The historical reliance on inference rather than observation for hydrogen positions deserves some elaboration. X-ray crystallography detects electrons, not nuclei, and hydrogen atoms, with their single electron participating in the bond, scatter X-rays weakly relative to carbon, nitrogen, and oxygen. In small-molecule crystallography, where crystals are typically highly ordered and data can be collected to very high resolution, hydrogen positions can often be refined directly or placed with confidence using geometric constraints. In macromolecular crystallography, by contrast, protein crystals diffract to lower resolution, contain enormous amounts of solvent-disordered content, and rarely permit direct localization of hydrogens on ligands. Neutron crystallography can locate hydrogen nuclei explicitly because neutrons scatter from nuclei rather than electron clouds, but neutron sources are scarce, beamtime is limited, and growing large protein crystals suitable for neutron work is difficult. As a result, the tautomeric state of a ligand bound in a protein structure is usually assigned by software or by the crystallographer’s judgment, using rules of thumb about which heteroatoms are likely protonated at physiological pH. These assignments are then inherited by downstream users of the structure, often without re-examination.

The scale of the Cambridge Structural Database mining effort described in the study is worth appreciating in context. The CSD has accumulated small-molecule crystal structures for over a century, and its high-resolution entries preserve the geometric detail needed to distinguish, for example, a hydroxyl oxygen from a carbonyl oxygen, or an imidazole nitrogen bearing a proton from one that does not. By systematically extracting more than 1.1 million tautomeric states from these structures, the researchers effectively converted a repository built for crystal engineering and structural chemistry into a training corpus for machine learning. This kind of data repurposing has become an increasingly productive strategy in computational chemistry, where experimental archives originally collected for one purpose turn out to contain the labels needed to train models for another. The orders-of-magnitude increase over prior tautomer datasets, which typically contained a few hundred experimentally characterized examples in solution, changes what is statistically learnable: rare tautomeric motifs, substituent effects, and heteroaromatic patterns that would be absent or underrepresented in smaller collections can now inform predictions.

The choice of a graph neural network operating on two-dimensional molecular input is also consequential. Two-dimensional representations encode the connectivity of atoms and bonds but not the three-dimensional conformation, meaning the model must learn tautomer preferences from topology and chemical context alone. This is a deliberate trade-off. Requiring a 3D structure would mean generating conformations first, introducing an additional source of error and computational cost, and would tie the model’s accuracy to the quality of conformer generation. By working directly from the 2D form, the tool can be applied at the earliest stages of library triage, before any geometry has been computed, which is exactly where tautomer decisions have the greatest leverage since they propagate into every subsequent calculation.

The finding that roughly 2.5 percent of ligands in the PDBbind set may carry incorrect tautomer assignments warrants careful interpretation. PDBbind is a curated collection of protein-ligand complexes drawn from the Protein Data Bank, widely used as a benchmark for docking and binding affinity prediction. If a few percent of its ligands have misassigned protonation or tautomeric states, then benchmark results that appear to reflect method performance may in part reflect data noise. The researchers are careful to note that the experimental structures themselves are not necessarily wrong; rather, the chemical annotation layered onto them may need revision. This distinction matters because the electron density in a crystal structure constrains the heavy-atom framework, and a reassigned tautomer often fits the same density equally well while offering more chemically reasonable hydrogen bonding. The example highlighted in the release, where reassignment enables additional hydrogen bonds with nearby protein residues, illustrates the practical payoff: a corrected tautomer can transform an apparently suboptimal binding pose into a coherent one.

Hydrogen bonding is particularly sensitive to tautomeric state because it depends on which atoms carry hydrogens and which act as acceptors. A pyridone and a hydroxypyridine, for instance, present entirely different donor-acceptor patterns to a binding site despite sharing the same atomic composition. In structure-based drug design, where hydrogen bonds are often the dominant specificity-determining interactions, an incorrect tautomer can lead medicinal chemists to draw the wrong conclusions about why a ligand binds or how to improve it. Free energy perturbation calculations and other rigorous simulation methods are similarly affected, since they require a fully specified molecular topology including every hydrogen. An error introduced at the tautomer assignment stage cannot be corrected later by more sampling or better force fields; it is baked into the simulation from the start.

