The device is built on MspA, a mushroom-shaped porin drawn from the bacterium Mycobacterium smegmatis. MspA has long been a favorite scaffold in nanopore research because of its conical geometry and its narrow, well-defined constriction, which sits at the narrowest point of the pore and acts as the readout zone where passing molecules modulate an ionic current. What the team led by Shuo Huang of Nanjing University did was to graft a bespoke chemical adaptor into that constriction. The adaptor, called maleimido-C2-formylphenylboronic acid, or maleimido-C2-FPBA, combines two reactive functions in one compact molecule: a maleimide group that covalently anchors the adaptor to a single cysteine residue engineered into the pore wall, and an ortho-formylphenylboronic acid group that juts into the lumen of the pore where it can interact with passing analytes.

The chemistry of that boronic acid head group is what gives the pore its versatility. Boronic acids are famous for forming reversible covalent bonds, known as boronate esters, with cis-diols — the paired hydroxyl groups found abundantly on sugars — which explains why FPBA-equipped pores excel at recognizing saccharides. But the formyl group adjacent to the boron adds a second recognition mode: it can condense with the N-terminal amino groups of amino acids and peptides to form iminoboronate adducts, a reversible linkage that has become a powerful tool in chemical biology. This dual chemistry means that amino acids, sugars and even nucleotides, which carry phosphate and hydroxyl groups capable of engaging the boron center, each generate their own characteristic electrochemical fingerprints as they dock at and escape from the constriction.

When a voltage is applied across a lipid membrane containing a single MspA-FPBA pore, ions stream through the opening and produce a steady baseline current. Each time an analyte molecule enters the pore and binds transiently to the FPBA adaptor, the ionic current drops in a brief blockade whose depth, duration, noise profile and shape encode information about the molecule’s identity. The researchers systematically characterized these events for each of their target molecules. Notably, the earlier versions of related sensors had struggled: an MspA pore carrying only a phenylboronic acid adaptor failed to detect the amino acid phenylalanine, and a nickel-nitrilotriacetic acid modified pore registered no events for fructose or guanosine monophosphate. Only the combined formyl-boronate chemistry of FPBA proved capable of capturing this chemically diverse set of analytes with strong, well-resolved signals.

The catalog of molecules the engineered pore can discriminate is remarkable. On the amino acid side, it resolves all twenty standard proteinogenic building blocks plus selenocysteine, the twenty-first genetically encoded amino acid, and it further distinguishes three post-translationally modified residues: N6-acetyllysine, O-phosphotyrosine and asymmetric dimethylarginine, representing acetylation, phosphorylation and methylation, respectively. On the nucleotide side, it identifies the four canonical nucleoside monophosphates — AMP, UMP, CMP and GMP — as well as three epigenetically relevant variants: pseudouridine monophosphate, N6-methyladenosine monophosphate and inosine monophosphate. Four monosaccharides, including iduronic acid, L-arabinose, D-fructose and N-acetyl-D-glucosamine, and five short peptides round out the panel. Some analytes produce more than one characteristic event type — cysteine, histidine and lysine each yield two event classes, and arabinose produces three — yet the machine learning layer absorbs this complexity without difficulty.

That machine learning layer is central to the system’s performance. The researchers extracted nine numerical features from each nanopore event, including the mean current blockade, its standard deviation, a local noise ratio, the minimum and maximum current values, the event range, the dwell time, and the skewness and kurtosis of the blockade shape, forming a nine-dimensional fingerprint for every single-molecule encounter. They then benchmarked several classifiers using ten-fold cross-validation and found that an ensemble of bagged decision trees — an approach rooted in Leo Breiman’s classic random forests methodology — performed best, reaching the headline accuracy of 98.7 percent across the full panel of 24 amino acids, 7 nucleotides, 4 sugars and 5 peptides. The resulting confusion matrix shows cleanly separated classes even for chemically subtle distinctions, such as proline versus methionine, which the model resolves with validation accuracies of 98.8 and 99.4 percent respectively.

Perhaps the most consequential demonstration is the sensor’s ability to generalize. Because each class of analyte — amino acids, nucleotides, saccharides, peptides — produces predictable families of event signatures, the trained model can recognize the category of a molecule it has never seen in its training data, a capability the team calls few-shot learning. When the researchers filtered a yeast cell extract into the pore chamber, the classifier correctly assigned events to their proper analyte classes and further identified individual amino acids and nucleotides within those classes, despite the mixture being far messier than any of the purified standards used in training. This out-of-database classification matters enormously for real-world applications, where an unknown sample cannot be presumed to contain only molecules with pre-archived reference signatures.

