein sequence space rather than rediscovered natural motifs. We subjected five complexes spanning the observed AF3 confidence range to 100-ns explicit-solvent molecular dynamics simulation. Interchain contacts were retained throughout, with stability varying substantially between systems. Confidence in the binding specificity of the nine highest-confidence candidates was bolstered by juxtaposition with AF3 evaluations of their interaction with CDK2, a negative-control kinase. This comparison yielded a significant, consistent reduction in ipTM (p = 0.0039) for the control binding partner. ToxinPred2 and AlgPred 2.0 screening suggested that one candidate was a potential allergen and two were potential toxins. These results support that de novo design of small, site-specific protein probes for an underserved disease target is achievable using free, publicly available computational tools and minimal resources, pointing to a greater role for the public and amateur scientists to contribute to biotechnological advancement.

The TMEM16 family of calcium-activated phospholipid scramblases (CaPLSs) and chloride channels (CaCCs) performs diverse physiological functions that include regulation of blood coagulation and apoptotic signaling, through a shared ten-transmembrane-helix (TM) architecture organized around a hydrophilic lipid-translocating groove. Mechanistic studies of TMEM16 family members have been hampered by the absence of a unified positional reference framework that would permit direct comparison of structurally equivalent residues across paralogs with different sequence numbering systems. Here we introduce a generic numbering scheme for TMEM16 scramblases (GNS-TMEM16), modeled on the Ballesteros & Weinstein system established for class A G protein-coupled receptors. A reference alignment (TMEM16-RA) was constructed from twelve human and mouse TMEM16 scramblases (TMEM16C/D/E/F/G/J) using structure-based ClustalW alignment of the ten TM helices. From this alignment, a TM-specific reference residue (TsRR) was identified for each helix by hierarchical application of three criteria: (1) 100% conservation in the core TMEM16-RA; (2) conservation in an augmented reference alignment (TMEM16-ARA) incorporating a group of phylogenetically more distant homologs composed of nhTMEM16, afTMEM16, TMEM16K, TMEM16A, and TMEM16B; and (3) structural and functional considerations, including helix-perturbing character, groove localization, conserved motif membership, and central TM position. The resulting ten TsRRs are Y1.50, W2.50, R3.50, E4.50, F5.50, P6.50, E7.50, D8.50, W9.50, and E10.50, and are illustrated in mTMEM16F. Each residue is assigned the identifier N.m(k), where N is the TM number, m is the position relative to the TsRR (for which m = 50), and k is the absolute sequence number. Loop residues receive dual identifiers referenced to the TsRRs of both flanking helices. Application of the GNS-TMEM16 is illustrated with the comparisons of the groove-opening measurements using pairwise distances between residues identified by their N.m indices to be corresponding across mTMEM16F, afTMEM16, and nhTMEM16. The results bring to light the advantages of corresponding residues identification in different TMEM16 proteins and show that the mammalian scramblase undergoes substantially larger separation at the extracellular groove entrance than either fungal homolog. Comparison of mutagenesis data guided by N.m correspondence shows at the conserved (E3.55,R6.26) salt-bridge locus, Ala substitution reduces activity more than 100-fold in nhTMEM16 but less than 2-fold in afTMEM16, illustrating that the GNS identifies structural equivalence of position without implying functional equivalence of the residue, which is a distinct advantage of GNS in providing mechanistic interpretation across paralogs. Also described is a protocol for extending the GNS-TMEM16 to uncharacterized protein sequences, including AlphaFold-predicted models, using structural superposition to mTMEM16F. Thus, the presented GNS-TMEM16 provides a stable positional reference for the integration and comparative analysis of structural, computational, and functional data across the TMEM16 family, utilizing a construction strategy applicable to yet other polytopic membrane protein families sharing a common transmembrane fold.

