Proteoform medicine: characterizing and targeting protein forms in human disease

Abstract

Proteoforms are the diverse molecular protein species produced from a single gene through genetic variation, alternative splicing and post-translational modifications. They are the crucial link between genotype and phenotype. There are estimated to be more than one million distinct protein variants produced from ~20,000 protein-coding genes in a given cell, making these proteoforms a vast and largely uncharacterized dimension in biomedical research. This Review focuses on the role of proteoforms in human genetic diseases. We highlight cutting-edge technologies for the identification and characterization of proteoforms, including long-read transcriptomics and emerging methods for direct protein sequencing, and we present a network biology framework to explain how proteoforms can perturb the molecular interactions and cellular pathways underlying disease phenotypes. We believe that precision medicine will require precision proteomics. An increasing knowledge of proteoform biology from molecular, systems and clinical perspectives will guide future research, ultimately contributing to a more precise understanding of the molecular basis of disease and refined therapeutic interventions.

Introduction

The concepts of ‘gene’ and ‘protein’ have evolved markedly since their original use. Early geneticists, beginning with Gregor Mendel in the 1800s, described invisible hereditary ‘factors’, and it was only in the mid-twentieth century that the molecular definition of genes was clarified through the discovery of DNA structure and the genetic code. Similarly, proteins were once viewed as amorphous colloids until the sequencing of insulin by Frederick Sanger in the 1950s revealed it to be a precisely ordered macromolecule¹. The ‘one gene–one protein’ model, proposed by George Beadle and Edward Tatum in the 1940s, initially seemed to be validated, but the experimental data began to tell a different story. Rather than single bands, smears or multiple spots for a given protein emerged on protein electrophoresis gels, which indicated the presence of many related protein species². The discovery of alternative splicing by Phillip Sharp and Richard Roberts in the 1970s, together with the advent of high-throughput RNA-sequencing technologies, exposed the combinatorial diversity of mRNA processing³. Studies in the early to mid-twentieth century — using chemical, electrophoretic and radiographic methods — revealed the existence of post-translational modifications (PTMs) on proteins^4–8. However, the tools available at the time were insufficient to map these modifications comprehensively or connect them to protein diversity. The advent of mass spectrometry-based proteomics in the 1990s enabled large-scale detection of PTMs and variant peptides, confirming that a single gene can give rise to a multitude of distinct molecular protein species.

Lloyd Smith and Neil Kelleher⁹ first proposed the term ‘proteoform’ in 2013 to denote a molecular form of a protein arising from genetic variation, alternative RNA processing or PTMs. The term proteoform has both a strict chemical definition and broader conceptual significance. Chemically, a proteoform is defined as the exact N-terminal to C-terminal amino acid sequence, as well as any covalent side-chain modifications. Conservative estimates suggest that the ~20,300 protein-coding genes in the human genome give rise to more than one million distinct proteoforms in a human cellular proteome, which is equivalent to roughly fifty protein variants per gene^10–12. The term also conveys a conceptual shift, reflecting that protein-level variation is extensive and both functionally and clinically relevant¹³. Disease-associated proteoforms are precise biomarkers and therapeutic targets across cancers¹⁴, neurodegenerative disorders¹⁵ and cardiovascular disease¹⁶.

This proteoform diversity poses crucial questions relating to how we should conceptualize the proteoform landscape and how individual protein variants can alter molecular interactions, cellular networks and clinical outcomes. Here, we provide a framework to address these questions by focusing on the proteoform as the fundamental unit of pathophysiology and by describing how the unique chemistry of protein variations can alter molecular pathways and drive disease phenotypes. We begin by outlining the molecular sources of proteoform diversity, illustrating how combinatorial mechanisms — such as genetic variation, alternative splicing and PTMs — can yield a large proteoform repertoire from a limited number of genes. We then introduce a systematic ‘map–perturb–predict’ framework¹⁷ for characterizing proteoforms and predicting their functional consequences. We also evaluate existing experimental and computational tools in the field relevant to these objectives. Next, we examine how to contextualize proteoforms within molecular networks, modelling them as discrete, disease-relevant entities. Finally, we discuss emerging strategies for proteoform-specific therapeutic targeting and outline a path towards clinical translation.

Sources of proteoform diversity

Tracing the molecular ontogeny of proteoforms can enable a deeper understanding of the multilayered regulatory logic that underlies proteoform diversity. The output of each gene is modified across four major regulatory layers: genetic variation of the DNA sequence, transcriptomic variation in the form of RNA processing, translational variation and post-translational variation in the form of PTMs (Fig. 1). Beyond single-gene products, proteoform diversity can also emerge from multi-gene protein families whose paralogues generate related yet functionally distinct proteoforms. These layers of variation can act both independently and in combination¹⁸ (Fig. 2 and Box 1), thereby expanding the products of a single genomic locus into a rich landscape of proteoforms.

Fig. 1 |. Overview of proteoform diversity, molecular ontogeny and functional examples of reference and alternative proteoforms from five well-characterized examples.

A, Genetic variation encompasses changes in the DNA sequence that affect the final proteoform. An example of this is the KRAS G12D genetic mutation. Ba, Transcriptomic variation includes alternative promoter use, as exemplified by brain and retinal variants of dystrophin (DMD). Bb, Transcriptomic variation also includes alternative splicing, as shown for the alternative inclusion of exon 18b in FOXP1-ES. C, Translational variation, for example using different translation start sites or an alternative reading frame, can synthesize alternative proteins from the same message. Alternative reading frame use at the CDKN2A locus encodes two unrelated proteins — p16INK4a and p14ARF. D, Post-translational variation encompasses chemical transformations such as phosphorylation (P), methylation (Me) and ubiquitylation (Ub). Phosphorylation of Ser32 and Ser36 residues in IκBα triggers its degradation and release of NF-κB to activate the transcription of inflammatory genes. Structures predicted using AlphaFold 3; AlphaFold Data Copyright (2022) DeepMind Technologies Limited.

Fig. 2 |. Combinatorial complexity of protein variation.

a–f, Examples to illustrate how protein function can be shaped by interactions between variations at different regulatory layers — genetic variation (mutations), transcriptomic variation (splicing changes), translational variation and post-translational variation (post-translational modifications such as phosphorylation (P), methylation (Me), trimethylation (Me3), ubiquitylation (Ub) and acetylation (Ac)). For further details, see Box 1.

Box 1 |. Combinatorial complexity of proteoform variation events.

Combinatorics refers to the idea that multiple sources of variation can coexist within a single proteoform, producing regulatory behaviours that are more than the sum of their parts. Rather than studying genetic variation (mutations), transcriptomic variation (splicing changes), translational variation and post-translational variation (post-translational modifications (PTMs)) in isolation, combinatorics emphasizes the interplay between these molecular events (Fig. 2). Even within a single regulatory layer, such as multiple PTMs on the same protein or multiple mutations in the same gene, these interactions can shape protein function in specific or nonspecific ways. In addition, combinatorial outcomes can be functionally diverse, being additive (each variation acts independently), synergistic (variations enhance each other’s effects), antagonistic (variations suppress each other’s effects), restorative (one variation compensates for a loss of function caused by another variation), ablative (variations disrupt a feature of the protein) or entirely novel (variations introduce a new interaction or function of the protein).