The throughput figures reported for the released tool give a sense of its practical reach. Processing roughly 4.6 million compounds in 3.2 hours on a single GPU-enabled node corresponds to on the order of hundreds of molecules per second, fast enough to annotate entire corporate or public screening libraries overnight. This matters because tautomer enumeration is combinatorial: a molecule with several heteroatoms and exchangeable protons can generate many candidate tautomers, and exhaustively evaluating each with quantum mechanics is infeasible at library scale. A learned model that ranks candidate tautomers by predicted stability provides a practical filter, allowing expensive quantum or experimental follow-up to be reserved for the most plausible candidates.

It is also worth situating this work within the broader movement toward physics-aware machine learning in chemistry. Rather than attempting to predict binding affinities or biological activity end-to-end, the model targets a discrete, well-defined chemical inference problem that sits upstream of many downstream tasks. Problems of this kind, where the answer is a structured chemical choice rather than a continuous property, can be validated against experimental ground truth and integrated into existing workflows as drop-in improvements. The open-source release of the tool as Tautomer-Predictor follows a pattern increasingly common in computational chemistry, where reproducibility and community adoption are treated as integral to the research contribution rather than afterthoughts.

Certain limitations naturally accompany this approach. The training data derive from small-molecule crystals, which reflect tautomeric preferences in the solid state, where packing forces and crystal environments can shift equilibria relative to solution or to the protein binding pocket. The study’s authors and the field more generally recognize that tautomer populations are environment-dependent: a form favored in a crystal may not dominate in water, and the protein microenvironment, with its own hydrogen bond donors, acceptors, and dielectric properties, can stabilize forms that are minor in bulk solution. Models trained on crystallographic data therefore capture one informative slice of tautomer stability, and their predictions are best understood as strong priors to be refined in context rather than final answers. Nevertheless, the availability of a fast, data-driven prior represents a meaningful advance over rule-based tautomer perception, which has historically relied on hardcoded chemical heuristics that vary between software packages and can disagree with one another.

The funding context, an NIH grant supporting methods development for molecular modeling, reflects the recognized importance of infrastructure-level improvements in drug discovery research. Individual breakthroughs in target identification or clinical validation attract considerable attention, but cumulative progress often depends on quietly fixing foundational layers of the computational pipeline. Tautomer assignment is one such layer. As machine learning methods continue to permeate molecular design, the quality of the chemical representations fed into them becomes a limiting factor, and efforts to clean and standardize those representations, whether through better data, better models, or better experimental characterization, compound across the entire discovery enterprise. The work from NYU exemplifies this genre of contribution: not a new drug or a new target, but a repair to the substrate on which countless downstream calculations rest.

Subject of Research: Scientists use AI to find stable forms of molecules for drug discovery

Article Title: Scientists use AI to find stable forms of molecules for drug discovery

Article References: Scientists use AI to find stable forms of molecules for drug discovery. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: tautomers, deep learning, drug discovery, chemical science, hydrogen positions, graph neural network, Cambridge Structural Database, Protein Data Bank, molecular modeling, NYU, Scientists, find

The proteasome is a massive 2.5-megadalton complex whose 20S catalytic core is capped by 19S regulatory particles that unfold ubiquitinated substrates and thread them into the destruction chamber. A narrow channel of roughly 15 angstroms in the alpha ring forces every substrate to enter in an extended, unfolded state. Structural studies of proteases bound to their substrates have long suggested that the polypeptide chain must flatten into an extended beta-strand conformation inside the catalytic groove, and modern inhibitors such as carfilzomib and bortezomib mimic exactly that geometry. What remained unclear was whether the conformation a substrate region adopts before it reaches the active site leaves any detectable signature on where the cuts actually fall.

Previous work on proteasomal specificity has focused almost entirely on sequence: the beta5 catalytic subunit favors hydrophobic residues, beta2 prefers basic ones, and the immunoproteasome’s swapped subunits further tune those preferences. But sequence motifs alone cannot explain why cleavage appears biased toward particular structural contexts. The researchers asked a deceptively simple question: how much cleavage signal survives when amino-acid identity is stripped away entirely, leaving only information about whether each residue sits in a helix, a beta-strand, or a coil?