Antibodies provide programmable molecular recognition, whereas enzymes enable repeated chemical transformation. Catalytic antibodies seek to combine these properties within a single protein scaffold. However, conventional approaches based on transition state analogue immunisation, library screening or local mutagenesis provide limited control over the atomic arrangement of catalytic residues. They also frequently produce antibodies that bind substrates without supporting efficient chemical turnover. Recent advances in generative protein design have enabled the construction of antibody complementarity determining regions and the scaffolding of functional motifs under structural constraints. A systematic strategy for transferring experimentally supported enzyme active site geometry into antibody variable domains is still lacking. Here, we present a computational framework that treats antibody and enzyme structures as distinct but complementary inputs. Developable Fv or VHH structures provide the immunoglobulin scaffold. Enzyme complexes containing substrates, products or transition state analogues provide catalytic residues, ligand conformations, metals, cofactors and key water networks. The selected catalytic atoms are mapped into antibody complementarity determining regions, while the surrounding loops are reconstructed using antibody compatible representations and constrained all atom diffusion. Sequence design and structural back prediction are followed by filters for antibody folding, catalytic geometry, ligand positioning, conformational stability and developability. The framework avoids direct fusion of intact enzymes and antibodies. Instead, it transfers only the local geometry required for catalysis. This separation of scaffold selection from catalytic motif selection creates a testable route for determining whether natural enzyme chemistry can be embedded within antibody formats. It also provides a practical basis for evaluating substrate binding, chemical conversion, product release and catalytic turnover as separate design objectives. Graphical Abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=91 SRC="FIGDIR/small/740676v1_ufig1.gif" ALT="Figure 1"> View larger version (45K): org.highwire.dtl.DTLVardef@16ef6d6org.highwire.dtl.DTLVardef@f7c23org.highwire.dtl.DTLVardef@9ee50borg.highwire.dtl.DTLVardef@1cf42d4_HPS_FORMAT_FIGEXP M_FIG C_FIG

Many computational methods to predict intrinsic disordered regions (IDRs) in proteins have been developed and their performances are blindly evaluated in community-driven assessment, Critical Assessment of protein Intrinsic Disorder (CAID). In this study, we developed PredIDR3 series, an updated version of PredIDR2 tested in CAID3 to accurately predict IDRs from protein sequences. It includes two methods depending on ensemble way. The performances of PredIDR3 series (AUC_ROC=0.953) are remarkably better than our previous PredIDR2 (AUC_ROC=0.936) on Disorder-PDB dataset of CAID3, which is thought to be mainly attributed to the use of more information for intrinsic disorder prediction based on deep convolutional neural network. In details, we introduced a new output-encoding scheme permitting a large sliding window (size=91) for the first time and extracted negative samples of the training set from non-IDRs of both PDB and DisProt databases, allowing to use more information for prediction of intrinsic disorder. PredIDR3 achieved comparable performance to the top-ranking methods of CAID3 in all criteria measured. PredIDR3 series can be freely available through the CAID Prediction Portal at https://caid.idpcentral.org/portal or downloaded as a Singularity container from https://biocomputingup.it/shared/caid-predictors/.

Structure and disorder predictors are increasingly used as decision-grade tools in protein engineering and in the analysis of newly emerged proteins, yet how the current state-of-the-art behaves on sequences outside the well-charted evolutionary space remains poorly characterised. We previously reported that AlphaFold2 confidence and the disorder predictor flDPnn produced discordant predictions for naturally evolved de novo Drosophila proteins and for shuffled sequences. Here, we revisit the comparison with AlphaFold3 and the best-performing disorder predictor PUNCH2 on the same sequence sets together with conserved Drosophila proteins and intrinsically disordered proteins. The discordance persists: pLDDT correlates positively with PUNCH2 disorder in random and de novo proteins and negatively with {beta}-strand fraction, opposite to the conserved and disordered baselines. A class-specific, score-defined driver subset jointly captures the unusual high-pLDDT, high-disorder, low-strand combination and contains 24.5% of de novo, 29.4% of random, 5.1% of conserved, and 1.3% of disordered proteins. Removing this subset normalises the correlations. A held-out classifier trained on architectural and compositional features that were not used in the driver definition recovers the subset, with helix and coil fraction, sequence length, entropy and hydropathy as the strongest predictors. The discordance is therefore not a sequence-class artefact but a localised, compositionally identifiable phenotype that current predictors handle in a non-canonical way - a concrete failure mode that protein designers and others working on sequences remote in sequence space should be aware of when relying on predictor outputs.

In this work, we focus on predicting fold-switching phenomena observed in certain protein structures. Proteins that undergo fold-switching are called metamorphic proteins. We model fold-switch prediction in proteins as a position/residue-wise classification task. We use various features extracted from a protein's sequence as well as structure to train the classifiers. We propose a novel class of handcrafted parametrised features that encode the structural context of a residue. We also use embeddings from existing protein language models (PLMs) to train the classifiers. We train binary and 3-class classifiers, with the 3-class classifiers providing granular information on the type of fold-switch. In our computations, higher classification performance is achieved for binary prediction as compared to 3-class granular prediction. The handcrafted features perform comparably with the PLM embeddings on our standard test set. A rigorous error analysis by us during the review period revealed a critical data-leakage vulnerability in the metamorphic protein test data. This analysis also revealed that the PLM-based classifier suffers sequence-memorisation bias and fails to generalise to novel fold-switching sequences. On the other hand, one of our domain-informed handcrafted feature based classifier detects fold-switching without relying on sequence memory.