Large-scale integration of acetylation, ubiquitylation and phosphorylation datasets has revealed extensive coordination of PTMs across protein complexes, suggesting that combinatorial modification is a fundamental regulatory mechanism in the human proteome²³⁸. Histone tails offer an example of PTMs that function in combination: acetylation at Lys18 and Lys23 can act synergistically, promoting transcriptionally active chromatin, whereas these same marks may antagonize repressive Lys9 trimethylation by interfering with heterochromatin formation¹⁹⁵ (Fig. 2a). In this example, specific PTMs reinforce or inhibit each other’s effects based on their positional context.

Combinatorics can also operate in a nonspecific manner, in which the combined biochemical properties of modifications — such as overall charge or steric bulk — alter proteoform behaviour. For example, the S305N mutation in MAPT (also known as tau) affects the final nucleotide of exon 10, destabilizing a regulatory stem loop to alter the splicing pattern and increase production of exon 10-containing 4R tau (tau protein isoform containing four repeat domains)²³⁹ (Fig. 2b). The altered splicing pattern also promotes hyperphosphorylation of tau, which is a hallmark of tau pathology. Here, the pathogenic effect of hyperphosphorylation likely arises from cumulative changes to the conformation or charge of the proteoform, rather than specific phosphorylation sites. This highlights the fact that combinatorial regulation of proteoforms can also operate through generalized tendencies rather than precise molecular interactions. Other examples of this include recombination-driven B cell receptor diversification in the mammalian immune system²⁴⁰, splice-mediated variation of Dscam in Drosophila for self-tolerance²⁴¹ and broad glycosylation patterns that make up the protective glycocalyx of cells²⁴².

The interplay between different variations in the same protein can also span different regulatory levels, extending across mutations, splicing and PTMs. For example, in colorectal cancer cell lines, the oncogenic KRAS(G13D) mutation restricts S-nitrosylation of the KRAS4b isoform at Cys118, a modification that is only observed in the wild-type protein, illustrating interplay between genetic variation and accessibility to PTMs²⁴³ (Fig. 2c). In tauopathies, the G303V mutation in MAPT promotes inclusion of exon 10, shifting isoform expression towards the 4R tau variant (Fig. 2d). This change in splicing of MAPT exposes specific lysine residues, such as Lys343 and Lys369, that are absent in the 3R isoform and that undergo ubiquitylation, which demonstrates that mutation-driven splicing can create opportunities for PTMs^244,245. A missense mutation in CFTR that impairs folding is partially rescued by a nearby silent single-nucleotide polymorphism (T2562G) that slows translation and provides additional time for the correct kinetic folding of the mutant protein (Fig. 2e). The example of CFTR is a case of intragenic epistasis, in which the effect of one variant is dependent on the presence of another within the same gene²⁴⁶. Variations at two different regulatory levels — amino acid sequence and translation speed — combine to restore protein function.

Whereas these examples highlight individual cases of the interplay between regulatory levels and intragenic epistasis, systematic screening approaches can reveal such dependencies at scale. In Escherichia coli, large-scale, deep mutational scanning of the enzyme β-lactamase (CXT-M) was used to assess all pairwise substitutions across 17 residues, identifying a double mutation that alters antibiotic resistance²⁴⁷ (Fig. 2f). This screen revealed evidence of widespread positive and negative intragenic epistasis, even indicating a new catalytic mechanism that emerged only when both E166Y and N170G mutations were present, with neither mutation conferring activity on its own. Although this work was carried out in a prokaryote model organism, it illustrates how large-scale combinatorial testing can uncover functional dependencies between variations, an approach that could be applied in human systems to map proteoform-specific regulatory logic.

Genetic variation

Proteoform diversity begins with genetic variation (Fig. 1A). Each human genome contains ~10,000 missense single-nucleotide polymorphisms (SNPs) that alter single amino acids, ~200 SNPs that affect start or stop codons, and ~600 small coding indels, both in-frame and frameshifting¹⁹. In addition, each individual genome has 8–10 structural variants in coding regions and occasional chromosomal rearrangements that can generate fusion proteins²⁰. Somatic mutations further diversify the proteome over time, introducing thousands of private variants in aging tissues and up to 100 substitutions per megabase in cancer cells²¹. These DNA-encoded changes alter linear amino acid sequences — from point mutations to truncations to chimeric fusions — and thus produce distinct proteoforms. An example is the KRAS(G12D) mutation, a single amino acid substitution that stabilizes protein conformation and enhances binding of KRAS to effector proteins to rewire downstream signalling and promote cell proliferation and survival²².

Transcriptomic variation

Downstream of genetic variants, differences in RNA processing — in the form of alternative promoter use, alternative splicing, alternative polyadenylation or alternative last exon use — introduce extensive variation that affects the final transcripts that are synthesized and, thus, the proteins that are encoded.

The use of alternative promoters by RNA polymerase produces distinct transcription start sites for ~75% of genes²³, which encode variable N-terminal sequences that are frequently tissue-specific²⁴ (Fig. 1Ba). An example of promoter-driven proteoform variation is provided by the dystrophin (DMD) gene²⁵; the full-length DMD protein maintains cytoskeletal integrity, whereas the retina-specific Dp260 isoform of DMD has distinct structural and signalling roles.

Alternative splicing is one of the most pervasive mechanisms for generating molecular diversity of the proteome, with more than 95% of genes producing multiple mRNA splice isoforms²⁶ (Fig. 1Bb). Skipped or mutually exclusive exons, alternative 5′ or 3′ splice sites and cryptic intronic cassettes together generate tens of thousands of open reading frame (ORF) variants²⁷. For example, in the case of transcription factor FOXP1, which regulates stem cell pluripotency and differentiation, inclusion of exon 18b maintains pluripotency, whereas exclusion of exon 18b creates a FOXP1 isoform with altered DNA-binding specificity to activate transcriptional programs of differentiation²⁸. Alternative polyadenylation and alternative last exon use can change the C-terminal coding region. Alternative polyadenylation can lead to early termination of the RNA sequence, producing a truncated protein, whereas alternative last exons produce different C-terminal end sequences²⁹.

RNA processing is an important source of proteoform diversity; the combination of mechanisms mediating transcriptomic variation is predicted to produce ~500,000 distinct protein-coding primary sequences in a human cell³⁰.

Translational variation

Even for a fixed mRNA transcript, ribosomes can diversify their output by deviating from the rules of canonical translation — in other words, ‘first AUG to in-frame stop codon’ — to synthesize alternative proteins (Fig. 1C). For example, leaky scanning or initiation at non-AUG codons creates start-site flexibility and thus variable N termini. Internal ribosome entry sites³¹ can initiate translation mid-transcript. And short ORFs localized upstream of, downstream of or even overlapping with the canonical ORF can encode micropeptides³². Ribosomes can also exhibit unusual behaviours such as translational frameshifting and codon recoding events^33,34.