To showcase practical analytical chemistry, the team applied MspA-FPBA to the compositional analysis of glycopeptides — peptides decorated with sugar chains, which are central to protein glycosylation, one of the most important and clinically consequential post-translational modifications in biology. They digested a glycosylated heptapeptide with leucine aminopeptidase, an enzyme that clips amino acids from the N-terminus until it stalls at the bulky glycosylation site. As the enzyme released glycine, methionine, glutamine and arginine one by one, the nanopore and its machine learning classifier identified each in real time, while the residual glycosylated tripeptide fragment generated its own distinctive events. In this way, a single measurement revealed both the amino acid composition preceding the sugar attachment site and the presence of the glycosylated remnant — information that would ordinarily require multiple orthogonal techniques.

The researchers are candid that this is a waypoint rather than a destination. In their abstract they note that future integration of a hydrolase enzyme with MspA-FPBA could turn the system into a self-consuming reader that degrades larger biomolecules into their building blocks and identifies them sequentially — an architecture that would echo the motor-protein ratcheting used in commercial nanopore DNA sequencing and that has already been explored conceptually for exopeptidase-driven protein sequencing. Challenges remain, including throughput, the concentration ranges required for reliable detection, and extending the panel beyond the current analyte set. Still, the convergence of a single multifunctional chemical adaptor, a rugged bacterial pore and modern machine learning delivers something the field has been missing: one sensor that speaks the chemical languages of proteomics, glycomics and transcriptomics at the single-molecule level. If the promised hydrolase coupling matures, the humble mycobacterial porin may find itself at the heart of a universal molecular reader, one blockade at a time.

Subject of Research: An engineered MspA nanopore with a formylphenylboronic acid adaptor for simultaneous single-molecule identification of amino acids, nucleotides, saccharides and peptides

Article Title: An engineered nanopore identifies saccharides, amino acids, peptides and ribonucleotides

Article References: Yao, L., Wang, Z., Chen, J., Sun, W., Wang, K., Xiao, Y., Zhang, H., Li, W., Wang, Y., Zhao, L., Dai, X., Qian, L., Zhang, P., & Huang, S. (2026). An engineered nanopore identifies saccharides, amino acids, peptides and ribonucleotides. Nature Biotechnology. https://doi.org/10.1038/s41587-026-03308-9

Image Credits: AI Generated

DOI: 10.1038/s41587-026-03308-9

Keywords: nanopore sensing, MspA, formylphenylboronic acid, amino acid identification, nucleoside monophosphates, saccharide detection, peptide sensing, machine learning, single-molecule analysis, glycopeptide, post-translational modifications, epigenetic modifications

Proteins are not static entities. After they are translated from messenger RNA, they undergo a dizzying array of chemical modifications: phosphate groups are added and removed, the protein backbone is clipped by proteases, small protein tags such as ubiquitin are attached, and amino acids can be chemically altered in dozens of other ways. Each unique combination of modifications and sequence variants constitutes a distinct proteoform. The trouble is that two proteoforms of the same protein can behave in radically different ways inside a cell, one folding into a harmless shape and another seeding the toxic aggregates that kill neurons. Standard proteomics methods, which typically chop proteins into small peptides before identifying them, lose the connectivity information that reveals which modifications coexisted on the same original molecule. As a result, the proteoform landscape of even a well-studied protein like tau has remained only partially charted.

Iterative Mapping of proteoforms tackles this problem by interrogating individual protein molecules directly, preserving the integrity of each proteoform throughout the measurement. The core idea is to perform repeated cycles of imaging-based readout on single immobilized molecules, building up a pattern of signals that serves as a molecular fingerprint. Because each molecule is observed on its own, the resulting data reflect genuine single-molecule heterogeneity rather than population averages. This matters enormously for tau, where rare proteoforms may be the biologically decisive species. A modification present on only one percent of tau molecules could be invisible to bulk measurements, yet a small pool of aberrantly modified molecules might be sufficient to nucleate the pathological aggregates that spread through the brain in Alzheimer’s disease.

The scale of the new approach is what sets it apart. Earlier single-molecule protein characterization methods, while conceptually elegant, were limited in throughput, making it impractical to survey the full diversity of proteoforms in complex biological samples. Iterative Mapping achieves large-scale measurement by combining highly parallel detection with an iterative readout strategy, allowing millions of individual molecules to be characterized in a single experiment. The researchers validated the technique using control samples of known composition, a critical step that established the method’s accuracy in quantifying predefined proteoform groups. Only after demonstrating that the technique could correctly identify and count proteoforms in mixtures of known makeup did the team apply it to more complex and clinically relevant material.

Tau is an unusually challenging target for such an analysis. In the human brain, the MAPT gene produces six major isoforms of tau through alternative splicing, differing in the number of microtubule-binding repeats and N-terminal inserts. On top of this isoform diversity, tau carries an enormous number of possible phosphorylation sites, with dozens of serine, threonine, and tyrosine residues that can be modified individually or in combination. The phosphorylation state of tau governs its normal function in stabilizing microtubules, the structural scaffolds of neurons, but hyperphosphorylation promotes tau’s detachment from microtubules, its misfolding, and ultimately its aggregation into the paired helical filaments that compose neurofibrillary tangles. Because the biological consequences of phosphorylation depend on which sites are modified together on the same molecule, knowing the total amount of tau phosphorylation in a sample is far less informative than knowing the actual distribution of proteoforms.