Protein structure databases have grown from thousands of experimentally determined structures to hundreds of millions of AI-predicted models, creating an urgent need for search methods that combine high accuracy with practical scalability. Here, I present the third generation of Reseek, a protein structure search algorithm achieving the highest overall accuracy (median rank 1) according to diverse metrics among tested methods including DALI, Foldseek and TM-align. Improved accuracy is obtained by parallel sequence alignment of many discrete alphabets capturing primary, secondary and tertiary features, giving a combined space of[~] 1022 possible states. Separate statistical models are optimized for family, superfamily and fold discrimination, respectively, revealing distinct combinations of features that characterize each level. Hundreds of query structures can be searched a against a multimillion-structure database on a server computer in minutes, making large-scale structure search at state-of-the-art accuracy practical on commodity hardware.

Short, linear sequence motifs within the human prion protein (PrP) may encode local aggregation tendencies that are not apparent from full-length sequence analysis. To map intrinsic amyloidogenic potential across PrP, we generated a complete one-residue-step library of overlapping 15-mer peptides from the 253-residue human PrP sequence and evaluated each peptide using WALTZ in both high-specificity and best-overall-performance modes, with full-length SNPeffect4/WALTZ output used for comparison. Peptide-level WALTZ analysis identified several candidate amyloidogenic regions, including an N-terminal signal-peptide segment spanning approximately residues 8-21/22 that was not detected in the publicly available full-length WALTZ/SNPeffect4 output. Predictions made using AlphaFold 3.0 also identified a short C-terminal area whose representative peptides formed either -helix or pair of {beta}-strands, indicating a region of potential higher susceptibility to conformational dynamics. This computational study supports the use of peptide tiling as a complementary screening strategy for identifying candidate short aggregation-prone motifs in PrP and other misfolding-associated proteins.

2-Deoxyribose-5-Phosphate Aldolase (DERA) is a key enzyme in the pentose phosphate pathway. Due to its C-C bond formation and stereoselective capabilities, DERA has been widely used for biocatalytic applications including the synthesis of chiral intermediates for antiviral and anticancer drugs. While protein engineering has expanded its substrate pool, improved yield, and enhanced stereoselectivity, the molecular basis of stereoselectivity remains unclear. Here, we determined the crystal structures of wildtype DERA from Geobacillus sp. and two of its variants with opposite stereoselectivity. Using a combination of structural biology, biochemistry, organic synthesis and molecular dynamic simulations, we show that the catalytic Lysine adopts two conformations and the Lysine conformation is a key determinant of DERA stereoselectivity. We also identified a mechanism of regulating stereoselectivity via a key amino acid. Using DERA from E. coli, we show that these findings are most likely conserved among bacteria.

Electrostatic interactions often play a role in determining the thermodynamic and kinetic properties of protein-protein interactions. However, the role of long-range electrostatic interactions in intrinsically disordered protein (IDP) binding is less clear, as they often bind in multiple steps including initial formation of a disordered encounter complex, followed by rearrangement into the bound state. We varied the salt concentration to probe the role of long-range electrostatic interactions in the binding of the highly charged AbpSH3 domain and the oppositely charged IDP ArkA. Using isothermal titration calorimetry, we observe that salt enthalpically destabilizes the bound complex. Molecular dynamics and NMR experiments reveal that salt has little effect on the bound state structure. However, simulations show that salt destabilizes the encounter complex intermediate, which primarily affects the association rate as measured by NMR. Consistent with these results, salt has the largest stabilizing effect on the apo SH3 domain, as cations substitute for the transient and long-range electrostatic interactions that can form with ArkA in the complex. We reveal a detailed picture of how a highly charged domain uses long-range, fuzzy, electrostatic interactions to help reach the bound state, a mechanism that is likely common among other highly charged domains that bind IDPs. TOC Image O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=96 SRC="FIGDIR/small/741257v1_ufig1.gif" ALT="Figure 1000"> View larger version (21K): org.highwire.dtl.DTLVardef@dfc7acorg.highwire.dtl.DTLVardef@1ae0438org.highwire.dtl.DTLVardef@19704adorg.highwire.dtl.DTLVardef@1b40b15_HPS_FORMAT_FIGEXP M_FIG C_FIG