An example of proteoform diversity driven by translational variation can be found at the CDKN2A locus, which encodes two unrelated proteins — p16INK4a and p14ARF — from overlapping sequences translated in different reading frames³⁵. Whereas p16INK4a inhibits cyclin-dependent kinases, p14ARF activates the p53 pathway, acting through distinct interaction partners.

Post-translational variation

After translation, newly synthesized proteins can undergo extensive chemical transformations through PTMs (Fig. 1D). More than 400 distinct PTMs have been identified, including phosphorylation, acetylation, glycosylation and ubiquitylation, as well as proteolytic cleavage, disulfide bond formation and N-terminal methionine excision³⁶. A well-characterized example is that of nuclear factor-κB (NF-κB) inhibitor-α (IκBα; encoded by NFKBIA). In response to cytokines, IκBα is phosphorylated at Ser32 and Ser36, triggering its degradation and releasing NF-κB to activate the transcription of inflammatory genes, which illustrates how site-specific PTMs can drive regulatory outcomes³⁷.

Variation in multi-gene protein families

Although we generally define proteoform groups as arising from a single gene product, proteoform diversity can also arise from multi-gene protein families in which closely related paralogues are expressed. These sequence variants, which are conserved within the population, can alter the structural properties and interaction interfaces of proteins, as well as their susceptibility to specific PTMs. For example, histone H3 variants H3.1, H3.2 and H3.3 differ by only a few residues yet have distinct genomic localizations, cell-cycle expression patterns and PTMs^38,39. Similarly, β-actin and γ-actin isoforms vary at just four positions but have unique intracellular distributions and interaction networks⁴⁰, and tubulin isoforms have small sequence changes that affect binding to microtubule-associated proteins and modifying enzymes⁴¹. Such paralogue diversity allows for distinct proteoform repertoires, enabling differential regulation and phenotypic outcomes despite high levels of sequence identity.

Combinatorial complexity

Crucially, the combination of genetic variation, alternative RNA processing, translational recoding and diverse PTMs results in an extraordinary combinatorial expansion of the proteome (Fig. 2 and Box 1). Importantly, this diversity is not simply additive — it reflects the interplay between multiple regulatory layers. Each layer — from DNA sequence to RNA processing to protein modification — functions as a checkpoint that has evolved to help the cell tailor functional molecules to meet contextual demands. In this sense, proteoforms embody the cell’s dynamic toolkit of fine-tuned molecular products that have been refined over very long periods of time to support homeostasis. This also means that proteoforms are not static entities. They are dynamically regulated in response to developmental stage, cell type or environmental stimuli, and may differ in subcellular localization, half-life or interaction potential. Without the mapping of these variations across regulatory layers, we risk missing the molecular players that are most relevant to physiology and disease. Cataloguing and understanding this proteoform diversity remains a central challenge for modern biological research, as illustrated by large-scale initiatives such as the Human Proteoform Project, which seeks to systematically characterize and annotate human proteoforms across biological contexts¹⁸.

Map–perturb–predict

Despite the recognized differences in functions and effects of individual proteoforms, most studies have traditionally focused on canonical, abundant or previously well-characterized proteoforms⁴². These represent only a tiny proportion of the estimated millions of proteoforms that are expressed across cells, tissues, development and disease conditions. As such, system-scale approaches are needed to identify the most phenotypically relevant proteoforms in any given context. In this section, we describe an iterative ‘map–perturb–predict’ framework to systematically characterize proteoform function¹⁷ (Fig. 3). Mapping involves high-resolution bioanalytical measurements of proteoforms. These proteoforms are then perturbed and assayed empirically to assess their impact on phenotype. Finally, the properties of untested proteoforms are predicted from experimental data, which can be used to prioritize targets for subsequent rounds of experimentation as well as to extract generalizable rules of proteoform biology. A summary of the experimental and computational tools available for the mapping, perturbation and prediction stages of this cycle can be found in Supplementary Tables 1–3.

Fig. 3 |. The map–perturb–predict conceptual framework for proteoform medicine.

a, Map. Systematic detection of proteoforms across physiological and disease contexts using transcriptomic and proteomic technologies to build a comprehensive proteoform atlas. b, Perturb. Functional characterization of proteoforms by experimental manipulation strategies (such as gene editing, transfection, RNA modulation and transduction) and comparative assays assessing, for example, molecular interactions, transcriptional regulation, cellular phenotypes and in vivo outcomes in animal models. c, Predict. Computational prediction of proteoform identity and function based on sequence and/or expression context and prior experimental data, enabling the modelling of proteoform behaviour and the identification of pathogenic variants. Ac, acetylation; Me, methylation; P, phosphorylation; SUMO, SUMOylation; Ub, ubiquitylation.

Proteoform mapping

Deep mapping of the human proteome¹⁸ (Fig. 3a) will require an integrated toolkit spanning multiple technological modalities (Supplementary Table 1). No breakthrough method currently captures all facets of the proteome; rather, each method provides a different view with its own strengths and weaknesses.

Sequence-based prediction.

Proteoform diversity can be estimated from sequencing-based approaches at genomic (DNA), transcriptomic (mRNA) or translatomic (ribosome) levels. Genomic sequencing provides the foundation for predicting proteoform diversity by cataloguing protein-altering coding variation (for example, missense SNPs and indels). High-throughput next-generation sequencing, including whole-genome sequencing and whole-exome sequencing⁴³, is the most commonly used method for variant detection at the genomic level. Over the past decade, long-read platforms that enable telomere-to-telomere assemblies (such as the Revio system from PacBio⁴⁴ and PromethION from Oxford Nanopore Technologies⁴⁵) have been increasingly adopted, allowing for complex structural variants to be resolved across diverse populations. Variant annotations are catalogued in gene annotation databases (such as UniProt⁴⁶) and variant databases (such as dbSNP⁴⁷ and COSMIC (Catalogue of Somatic Mutations in Cancer)⁴⁸), and can be used as inputs for proteogenomic workflows that translate genomic variation into predicted proteoform sequences⁴⁹.

Transcriptome sequencing reveals which mRNAs are produced in a cell or tissue and, by extension, which ORFs are available for translation. The advent of short-read RNA-sequencing technologies (such as Illumina⁵⁰) brought unprecedented sequencing depth of millions to billions of reads per sample⁵¹. This revolutionized the field of transcriptomics, revealing that ~95% of multi-exon genes undergo alternative splicing²⁶ and exposing widespread 3′-end variation and RNA editing, and thousands of previously unannotated exons and isoforms. However deep, short reads (of 50–250 nucleotides) paint a fragmentary picture of single variation ‘events’ — such as exons, splice junctions and editing sites — without resolving their co-occurrence on full-length transcripts. Hence, isoform reconstruction from short-read sequencing data is probabilistic and cannot determine combinatorial splice patterns⁵².

Third-generation long-read sequencing addresses this issue by enabling contiguous, full-length inferences of protein isoform sequences. Platforms such as PacBio HiFi⁴⁴ and Oxford Nanopore Technologies (ONT)⁴⁵ generate full-length cDNA or native RNA reads (dRNA-seq⁵³) that resolve exon connectivity and RNA modifications⁵⁴. Efforts to build tissue atlases using these technologies have uncovered tens of thousands of previously unseen coding transcripts⁵⁵. For high-accuracy platforms such as PacBio, reads can be translated into ORFs to predict protein isoforms within individual samples, which can then be validated using mass spectrometry^56,57.