To find out, the team first mined a corpus of 1,505 non-redundant human and viral protein structures from the Protein Data Bank, all resolved at 2.5 angstroms or better. They scored every residue with established sequence-based cleavage predictors, including NetChop 20S, NetChop C-term, ProteaSMM-C, and ProteaSMM-I, then compared the secondary-structure composition of the top 10 percent of predicted cleavage sites against the background composition of the whole corpus. The result was striking: high-confidence predicted cleavage sites were consistently enriched in beta-strand, or extended, conformations and depleted in coil-like regions. The beta-strand signal peaked near the cleavage site itself, forming a bell-shaped profile that faded with distance, and it was even stronger for predictors trained on immunoproteasome and MHC class I ligand data.

The reciprocal pattern appeared at unfavored sites. There, coils were enriched while helices and beta-strands were underrepresented relative to expectation. Crucially, the same tendencies held when the researchers analyzed true experimental cleavage sites from a 20-protein digestion dataset combining published in vitro data, and they held whether secondary structure was predicted from sequence by NetSurfP-3.0 or assigned directly from crystallographic coordinates by STRIDE. A protein-level analysis reinforced the point: mean predicted coil content correlated inversely with mean predicted cleavage probability across the 1,505-protein corpus and across 22 experimental digestion maps, suggesting the association is not an artifact of residue-level statistical non-independence.

The team then went a step further and built a machine learning model that saw no amino acids at all. Their proof-of-concept classifier, dubbed SSO-8, represented each candidate cleavage site as an eight-residue window spanning positions P4 to P4′ on either side of the scissile bond, encoded solely as 64 eight-state secondary-structure probabilities predicted by NetSurfP-3.0. Trained on the 20-protein digestion dataset, which contains 1,215 experimentally determined cleavage sites, the final SSO-8 model was a linear support-vector machine selected by protein-grouped cross-validation. On held-out validation benchmarks derived from 16 HIV-derived peptides, SSO-8 achieved an area under the ROC curve of 0.649 on the constitutive proteasome dataset and 0.737 on the immunoproteasome dataset.

Those numbers become remarkable in context. A matched sequence-only control, trained with the identical protocol but using full one-hot amino-acid encodings, reached AUC values of 0.674 and 0.736 on the same benchmarks, and paired statistical tests found no significant difference between the structure-only and sequence-only models. In other words, secondary structure alone captured nearly the same cleavage signal as complete amino-acid identity, while using 2.5-fold less input information. On the immunoproteasome benchmark, SSO-8 even edged out the widely used NetChop 20S predictor, which reached an AUC of 0.711, and matched NetChop C-term at 0.731.

To test whether structural information adds anything beyond sequence, the researchers constructed hybrid position-specific scoring matrices, or PSSMs, that aggregated predictions from multiple methods. Adding SSO-8 to a sequence-only ensemble nudged the immunoproteasome benchmark AUC from 0.781 to 0.797, and the best five-method ensemble reached 0.806. The improvement was directionally consistent but not statistically significant, with a paired permutation p-value of 0.30, so the authors are careful to present it as suggestive rather than conclusive. Still, the pattern supports the idea that local backbone conformation carries cleavage information partly orthogonal to sequence motifs.

The team also ran a rigorous audit for data leakage. Pairwise Needleman-Wunsch alignments confirmed that no protein pair between the two training subsets, or between the training set and the validation peptides, exceeded the standard 30 percent identity threshold for functional relatedness. And because the PSSM octamer labels were derived purely from predicted scores rather than experimental annotations, experimental data from training proteins could not circularly inflate the corpus-based results. A post hoc analysis added another intriguing layer: beta-strand frequency at true immunoproteasome cleavage sites was 32.5 percent, significantly higher than the 18.7 percent observed at constitutive cleavage sites, indicating that the two proteasome types carry distinguishable structural signatures.