Machine learning-guided protein sequence redesign is now routinely used to optimize multiple properties relevant for protein engineering, most prominently thermostability and recombinant expression levels. Salt tolerance is a valuable property for "blue-biotechnology"- enabled biomanufacturing, yet no generative computational method exists to redesign proteins for increased salt tolerance. We hypothesized that a training dataset heavily biased toward salt-adapted proteomes would yield a model capable of designing proteins with halophilic properties. To test this, we retrained the sequence redesign model ProteinMPNN on proteins from "salt-in" extreme halophiles such as Haloarcula marismortui, a Dead Sea archaeon that grows optimally near 3-4 M NaCl, roughly six times the salinity of seawater, and accumulates molar concentrations of salts in its cytoplasm. Our model, HaloMPNN, redesigns non-halophilic proteins so that their properties shift towards those of natural halophilic proteins: lower predicted isoelectric point, greater surface acidity, and reduced surface and core hydrophobicity. Redesigning a broad range of non-halophilic proteins with SolubleMPNN, ProteinMPNN, and HyperMPNN shows that this shift is specific to HaloMPNN rather than a generic consequence of sequence redesign. HaloMPNN therefore offers both a route to designing candidate salt-tolerant enzymes and a means of identifying the characteristics that underlie halophilic adaptation.

Recent advances in AI- RNA structure prediction have increased the availability of candidate models, but their accuracy remains variable. Developing reliable quality assessment (QA) methods requires datasets that represent contemporary prediction errors across diverse targets and molecular contexts. Here, we present RNA3D-Decoy, a large-scale benchmark dataset comprising 103,549 decoys for 677 experimentally determined targets, including 416 RNA-only structures,117 RNA structures with metal ions, and 144 RNA-protein complexes. Decoys were generated using five biomolecular predictors: AlphaFold 3, ESMFold2, Protenix, Chai-1, and Boltz-2.Each decoy is evaluated against its experimental reference us-ing TM-score, lDDT, and RMSD, with global and local quality annotations where applicable. DockQ scores additionally characterize RNA-protein interface quality. Target grouping incorporates both RNA and protein sequence relationships to support dataset partitioning. Analyses reveal distinct quality distributions across dataset levels and target-dependent differences among predictors, with each predictor contributing the highest-TM-score model for a subset of targets. Broad pooled score distributions also coexist with limited within-target score variation, particularly for RNA-protein complexes, emphasizing the importance of target-level characterization. By combining AI-generated decoys, standardized quality labels, and structured metadata, RNA3D-Decoy provides a resource for developing and evaluating methods for global quality estimation, local error prediction, and candidate-model ranking. The data set and accompanying code are available at: https://github.com/HamiltonChenLab/RNA3D-Decoy.

Many functional RNA motifs adopt structures that deviate from the canonical A-form helix and are emerging targets for RNA-directed therapeutics. The microtubule-associated protein tau (MAPT) A-bulge motif (5'-GCAGU/5'-ACGU) is one such motif. Because its structure is stabilized by a delicate balance of local interactions, its accurate modeling remains a major challenge for molecular dynamics (MD) simulations. The experimentally determined nuclear magnetic resonance (NMR) structure of the MAPT A-bulge motif provides a stringent test of whether RNA force fields can accurately reproduce the experimentally observed conformation. Most current AMBER-family RNA force-field models have incorrectly favored a non-native base-triple state of the MAPT A-bulge motif over the experimentally observed stacked state. Structural comparison of the stacked and base-triple conformations revealed that overly favorable NH-N hydrogen bonds between the bulged adenosine and an adjacent Watson-Crick base pair were the primary source of this imbalance. We developed gHBfix-18Ab, an 18-component hydrogen-bond correction that distinguishes NH and NH2; donors. gHBfix-18Ab was combined with the previously developed OL3CP and NBfix0BPh corrections to generate the composite model gHBfix-18Ab*. This model restored the experimentally observed stacked state as the global minimum in the calculated free-energy profile and improved agreement with NMR-derived distance data for the A-bulge region. Importantly, gHBfix-18Ab* did not produce marked structural destabilization of the cUUCGg tetraloop, a widely used benchmark for RNA force-field validation, suggesting that the refinement preserves the stability of the unrelated RNA motif. These results demonstrate that targeted refinement of hydrogen-bond interactions provides a practical strategy for systematic improvement of RNA force fields toward more accurate modeling of noncanonical RNA motifs.