However, not all mRNA isoforms are translated. Ribosome profiling (Ribo-seq) provides empirical evidence of translation by sequencing ribosome-protected mRNA fragments⁵⁸. Ribo-seq analyses delineate ORF boundaries and have enabled diverse translational events to be observed, including the use of upstream ORFs and non-AUG start codons (for example, as catalogued in TranslatomeDB⁵⁹), internal initiation events, and recoding events such as ribosomal frameshifting. Such events are missed by approaches that assume canonical ‘AUG-to-stop codon’ translation^58,60. Recent extensions of Ribo-Seq, such as Frac-seq (capturing both cytoplasmic and translated mRNAs)⁶¹ and Ribo-STAMP⁶², have measured isoform-resolved translation, supporting the existence of alternative protein isoforms.

Direct detection.

Only direct protein measurements can validate which proteoforms are stably expressed in vivo. Among the landscape of protein detection tools — including affinity-based measurements, mass spectrometry approaches (bottom-up and top-down) and next-generation protein sequencing — each method provides different insights into proteoform-level features. No single platform can fully capture all proteoform diversity, with each method having a different balance of resolution versus throughput. Dissecting the combinatorial complexity of multiple concurrent modifications on a single proteoform remains challenging owing to coverage limitations. To address these challenges⁶³, integrated approaches are increasingly being used to leverage the strengths of complementary technologies. For example, proteogenomics combines RNA sequencing (RNA-seq) with proteomic evidence to validate predicted protein variants^64–66; long-read RNA-seq provides full-length isoform models that increase the confidence of protein isoform predictions⁶⁷.

Affinity-based methods are widely used for protein detection. They range from western blots and enzyme-linked immunosorbent assays (ELISAs) to more recently introduced, high-throughput platforms such as proximity extension assays (Olink⁶⁸) and aptamer-based assays (SOMAscan⁶⁹). However, most affinity reagents bind to protein regions that are shared across variants or can be altered by sequence changes or PTMs, which limits their ability to reliably distinguish similar proteoforms. Proteoform-specific detection is possible using affinity-based methods, but it requires prior knowledge of the target and the availability of high-quality reagents against unique epitopes.

Bottom-up mass spectrometry proteomics (BUP)⁷⁰, also known as ‘shotgun’ mass spectrometry, remains the gold standard for high-throughput protein identification and quantification at the peptide level, enabling the generation of comprehensive proteome maps of human cell lines⁷¹, healthy tissues^72–74 and cancers^75,76. In a typical workflow, proteins are enzymatically digested, the resulting peptides are analysed by liquid chromatography–tandem mass spectrometry (LC–MS/MS), and peptides are identified from LC–MS/MS spectra through searching sequence databases. State-of-the-art mass spectrometry instruments, such as Orbitrap from Thermo Fisher Scientific and timsTOF instruments from Bruker, combined with ultra-fast acquisition modes, now routinely identify more than 200,000 peptides and more than 12,000 protein groups in a single LC–MS/MS run⁷⁷. A recent six-enzyme survey of Tier-1 ENCODE human cell lines (GM12878, K562 and H1) reported the presence of ~180,000 peptides, covering 80–90% of the canonical proteome⁷¹.

BUP excels at cataloguing the landscape of individual variation ‘events’, such as genetic variants and variants in splice boundaries or proteolytic cleavage sites, across large cohorts and experimental conditions. Specialized workflows further enable the site-specific detection of diverse PTMs, including phosphorylation, proteolytic cleavage events, glycosylation, N-myristoylation and S-acylation⁷⁸. However, in an analogous manner to short-read RNA sequencing, BUP is intrinsically fragmentary. The proteolytic digestion step disconnects peptides from their source proteoform, abolishing the proteoform context⁷⁹. As a result, peptide-level data cannot specify which combinations of variants co-occur on the same protein molecule (with the exception of rare cases in which multiple variation events reside within the same peptide). Furthermore, certain sequences or events that occur in specific proteoforms may not be amenable to detection by mass spectrometry owing to their inherent physiochemical properties. To bypass some of these limitations, BUP technologies can attempt to infer ‘proteoform groups’ by analysing peptide covariation patterns across multiple biological conditions. Although these approaches do not resolve complete sequence or PTM composition, they can reveal biologically meaningful distinctions at high throughput and deep proteomic coverage. Recent studies have used this strategy to investigate protein complex assembly⁸⁰, thermal stability⁸¹ and expression across tissues⁸² of proteoform groups, highlighting its potential as a complementary means to gain insight into intact proteoforms when direct resolution is not feasible.

Top-down mass spectrometry proteomics⁸³ (TDP) is one of the few technologies that directly measures intact proteoforms. Unlike epitope-centric or peptide-centric workflows that must infer proteins from polypeptide fragments, TDP analyses intact proteins, thus preserving the native combination of sequence variants, PTMs and proteolytic events on a single molecule⁸⁴. As it is the case for BUP, TDP typically uses LC–MS/MS. In TDP, however, intact proteoforms rather than peptides are introduced into the mass spectrometer and fragmented using electron-based, collision-based or photon-based methods (such as electron capture dissociation, electron transfer dissociation or ultraviolet photodissociation), generating MS/MS spectra of proteoform fragment ions that retain labile modifications and support high-resolution backbone mapping⁸³. Each spectrum theoretically corresponds to a distinct proteoform, enabling protein sequence and modifications to be unambiguously assigned in a single experiment. This capability led to Smith and Kelleher first proposing the concept of proteoforms. Most TDP analyses, unless otherwise specified, are carried out under denaturing conditions to maximize sequence coverage of individual proteoforms. By contrast, native TDP (also known as native MS) preserves higher-order assemblies to investigate proteoform stoichiometry and function within protein complexes⁸⁵.

