New Web Server Predicts Amyloid Fibril Structures from Peptide Sequences

Scientists in Warsaw have unveiled a computational tool that tackles one of the most stubborn problems in structural biology: predicting how short, disordered peptides stack together into the amyloid fibrils implicated in Alzheimer’s disease, Parkinson’s disease, prion disorders, and type II diabetes. The tool, called InRegPep, is a freely accessible web server described in Frontiers in Molecular Biosciences by Wojciech Puławski, Szymon Niewieczerzał, Andrzej Koliński, Grzegorz Firlik, and Michał Koliński of the Polish Academy of Sciences and the University of Warsaw. Rather than betting on a single structure, the method generates large ensembles of candidate aggregates and then filters them down to the most physically plausible protofilament models.

Amyloid fibrils are built from layers of peptides stacked like bricks, most commonly in so-called in-register parallel beta-sheets, in which identical residues in neighboring chains align laterally and the strands run in the same direction. The precise packing of these sheets determines a fibril’s toxicity, its ability to seed new aggregates, and its capacity to propagate between cells. Experimental techniques such as cryo-electron microscopy, atomic force microscopy, and solid-state nuclear magnetic resonance have delivered a growing catalogue of high-resolution fibril structures, and dedicated databases now support systematic comparison of these assemblies. Yet the known structures remain sparse relative to the enormous diversity of amyloidogenic sequences and the wide range of conditions under which fibrillation can occur.

Artificial intelligence systems such as AlphaFold have transformed the prediction of folded proteins and their complexes, but amyloids present a fundamentally different challenge. These methods are trained largely on well-folded, monomeric proteins and typically aim to identify a single native conformation at the global minimum of an energy landscape. Amyloid fibrils, by contrast, are polymorphic: a single peptide sequence can adopt multiple structurally distinct fibril architectures depending on environmental conditions, and the energy differences between these alternative packings can be subtle. Short, intrinsically disordered peptides also lack the evolutionary signals that power modern AI predictors, and benchmarking studies have shown reduced accuracy for sequences dominated by flexible coil regions. A recent AI approach called RibbonFold has begun to address polymorphism by embedding parallel in-register constraints and a polymorphism-aware loss function into an AlphaFold2-derived framework, but the field clearly needs complementary strategies.

InRegPep takes a deliberately different route, built around multiscale simulation rather than machine learning. At its core lies the CABS coarse-grained model, a long-established framework in which each amino acid residue is represented by up to four pseudo-atoms: the alpha carbon, the beta carbon, a united atom for the remainder of the side chain, and a pseudo-atom used to define backbone hydrogen bonds. The alpha carbons sit on a dense three-dimensional lattice while the remaining pseudo-atoms float off-lattice, a hybrid scheme that preserves near-atomistic backbone resolution of roughly half an angstrom while dramatically cutting computational cost. Because lattice positions restrict conformational space to a set of inexpensive local moves, the model can sample vast stretches of conformational landscape, and its statistical force field explicitly favors extended beta-strand conformations and inter-chain contacts, exactly the interactions that stabilize amyloid-like arrangements.

The workflow begins when a user submits a peptide sequence of 10 to 100 amino acids along with parameters such as secondary structure assignments, the number of docking runs, Monte Carlo cycles, temperature range, pH, and optional restraints mimicking disulfide bridges. Distance restraints between corresponding residues in neighboring chains bias the simulation toward the in-register parallel arrangement. Five identical peptides are then docked simultaneously, with every chain remaining fully flexible throughout the process. Up to 100 independent docking runs can generate as many as one million unique aggregate models, each represented as a simplified alpha-carbon trace.

Selection is where the pipeline earns its keep. Twenty candidate aggregates are chosen from the generated pool using a combination of energy-based criteria, a structural symmetry measure called pcaRMSD, and k-means clustering. These survivors are reconstructed into all-atom representations, geometry-optimized, and subjected to molecular dynamics simulations with the OpenMM engine. Stability is judged from changes in RMSD and pcaRMSD along the trajectories, intermolecular contact analysis, and conservation of secondary structure. The top ten models are ranked, each classified as plausible or unstable, and for plausible aggregates the server assembles short protofilaments of fifteen peptides by replicating the central layer along the fibril’s long axis using transformation matrices derived from the predicted aggregate geometry.

The benchmark results are striking. Tested against fifteen experimentally determined protofilament structures solved by NMR or cryo-EM, all with in-register parallel beta-sheet architecture and peptide lengths from 22 to 40 residues, InRegPep produced a model within 5 angstroms RMSD of the experimental structure among its ten top-scoring predictions for ten of the fifteen systems. In one case, a fibril from a beta-2 microglobulin fragment, the best model came within 2.52 angstroms. The authors note that near-native structures appeared in the docking trajectories for almost all systems, but that ranking them correctly remains the hardest step, a limitation they acknowledge openly. The method also performs better on shorter peptides; tests on longer targets such as an 88-residue tau fragment found polymorph-like structures in the trajectories, but the scoring function failed to elevate them into the top ten.

Beyond single-structure prediction, the ensemble-oriented design makes InRegPep a platform for exploring amyloid polymorphism itself. Analysis of docking trajectories for tau and amyloid-beta 42 showed that the generated ensembles include structures closely resembling experimentally determined polymorphic conformations, or at least capture key structural motifs of these assemblies. Because amyloid-forming peptides populate multiple competing minima corresponding to fibrils that can differ in morphology, stability, and biological behavior, viewing amyloid structures as ensembles of accessible states rather than single architectures is increasingly seen as essential. The server’s downloadable output includes cluster centroids from independent docking trajectories, giving researchers a practical map of the conformational landscape for their sequence of interest.

The practical implications extend across medicine and nanotechnology. Accurate protofilament models can guide the design of aggregation inhibitors, molecules that modulate fibril growth, and compounds that bind selectively to particular fibril forms, which matters because different polymorphs of the same protein are associated with different diseases and disease subtypes. The rational engineering of peptides with altered aggregation propensity also opens doors to biomaterials and nanostructured devices built from amyloid-like assemblies. The authors suggest their models can serve as foundations for mutational design and for downstream experimental or computational validation.

InRegPep is freely available at bioinfo.imdik.pan.pl/inregpep, with benchmark data deposited on Zenodo, and the work was supported by the National Science Center of Poland. It arrives at a moment when cryo-EM is revealing ever more fibril structures, including brain-derived polymorphs of amyloid-beta and tau, yet the computational toolkit for predicting new assemblies from sequence alone remains thin. By combining efficient coarse-grained sampling with symmetry-aware scoring, clustering, and atomistic refinement, the Warsaw team has delivered a workflow that complements rather than competes with AI-based prediction, and one that embraces the polymorphism that makes amyloids both so dangerous and so scientifically fascinating.

Subject of Research: Computational multiscale modeling of amyloid peptide aggregate and protofilament structures

Article Title: InRegPep: a web server for modeling of parallel beta-sheet fibrillar aggregates of amyloidogenic peptides

Article References: Puławski, W., Niewieczerzał, S., Koliński, A., Firlik, G., & Koliński, M. (2026). InRegPep: a web server for modeling of parallel beta-sheet fibrillar aggregates of amyloidogenic peptides. Frontiers in Molecular Biosciences, 13, Article 1908797. https://doi.org/10.3389/fmolb.2026.1908797

Image Credits: AI Generated

DOI: 10.3389/fmolb.2026.1908797

Keywords: amyloid fibrils, InRegPep, CABS model, coarse-grained simulation, molecular dynamics, protein aggregation, amyloid polymorphism, protofilament, structure prediction, web server, neurodegenerative disease, computational biology