Mechanistically, the authors propose that strand-like geometry may simply reflect compatibility with the extended conformation the proteasome demands of its substrates, rather than selection for intact native beta-strands before engagement. The signal was consistently stronger for the immunoproteasome, the isoform induced by interferon-gamma during inflammation that specializes in producing ligands for MHC class I presentation. Because proteasomal cleavage generates the C-terminal anchor residue of presented peptides, the researchers hypothesize that this beta-strand bias may steer the immunoproteasome’s output toward structurally constrained regions of proteins, where escape mutations are less tolerated by the virus or tumor. If confirmed, incorporating structural features into next-generation cleavage predictors could improve epitope and neoantigen discovery for vaccine development and cancer immunotherapy, though the authors caution that their model was trained on limited in vitro data lacking cellular context, and that protein language models may offer an even richer structural representation for future work.

Subject of Research: Association between substrate local secondary structure and proteasomal cleavage propensity

Article Title: Substrate secondary structure is associated with proteasomal cleavage propensity

Article References: Gutierres, M. D. B. B., Beruldsen, F., Antunes, D. A., & Vieira, G. F. (2026). Substrate secondary structure is associated with proteasomal cleavage propensity. Frontiers in Bioinformatics, 6, Article 1730535. https://doi.org/10.3389/fbinf.2026.1730535

Image Credits: AI Generated

DOI: 10.3389/fbinf.2026.1730535

Keywords: proteasome, immunoproteasome, secondary structure, beta-strand, machine learning, MHC class I, T-cell epitopes, cleavage prediction, immunoinformatics, neoantigen, Protein Data Bank, NetSurfP

The new server builds on COCaDA, short for COntact search pruning by C-alpha Distance Analysis, a command-line tool the group previously developed for detecting interatomic contacts at large scale. The core insight behind COCaDA is elegantly simple: rather than exhaustively checking every possible pair of atoms in a protein — a computation that grows explosively with protein size — the algorithm uses the distances between C-alpha atoms, the backbone reference points of each amino acid residue, to prune the search space. If two residues’ C-alpha atoms are too far apart for any of their side-chain atoms to plausibly interact, the algorithm never bothers examining those atom pairs at all. This pruning strategy slashes the computational cost of contact detection dramatically, making it feasible to analyze entire structural databases rather than one protein at a time. COCaDA-web now wraps that engine in an interactive interface and adds a precomputed database of roughly 746 million contacts, divided into about 690 million contacts within single protein chains and 55 million contacts between different chains.

Speed is one of the platform’s headline claims. According to the authors, COCaDA-web takes less than a second to process proteins of up to 1,000 residues, a size range that covers the vast majority of entries in the Protein Data Bank. That kind of responsiveness transforms the workflow of exploratory analysis. Instead of submitting a job, walking away, and returning hours later to a static text file, researchers can adjust parameters, rerun the analysis, and watch the results update in real time. The results for each protein entry are presented in a dynamic table that can be filtered and interrogated, alongside an interactive three-dimensional visualization that lets users see exactly where the detected contacts sit in the folded structure. For those who prefer to work in professional molecular graphics software, the server also generates annotated PyMOL sessions, pre-packaged so that the contact analysis can be opened, inspected and repurposed directly in one of the field’s standard visualization environments.

Two features distinguish COCaDA-web from earlier contact-analysis efforts. The first is pH value customization for protonation-aware analysis. Protonation — whether a particular chemical group on an amino acid carries an extra proton or has lost one — changes with the acidity of the surrounding environment, and those changes alter which atoms can form hydrogen bonds and salt bridges. A histidine residue that is positively charged at pH 5 may be neutral at pH 7.5, and that difference can reshape an entire interaction network at an active site or a protein-protein interface. By letting users specify the pH of interest, the server can produce contact maps that reflect the chemistry of a specific biological condition rather than a single default state. The second novelty is support for biological assemblies. Many proteins only do their real work as multimers — symmetric complexes of several identical or complementary chains — and analyzing contacts in the full biological assembly, rather than in a single asymmetric unit extracted from a crystal, can reveal interfaces that would otherwise be missed entirely.