Denaturing TDP has been used to characterize human proteoforms in cell lines⁸⁶, tissues^87,88, biofluids^89,90 and cancers^91,92. The number of annotated proteoforms in public databases, such as those curated by the Consortium for Top-Down Proteomics (CTDP), continues to grow¹⁸. TDP requires optimized sample preparation to preserve intact proteins and reduce complexity, and its throughput is somewhat lower than that of BUP^93,94. When comparing denaturing TDP and BUP under matched, single-injection cell lysate conditions, current state-of-the-art denaturing TDP identifies ~400 proteins and ~800 to several thousand proteoforms, with coverage largely limited to protein species weighing less than ~30–45 kDa⁹³; by contrast, state-of-the-art untargeted BUP routinely detects 7,000–10,000 protein groups, but it yields no fully resolved proteoforms owing to loss of sequence context during digestion⁹⁵. Indeed, deep proteome coverage using TDP still requires extensive pre-fractionation and sophisticated algorithms to deconvolve complex MS/MS spectra. In addition, proteins with a molecular mass greater than ~30 kDa remain challenging to analyse with high sequence coverage owing to retained higher-order protein structure, lower fragmentation efficiency and increased spectral complexity⁹⁶. Despite this limitation, TDP pipelines and datasets are becoming more standardized and accessible across the community. Raw data can be deconvolved and scored for proteoforms using software packages such as ProSightPD⁹⁷ and MetaMorpheus-TC⁹⁸. All implement the CTDP five-level confidence scheme⁹⁹, ensuring consistent reporting of data. Guidelines and benchmarking efforts for intact protein analysis are now available to support rigorous experimental design in basic research settings¹⁰⁰, and recent efforts have translated these workflows for clinical sample preparation to enable reliable detection of proteoform biomarkers¹⁰¹. TDP spectra and proteoform identifications feed directly into public archives (such as TDPortal and MassIVE), and the Human Proteoform Project¹⁸ is driving community-wide adoption of standards and cross-laboratory integration. TDP-based workflows are also being integrated with BUP to increase confidence in proteoform identification¹⁰². Additionally, recent advances now extend top-down mass spectrometry to spatial and single-cell contexts. Proteoform imaging by individual ion mass spectrometry enables mapping of intact proteoforms directly from tissue sections¹⁰³, and single-cell TDP has been shown to resolve proteoform heterogeneity at the subcellular level¹⁰⁴. These capabilities can provide finer resolution of proteoform distributions within tissue architectures and at the level of individual cells — analogous to the benefits of spatial and single-cell transcriptomics — enabling deeper insights into both normal and pathological proteoform states.

The next phase of proteoform research goes beyond mapping and identification, aiming to uncover biologically relevant differences in proteoform composition between conditions and to link these differences to functional outcomes. Capturing combinatorial complexity — for example, multiple PTMs on the same proteoform — requires technologies that preserve sequence context while enabling accurate quantification¹⁰⁵. This quantitative dimension is a crucial aspect of proteoform research, as changes in proteoform abundance, stoichiometry or modification status can have direct implications for disease mechanisms and therapeutic targeting¹⁰⁶. Whereas bottom-up workflows can quantify individual PTM sites⁷⁰ or splice variants¹⁰⁷, only top-down approaches or hybrid methods can directly resolve and quantify the full combinatorial landscape of sequence variants and modifications. Advances in quantitative top-down workflows — including both label-free and label-based strategies — are enabling precise measurement of intact proteoform abundances across biological conditions^108,109.

Beyond mass spectrometry, the next generation of protein analysis is emerging, with the potential for proteoform-resolved sequencing¹¹⁰. These technologies include platforms from Quantum-Si¹¹¹ and Oxford Nanopore Technologies¹¹² and rely on novel measurement principles capable of sequencing peptides and proteins with resolution of single amino acids and even PTMs. Proof-of-concept studies suggest the potential for direct detection of proteoforms and PTMs^113,114. Recent work from Nautilus Biotechnology has demonstrated this in practice, by resolving 130 distinct full-length tau proteoform groups, including alternative splicing and PTMs, from model systems of neurodegenerative disease and from human brain tissue¹¹⁵. These findings underscore the promise of emerging, non-mass spectrometry methods for mapping proteoform diversity in disease contexts, although further advances are needed to improve throughput and coverage in complex biological samples.

Proteoform perturbation

Proteoform maps enumerate the chemical identities of proteoforms but interest lies ultimately in their biological effects and, in particular, how proteoform diversity contributes to function. To study this, the functional properties of a ‘reference’ proteoform must be compared with the measured properties of an alternative proteoform. In turn, this requires experimental approaches to manipulate the expression of individual proteoforms and a range of assays to probe the functional relationships between proteoforms (for example, loss of function or gain of function) (Fig. 3b). A representative list of techniques for proteoform perturbation assays can be found in Supplementary Table 2.

Experimental manipulation.

Studying proteoform function requires tools that go beyond whole-gene knockouts. Instead, one must selectively manipulate the abundance or activity of specific splice isoforms, translation products or PTM states, or even the stoichiometric distribution of multiple products.

A versatile molecular toolkit for the targeted perturbation of proteoforms at the DNA, RNA and protein levels continues to expand (Supplementary Table 2). At the DNA level, CRISPR-based tools, including Cas9 nucleases¹¹⁶, base editors¹¹⁷ and prime editors¹¹⁸, allow for the precise introduction of DNA variants that generate or abolish proteoforms. CRISPR interference and CRISPR activation can modulate the expression of isoform transcripts by modulating promoter usage (and hence the transcription start site)¹¹⁹. At the RNA level, approaches such as RNA interference (using small interfering RNA or short hairpin RNA)¹²⁰ and antisense oligonucleotides (including engineered splice-switching oligonucleotides¹²¹ and synthetic splice factors¹²²) can degrade or redirect the splicing of specific transcripts. More recently, dCas13-based systems have enabled knockdown of specific transcript splice forms by targeting specific exons or splice junctions^123,124. To induce the expression of particular proteoforms, engineered DNA constructs encoding genetic variants or representing alternatively spliced forms can be introduced to cells via plasmids¹²⁵, viral vectors^126,127 or transposon-based systems¹²⁸, in which the ORF constructs harbour mutations, epitope tags or specific splice sequences¹²⁹. Codon expansion and RNA editing allow for site-specific incorporation or removal of modified residues^130,131, and chemically induced proximity can be used to recruit modifying enzymes to defined RNA or protein sites¹³².

At the protein level, bioorthogonal chemistries enable selective labelling and tracking of PTM-defined proteoforms¹³³. The holy grail of proteoform manipulation is being able to fully synthesize any proteoform, but this remains technically challenging for many biological systems^134,135. One major exception is the field of chromatin biology, in which semisynthetic and chemical biology strategies have enabled precise, proteoform-specific reconstitution of nucleosomes. These approaches allow for site-specific installation of PTMs or noncanonical residues, producing homogeneously modified histones with defined sequence and modification states¹³⁶. Such PTM-defined nucleosomes have been used to assemble complexes with exact proteoform stoichiometries for in vitro mechanistic studies — for example, showing how acetylation can influence the ‘reading’ and ‘writing’ of H3K4 methylation marks¹³⁷ or how histone H3 tail conformation mediated by PTMs can inhibit association of specific chromatin-binding domains¹³⁸. Recent advances in intracellular protein editing, including transposition-based protein editing¹³⁹ and site-specific editing of endogenous proteins¹⁴⁰, may further expand the ways in which proteoforms can be generated, interrogated and even leveraged therapeutically.

Of note, no single tool can generate or edit all types of proteoform variation. Moreover, proteoforms from the same gene rarely act in isolation and can compete for partners, form dominant-negative complexes or form feedback loops, so these intra-gene, inter-proteoform interactions should be accounted for to avoid misattribution of phenotypes.

Experimental characterization.

The functional consequences of proteoform perturbation can be interrogated along the continuum of molecular-to-organismal scales. Altered molecular interactions measured by biophysical or biochemical assays are propagated through to higher-order biological levels, affecting cell phenotypes (for example, as measured by localization or drug screens) and, ultimately, organismal phenotypes and clinical outcomes (see the Network biology of proteoforms section).

