Comprehensive evaluation of artificial intelligence-empowered approaches for protein–aptamer complex prediction 1. The review benchmarks four AI structure predictors (AlphaFold3, Chai-1, Boltz-2, RoseTTAFold2NA) plus a template-based method (3dRNA/DNA) on protein–aptamer complexes, focusing not only on geometric accuracy but also on MD-based stability and MM/GBSA energetic consistency. 2. A key finding is limited aptamer specificity: across 11 curated post-AF3-cutoff PDB complexes, predicted binding free energies for native (positive) pairs often overlapped with intentionally mismatched protein–aptamer pairs, suggesting models can generate “bound-like” complexes even for non-cognate sequences. 3. AF3 was additionally stress-tested with shuffled-sequence controls (same length and nucleotide composition; ≥80% positions changed). Even these nonbiological aptamers frequently produced complexes with seemingly reasonable MM/GBSA trends, reinforcing that sequence-level specificity is not reliably captured. 4. The benchmark design is a central contribution: it includes a strict nonbinding control set (rotated aptamer assignment with constraints to avoid homologous/cognate swaps, e.g., thrombin aptamers not cross-paired) and evaluates 25 samples per target (5 seeds × 5 diffusion samples) under a matched sampling regime across AF3/Chai-1/Boltz-2. 5. Interface accuracy was assessed with iLDDT (OpenStructure). Boltz-2 achieved the highest overall iLDDT distribution on the full set, but accuracy dropped on complexes outside its training window; several hard targets (e.g., 8TFD, 9GXH) remained challenging across models. 6. Interfacial hydrogen-bond networks were evaluated against experimental structures (HBPLUS precision/recall/F1). Even when total H-bond counts looked similar, models often missed native H-bonds and introduced non-native ones, indicating partial recovery of interaction determinants rather than faithful interface reconstruction. 7. MD simulations (20 ns, CHARMM36m, standardized NaCl environment) showed many predicted complexes remain physically stable (Rg/RMSD, persistent contacts), including mismatched negatives—highlighting that “stability in MD” does not imply correct molecular recognition. 8. Confidence calibration is flagged as a practical issue: comparing ipTM vs iLDDT shows that high internal confidence does not guarantee accurate protein–aptamer interfaces; Boltz-2 in particular tended to assign uniformly high ipTM even when iLDDT was low, consistent with overconfidence in this setting. 9. Ion inclusion was tested for ion-containing complexes by providing explicit ions to AF3/Chai-1/Boltz-2. Interface accuracy changed little (small improvement only for Boltz-2), and predicted ion placement often disagreed with experimental arrangements, implying ion-dependent aptamer folding/recognition is still not well captured by current pipelines. 📜Paper: doi.org/10.1093/bib/bbag206 #Aptamers #ProteinStructure #ComplexPrediction #AlphaFold3 #RoseTTAFold #Benchmarking #MolecularDynamics #ComputationalBiology #DrugDiscovery #NucleicAcids