The demonstration in model systems used in tauopathy research provides a bridge between controlled validation experiments and human tissue. Cell and animal models of tauopathy are workhorses of the field, used to test hypotheses about how tau becomes pathological and to screen candidate therapies. Applying Iterative Mapping to these systems allows researchers to characterize how the tau proteoform landscape shifts as disease-like states develop, and to compare the proteoform signatures of different models against one another. Such comparisons could help resolve a persistent problem in the field: different model systems recapitulate different aspects of tau pathology, and it has been difficult to know which models most faithfully reflect the human disease. A quantitative, single-molecule proteoform census offers a new common currency for making those comparisons.

The most striking application, however, is the analysis of human-derived brain tissue samples. Post-mortem brain tissue from individuals with Alzheimer’s disease and related tauopathies is a precious and technically difficult resource, often available in limited quantities and frequently affected by post-mortem delays and variable tissue quality. Demonstrating that Iterative Mapping can extract meaningful proteoform quantification from such material establishes the method’s readiness for real-world translational research. The ability to profile tau proteoform groups directly in human brain tissue means that hypotheses generated in models can now be tested against the actual molecular substrate of disease, and that proteoform patterns associated with specific diagnoses, disease stages, or clinical outcomes can be systematically searched for.

The implications for drug development could be substantial. A growing number of therapeutic strategies target tau directly, including antisense oligonucleotides designed to reduce tau production, immunotherapies intended to clear pathological tau species, and small molecules aimed at inhibiting the kinases that phosphorylate tau. Each of these approaches would benefit from a measurement technology that can report precisely which proteoforms are reduced or altered following treatment. Bulk phosphorylation assays can indicate that total tau phosphorylation has decreased, but they cannot reveal whether the specific proteoform groups thought to drive toxicity have been affected. Single-molecule proteoform quantification provides exactly that granularity, potentially enabling biomarker-guided clinical trials in which molecular responses are monitored at the level of individual protein species.

Beyond tau, the demonstration establishes a general template for large-scale single-molecule proteoform analysis that could be extended to other proteins of biomedical importance. Alpha-synuclein in Parkinson’s disease, huntingtin in Huntington’s disease, TDP-43 in amyotrophic lateral sclerosis, and amyloid precursor protein in Alzheimer’s disease all share the same basic challenge: their pathological behavior depends on proteoform-level details that bulk methods obscure. If Iterative Mapping can be adapted to these targets, the technology could catalyze a broader shift in proteomics toward intact-protein, single-molecule measurement, complementing the peptide-centric workflows that have dominated the field for decades. The convergence of single-molecule imaging, iterative biochemical readout, and computational analysis reflected in this work suggests that the long-sought goal of routinely reading complete proteoforms is moving from aspiration toward practice.

Challenges remain before such methods become routine in laboratories and clinics. Sample preparation for single-molecule analysis must preserve labile modifications, the computational pipelines for interpreting iterative readout patterns must be robust across diverse sample types, and the proteoform groups quantified today represent a subset of the full molecular diversity that likely exists in brain tissue. Nevertheless, the demonstration that large-scale, single-molecule proteoform measurement is achievable, validated against known controls, and applicable to human tissue marks a genuine advance. For a protein like tau, whose transformation from a neuronal workhorse into a killer aggregate has puzzled researchers for decades, the ability to count and classify its molecular forms one molecule at a time may finally provide the resolution needed to understand, and ultimately interrupt, the progression of tauopathy.

Subject of Research: Large-scale single-molecule measurement of intact tau proteoforms using Iterative Mapping

Article Title: Large-scale single-molecule analysis of tau proteoforms

Article References: Joly, J., Budamagunta, V., Zhang, Z., Nortman, B., Jouzi, M., Bhatnagar, R., Egertson, J. D., Flaster, M. E., Grothe, R., Guha, S., Kaneshige, K., McVey, K., Nelson, N., Perera, R. T., Tan, S. J., Trinh, T., Arnott, D., Lipka, J., Pandya, N. J., … Mallick, P. (2026). Large-scale single-molecule analysis of tau proteoforms. Nature Methods, 23(9), 1786-1797. https://doi.org/10.1038/s41592-026-03188-6

Image Credits: AI Generated

DOI: 10.1038/s41592-026-03188-6

Keywords: tau, proteoforms, single-molecule analysis, Iterative Mapping, tauopathy, Alzheimer's disease, phosphorylation, proteomics, neurodegeneration, brain tissue, biomarkers, drug development