At the molecular level, proteoforms differ in their capacity to interact with other macromolecules. Large-scale protein–protein interaction maps (such as the Human Reference Interactome (HuRI)¹⁴¹ and BioPlex Interactome¹⁴²) were built using a single canonical proteoform per gene; these maps serve as reference interactomes and as a baseline for comparison of alternative proteoforms. Indeed, large-scale studies have begun to systematically map proteoform-specific interactions, particularly those altered by mutations¹⁴³, splicing events¹⁴⁴ and PTMs¹⁴⁵, all of which have demonstrated widespread modulation of protein–protein interactions — often leading to gain or loss of binding partners, creation of new interaction hubs or context-specific complexes. Consistently, these protein–protein interaction studies uncover a large number of novel interactions, which are being catalogued in databases such as The International Molecular Exchange Consortium (IMEx)⁴² and CanIsoNet¹⁴⁶.

Of course, proteoforms do not only interact with protein partners but also participate in interactions with DNA, RNA and other biomolecules. This was recently demonstrated for transcription factor isoforms that bind different genomic elements¹⁴⁷. Although many binding and activity assays operate at the gene level (Supplementary Table 2), their underlying chemistries are readily adaptable to proteoform-specific application using selective enrichments or engineered tags. For example, epitope-tagged isoforms (QKI-5 and QKI-6) of the RNA-binding protein QKI were profiled using ultraviolet crosslinking and affinity purification (uvCLAP), revealing isoform-specific RNA-binding motifs and localization patterns¹⁴⁸. Similar approaches have been applied to enhanced crosslinking and immunoprecipitation (eCLIP)¹⁴⁹ and chromatin immunoprecipitation followed by sequencing (ChIP-seq)¹⁵⁰, using epitope-tagged proteins when specific antibodies are unavailable. Site-specific introduction of PTMs via genetic code expansion has enabled direct comparison of catalytic activity between proteoforms¹⁵¹, highlighting the feasibility of adapting diverse binding and enzymatic assays to the proteoform level.

Many genes produce proteoforms that have different subcellular distributions, depending on the localization signals that are retained or masked in the final protein¹⁵². For example, one study reported widespread localization differences of proteoforms resulting from pathogenic genomic variants¹⁵³. However, large-scale imaging projects such as OpenCell¹⁵⁴ and The Human Protein Atlas¹⁵⁵ have so far profiled protein subcellular localization for only one proteoform per gene.

Studies of proteoform biology at the cellular scale are only just emerging. A dCas13-based system was recently adopted for high-throughput isoform-resolved CRISPR screens^121,156. In addition, phenotypic drug screening has identified compounds that modulate proteoform functions. Notably, a thermal proteome profiling assay, which monitored proteome-wide shifts in protein stability upon compound treatment, was used to detect direct and indirect targets at proteoform resolution¹⁵⁷. Other screens, encompassing readouts, including proliferation, drug response, and bulk or single-cell gene expression, could also be used in theory to study proteoform biology^158–160. As proof-of-concept of such a study, lentiviral transduction of paired wild-type and mutant plasmids was used to generate stable cell lines, enabling functional annotation of more than 1,000 cancer-associated mutations and confirming more than 200 of these as activating mutations¹⁶¹. In vivo studies, including xenograft-based mutational screens^161,162 and mouse models allowing for manipulation of alternative splicing¹⁶³, also demonstrate a proof-of-concept for organism-based proteoform assays.

Proteoform prediction

In practice, measuring the function of every proteoform across all contexts and conditions is prohibitively resource-intensive, if not infeasible. A long-term objective, therefore, is to develop computational approaches that predict proteoform function without the need for extensive experimentation (Fig. 3c). Within the map–perturb–predict framework, predictions serve both to generate hypotheses and to elucidate fundamental molecular principles of proteoform biology (Supplementary Table 3). Prediction tasks typically fall into three areas, defined by the functional property of interest: intrinsic biophysical properties, molecular interactions and composite functional outputs.

Intrinsic properties such as structure and stability are closely associated with proteoform function. Generative, deep-learning models such as AlphaFold¹⁶⁴ can predict protein structure from sequence, so, in principle, they can be applied to distinct proteoforms. The latest iteration of AlphaFold, AlphaFold 3 (ref. 165), can model proteoforms containing numerous PTMs — including phosphorylation, acetylation, methylation, citrullination, palmitoylation and glycosylation — providing structural predictions for both unmodified and heavily modified forms. The predicted structures can then be compared to infer how mutations, PTMs or alternative splicing could alter proteoform function. Tools such as ProSTAGE¹⁶⁶ predict mutation-induced stability variations, and CHESS3 (ref. 167) is a database of splice isoform structures. Although challenges remain, such as the inability to accurately model disordered regions, these tools offer a first-pass estimate of the folding and stability of proteoforms.

Molecular interactions are often predicted by tracing how molecular changes (for example, mutations and PTMs) impact determinants of binding (for example, domains and short linear motifs). Many tools assess whether such molecular changes alter the presence, position or conformation of these elements, and then infer how such alterations might lead to the gain, loss or modulated affinity of the interaction (see Supplementary Table 3 for examples). Several tools predict how such molecular changes affect protein–protein interactions, based on how amino acid residues affect known binding domains or motifs¹⁶⁸. For example, 3did¹⁶⁹ predicts binding via experimentally determined (in other words, reported in the Protein Data Bank) and predicted domain–domain interactions. Expanding these approaches to other types of interaction (for example, with DNA, RNA or lipids) will be important for future discovery.

Tools that predict composite functional outputs comprise a heterogeneous class of approaches that infer function from a range of features, such as structural stability, network perturbation, disease relevance, gene ontology term similarity and others. Classic examples include variant effect prediction (for example, VEP¹⁷⁰ and PolyPhen¹⁷¹), which predicts loss of function of nonsynonymous variants. Other tools focus on isoform-specific functions, taking into account expression levels and protein features (for example, iMILP¹⁷²), or focus on PTM-driven changes, such as phosphorylation effects (for example, PTMfunc¹⁷³).

Despite the diversity of proteoforms, most predictive models have so far been trained on reference proteins, which limits their applicability to study alternative variants. This stems from a fundamental bias in training data: public databases overwhelmingly annotate reference forms, with a scarcity of information for alternative proteoforms. So, as the datasets generated by proteoform mapping and perturbation assays grow, predictive models should become more generalizable across different proteoforms.

Network biology of proteoforms

The map–perturb–predict framework focuses on the family of proteoforms encoded by a single gene, but proteoforms rarely act alone, instead operating within multi-scale networks. Network biology provides a framework to model proteoform behaviour within these complex systems¹⁷⁴. Furthermore, the tools and mathematical approaches of network biology can be applied to the study of complex human diseases — representing the field of proteoform medicine — to discover crucial pathways, biomarkers and drug targets¹⁷⁵.