To demonstrate what the platform can do, the authors turned to a pair of proteins that make for a striking natural experiment: Transthyretin, known as TTR, and 5-Hydroxyisourate Hydrolase, or HIUase. The two proteins are evolutionarily and structurally related, sharing a common fold, yet they perform utterly different biological jobs. TTR is a transport protein found in blood and cerebrospinal fluid that carries thyroxine and retinol-binding protein — and, infamously, it is prone to misfolding and aggregation into amyloid fibrils, the hallmark of diseases including familial amyloid polyneuropathy and senile systemic amyloidosis. HIUase, by contrast, is an enzyme involved in purine metabolism in many organisms, catalyzing the hydrolysis of 5-hydroxyisourate as part of the pathway that degrades uric acid; loss of its function is associated with gout, a painful condition caused by urate crystal deposition. Two proteins with such similar architecture but such divergent — and clinically consequential — behaviors are exactly the kind of case where contact-level analysis can expose what the static fold alone cannot.

Running both proteins through COCaDA-web, the researchers found meaningful differences in the patterns of interatomic contacts, particularly in the central cavities of the two proteins. These cavities — the internal spaces that in TTR form the thyroxine-binding channel and in HIUase contribute to the enzyme’s active-site environment — showed distinct contact signatures that help explain how two proteins built on the same structural scaffold came to serve such different functions. The comparison illustrates a broader principle in structural biology: sequence and fold tell only part of the story. The fine-grained geography of atomic interactions, especially within cavities and at interfaces, encodes the functional specialization that evolution has layered onto a shared architectural blueprint. Tools that make those signatures easy to compare, side by side and residue by residue, give researchers a new lens on questions of functional divergence, protein engineering and drug targeting.

The scale of the underlying database deserves emphasis. Three-quarters of a billion contacts, precomputed and indexed, means that much of the drudgery of contact analysis has already been done before a user ever opens the website. A researcher investigating a protein family can pull up precomputed contact data for hundreds of related structures instantly, comparing interaction networks across species, mutants or ligand-bound states without running a single calculation. For proteins not yet in the database, users can submit their own queries and have the server compute contacts on demand, with results delivered in the same dynamic table and 3D visualization environment. The server also integrates with AlphaFold, the deep-learning system that has predicted structures for hundreds of millions of proteins, extending the reach of contact analysis beyond experimentally determined structures into the vast space of computationally predicted ones.

Accessibility is a deliberate design goal. COCaDA-web is freely available at bioinfo.dcc.ufmg.br/cocada-web with no login requirements, and the team has included step-by-step usage guides to lower the barrier for newcomers. That matters because contact analysis has historically been a gatekept skill: the traditional implementations, as the authors note, remain computationally expensive and pose significant scalability and usability challenges. A graduate student in a biology lab without a computing cluster, a clinician curious about the structural basis of an amyloid disease, or a teacher building a lesson on protein structure can all now interrogate the same data that specialists use. The work was funded by Brazilian agencies CAPES, FAPEMIG and CNPq, and the authors acknowledge advice from Carlos Henrique da Silveira, Lucas Bleicher and Miguel Ortega in developing the new features.

The broader significance of tools like COCaDA-web lies in how they reshape the rhythm of discovery in structural bioinformatics. As predicted structures from AlphaFold and experimental structures from cryo-electron microscopy flood the public databases, the bottleneck is shifting from obtaining structures to interpreting them. Contact networks are one of the most information-dense summaries of a structure — a fingerprint that captures chemistry, geometry and environment in a single representation — and making those fingerprints searchable, comparable and visualizable at web speed turns a specialist computation into a routine exploratory act. For diseases of protein misfolding, where the transition from a soluble transporter to a pathological aggregate hinges on subtle shifts in which atoms find which neighbors, that accessibility could accelerate the search for the molecular tipping points. And for the everyday work of characterizing enzymes, engineering binders and annotating newly sequenced genomes, a tool that answers contact questions in under a second may simply become part of how structural questions get asked.

Subject of Research: Interactive web server for exploratory analysis of interatomic contacts in protein structures

Article Title: COCaDA-web: an interactive web server for exploratory analysis of interatomic contacts in proteins

Article References: Lemos, R. P., Mariano, D., Bastos, A. L. A., Silveira, S. D. A., & de Melo-Minardi, R. C. (2026). COCaDA-web: an interactive web server for exploratory analysis of interatomic contacts in proteins. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06624-8

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06624-8

Keywords: COCaDA-web, protein contacts, structural bioinformatics, Protein Data Bank, web server, interatomic interactions, Transthyretin, amyloidosis, 5-Hydroxyisourate Hydrolase, protein structure, biological assemblies, protonation