Conventionally, in biological networks, the nodes are modelled as genes¹⁷⁶. By contrast, in proteoform-resolved networks, nodes represent discrete proteoforms, and the edges represent diverse relationships between these proteoforms — biophysical, regulatory, epigenetic or genetic — as determined by proteoform assays or derived from proteoform predictions¹⁷⁷. Comparing how nodes representing ‘reference’ proteoforms occupy a network compared with nodes representing ‘alternative’ proteoforms can reveal the effects of subtle molecular differences on rewiring higher-order pathways and driving phenotypic switches (Fig. 4). Beyond the effects of individual proteoforms, the collective behaviour of co-regulated proteoforms within a network may have a stronger influence, with distributed perturbations acting in parallel across the system. For dynamically maintained states such as reversible PTMs, introducing a single proteoform to a network outside of its native context may have limited effect, as it will quickly revert to the equilibrium determined by the surrounding regulatory network; in such cases, altering the upstream or downstream interactions that maintain this balance may be a more effective way to shift the proteoform landscape. Viewing these changes within the context of a network structure enables inference of causal relationships, not just associations, between proteoform variation and higher-order phenotypes, including clinical phenotypes in a translational context¹⁷⁸.

Fig. 4 |. Proteoform-resolved networks reveal functional rewiring beyond gene-level models.

A–C, Network models comparing the ‘reference’ proteoform network to ‘alternative’ proteoform networks provide insights into relative interaction differences and associated functional outcomes, including loss of interaction (A), gain of interaction (B) and change of interaction (C). Aa, Loss of function by loss of interaction: CREB5–204 activates transcription through MAPK9, whereas CREB5–202 loses this interaction, which impairs transcriptional activation. Ab, Gain of function by loss of interaction: PIK3CA with E545K and/or H1047R mutations loses its interactions with BPIFA1 and SCGB2A1, which promotes conversion of PIP₂ to PIP₃ and thus activates the PI3K–AKT pathway. Ba, Loss of function by gain of interaction: the non-phosphorylated IκBα proteoform gains an interaction with NF-κB, thereby inhibiting the NF-κB-mediated transcription of inflammatory genes. Bb, Gain of function by gain of interaction: BRAF V600E mutant acquires a new interaction with KEAP1–NRF2, which promotes oncogenic transcriptional activation. Ca, Change of function by change of interaction: FOXP1-ES activates the transcription of pluripotency-associated genes (such as OCT4 (also known as POU5F1), NANOG and NR5A2), whereas FOXP1 activates the transcription of differentiation-related genes (such as GAS1, HESX1 and SFRP4). Cb, Change of localization by change of interaction: RBFOX1–240 binds nuclear RNAs associated with neurodevelopment, whereas RBFOX1–207 binds cytoplasmic RNAs associated with neuronal excitability, which are enriched in loci identified by genome-wide association studies for neurological disease such as psychiatric disorders and epilepsy.

Causes of network rewiring

Genetic variation.

Genetic mutations frequently drive proteoform-specific network rewiring by altering interaction patterns rather than simply causing loss of protein function or misfolding. A landmark study provided early evidence for this, showing that most disease mutations do not simply destabilize the mutated protein but rather result in perturbations of protein–protein interactions that are characteristic of the genetic disorder¹⁴³. This result was reinforced by a later study showing that mutations affecting protein domains or motifs can rewire network connectivity at the proteoform level¹⁷⁹. Another study used interactome analysis of genes containing loci that had been identified by genome-wide association studies to detect high-impact functional network modules underlying complex diseases¹⁸⁰. In cancer, mapping of protein interactomes in head and neck tumours revealed that specific mutations of PIK3CA induce unique patterns of network rewiring, uncovering mutation-specific therapeutic vulnerabilities¹⁸¹. By integrating genomic mutations and splicing events into functional association networks, a pan-cancer analysis using the network-based strategy DrAS-Net linked more than 2.5 million somatic variants to more than 40,000 driver mutations and 80,000 splicing changes across 33 tumour types, revealing that single-nucleotide variants can reshape isoform profiles and underlie cancer heterogeneity. This highlights how genetic variation can drive widespread changes in alternative splicing, providing a direct link between mutations and isoform-mediated network remodelling¹⁸².

Alternative splicing.

Alternative splicing is another important mechanism driving proteoform network remodelling. Tissue-specific splicing events can reshape protein interactomes, in which intrinsically disordered regions of proteins participate in interactions based on the differential presentation of binding motifs in different proteoforms, dynamically altering network connectivity^183,184. Systematic analyses have emphasized how alternatively spliced exons can remodel interactomes, often creating multitasking proteoforms¹⁸⁵, a conclusion that agreed with a study reporting the extensive expansion of protein interaction capabilities generated through alternative splicing¹⁴⁴. In another study, isoform-specific interaction networks were used to reveal pathogenic isoform switching across 1,209 cancer samples spanning multiple cancer types, with widespread clinical implications¹⁸⁶. Sometimes, such information can be inferred from the network topology itself, as in the tool NEASE (network-based enrichment method for alternative splicing events), which incorporates collective interaction changes to predict isoform functions¹⁸⁷.

Post-translational modifications.

PTMs — particularly phosphorylation — modulate network architecture at short time scales relevant to cellular responses. For example, phosphorylation of the molecular chaperone HSP90 markedly altered its interaction network, with broad downstream effects on global protein folding¹⁸⁸. Similarly, kinase-regulated networks were shown to be influenced by phosphorylation of key hubs in the signalling network, such as glycogen synthase kinase 3β (GSK3β)¹⁸⁹. Beyond phosphorylation, the necessity of integrating phosphoproteome data with glycoproteome data has been demonstrated, revealing how crosstalk between PTMs can affect disease-related protein networks¹⁹⁰.

Intra-gene proteoform relationships.

Proteoforms encoded by the same gene are often mechanistically linked, with sequence variants or PTM patterns acting in hierarchical or cooperative ways. For example, in the NFAT transcription factor family, phosphorylation at specific serine residues prevents nuclear import, and sequential dephosphorylation by calcineurin is required for activation — a clear example of ordered PTM dependencies within a proteoform series¹⁹¹. Similarly, in p53, phosphorylation at Ser15 enhances acetylation at Lys382, which in turn promotes DNA binding and transcriptional activation, illustrating how one type of proteoform modification can potentiate additional regulatory layers¹⁹². Alternative splicing can generate isoforms that directly regulate each other’s function; for example, the BCL-xL and BCL-xS splice variants interact physically with each other, with the pro-apoptotic BCL-xS antagonizing the anti-apoptotic activity of BCL-xL¹⁹³. Enzymes that ‘write’, ‘read’ and ‘erase’ PTMs often have proteoform specificity, such as histone methyltransferases that recognize distinct combinations of histone variant and PTM¹⁹⁴.

Even when the underlying mechanisms remain unresolved, proteoform dependencies can be inferred quantitatively. TDP of intact proteins and middle-down proteomics of large peptides have enabled the co-occurrence patterns of PTMs to be measured, as well as, in some cases, the directionality of their relationships. For example, large-scale analyses of histone tails have revealed global patterns of crosstalk between co-existing PTMs^195,196, and visualization tools¹⁹⁷ can capture and interpret these dependencies. Incorporating such quantitative relationships — whether mechanistically mapped or statistically inferred — will be essential for building accurate, predictive models of proteoform networks.

Effects of network rewiring

Knowledge of proteoform-specific network perturbations provides a scaffold to understand how changes to molecular interactions affect phenotype through loss, gain or change of network function, in a manner that is often nonlinear or non-obvious. To illustrate these mechanistic principles of proteoform network rewiring, we highlight ways in which proteoform-driven interaction changes can alter network topology and lead to distinct phenotypic outcomes (Fig. 4). These insights suggest new therapeutic possibilities; because each pathological outcome results from a specific topological disturbance that is caused by a proteoform — resulting in the loss, gain or rerouting of network interactions — the corrective strategy must in some way restore the network back to its physiologically normal state.

Loss of interaction.

Loss of interaction is a common mechanism by which proteoforms alter cellular behaviour, with the functional outcome depending on the network context (Fig. 4A). For example, CREB5–202 cannot bind MAPK9 and therefore has attenuated transcriptional activation relative to the canonical CREB5–204 isoform¹⁴⁷ (Fig. 4Aa). Counterintuitively, loss of interaction can also result in gain of function. This is illustrated by PIK3CA, which normally is engaged in inhibitory interactions with BPIFA1 and SCGB2A1 that constrain activation of the PI3K–AKT pathway. Oncogenic PIK3CA proteoforms, containing mutations such as E545K and/or H1047R, abolish these inhibitory interactions to activate constitutive signalling, leading to tumour proliferation¹⁹⁸ (Fig. 4Ab). Finally, a change of function can also result from loss of interaction, as exemplified by TDP-43, a protein with a central role in neurodegeneration. Phosphorylation at specific serine residues (Ser379, Ser403 and Ser409) alters the localization and RNA-binding profile of TDP-43, producing an aggregative cytoplasmic proteoform that is unable to regulate nuclear splicing of tau pre-mRNA. The resulting shift in tau isoform ratios contributes to multiple neurodegenerative disease phenotypes¹⁹⁹.

Gain of interaction.

Gain-of-interaction events can create new regulatory links that alter cellular behaviour. For example, the non-phosphorylated IκBα proteoform gains an interaction with NF-κB, sequestering it in the cytoplasm and thereby preventing transcriptional activation of inflammatory genes (Fig. 4Ba). This gain of interaction leads to a loss-of-function phenotype, in contrast to the phosphorylated IκBα proteoform that is targeted for degradation. Similarly, the BRAF V600E proteoform acquires the ability to interact with KEAP1, a repressor of NRF2 (ref. 200) (Fig. 4Bb), which prevents NRF2 degradation; in turn, this activates NRF2-dependent, antioxidant transcriptional programs that support tumour survival. Here, a single proteoform-specific interaction rewires the network towards an oncogenic state, highlighting how a gain of function can result from novel molecular contacts.

Change of interaction.

A change of interaction for different proteoforms resulting in a change of function is illustrated by FOXP1 isoforms, for which proteoform-specific alterations in DNA-binding specificity can drive opposing regulatory programs (Fig. 4Ca). The FOXP1-ES splice isoform preferentially binds the promoters of pluripotency-associated genes (such as OCT4 (also known as POU5F1, NANOG and NR5A2), thereby repressing cellular differentiation pathways and maintaining pluripotency²⁸. By contrast, the canonical FOXP1 isoform promotes cell differentiation by activating target genes associated with development (such as GAS1, HESX1 and SFRP4).

Changes in interaction partners leading from a change in localization are also commonly observed for different proteoforms. For example, RBFOX1 isoforms engage different target RNAs and have different localizations²⁰¹. RBFOX1–207 localizes to the cytoplasm and binds RNA transcripts linked to neuronal excitability (such as GOT1, SNAP25 and SCN8A), whereas RBFOX1–240 remains nuclear and regulates splicing programs associated with neurodevelopment (such as MAPK10, CDK14 and DCTN1) (Fig. 4Cb). These spatially driven changes in RNA-binding activity link proteoform variation to distinct neuropsychiatric outcomes.

Proteoform-specific therapeutic targeting

The concept of proteoform medicine calls for a shift in perspective — from viewing disease in the context of genetic changes to seeing it as a consequence of altered nodes (proteoforms) and edges (interactions) in molecular networks. To guide therapeutic design in this context, we introduce a conceptual framework for prospective proteoform-targeted intervention strategies, organized according to increasing network scope: (1) node correction, which targets the disease-associated proteoform directly; (2) edge correction, which aims to restore or block specific interactions of disease-associated proteoforms; and (3) neighbourhood correction, which modulates the network surrounding a disease-associated proteoform to buffer or redirect dysregulated pathways. We also introduce proteoform-resolved precision medicine approaches for diagnostic and prognostic biomarkers. Clinically relevant examples are used to illustrate each strategy (Fig. 5).

Fig. 5 |. Proteoform-resolved therapeutic strategies across molecular and network scales.

Proteoform-level dysregulation underlies diverse pathological states and can be targeted through a range of therapeutic strategies. When possible, clinically relevant examples are used to illustrate these strategies. a, Node correction. Strategies including gene or protein therapy, chemically induced modulation of post-translational modifications (PTMs), antisense oligonucleotides (ASOs; including engineered splice-switching oligonucleotides (SSOs)) and proteolysis-targeting chimeras (PROTACs) can be used to directly alter the abundance or state of the pathological proteoform. b, Edge correction. Strategies to target interactions that are altered by disease-associated proteoforms include interaction inhibitors or stabilizers. c, Neighbourhood correction. The activity or availability of the primary interacting partners of a proteoform can be modulated to compensate for proteoform dysfunction. d, Extended neighbourhood correction. Targeting downstream or upstream nodes (secondary interactors) or bypassing the pathological module in the proteoform-resolved network can be used to restore system-level balance.

Node correction

An abnormal reduction in the concentration of a particular proteoform requires strategies to replenish the expression of functional proteoforms (Fig. 5a). Gene therapies are the most direct route to achieve this, by delivering DNA or RNA constructs to replace mutated genes. Notable clinical successes of this approach include the adenovirus-based gene therapy onasemnogene abeparvovec, which introduces a functional copy of the SMN1 gene in patients with spinal muscular atrophy²⁰². Similarly, wild-type proteoforms can be replaced by delivering recombinant proteins or functional mRNA; for example, the recombinant protein isoform EDA-A1 is introduced during pregnancy to treat perinatal X-linked hypohidrotic ectodermal dysplasia²⁰³.

Proteoform modification involves making changes to the protein molecule directly to convert it to a related proteoform, for example by modulating PTMs. Therapeutic strategies include the use of targeted covalent inhibitors²⁰⁴ or chemically induced proximity¹³² to selectively inhibit or introduce PTMs that alter proteoform activity. Clinical applications include kinase inhibitors such as osimertinib, which is used to target mutant EGFR proteoforms in lung cancer and inhibit their phosphorylation²⁰⁵.

Proteoform removal aims to eliminate aberrant, mis-expressed or toxic protein species. RNA-targeting approaches, such as antisense oligonucleotides and splice-switching oligonucleotides, manipulate RNA splicing patterns to selectively reduce or eliminate pathological proteoforms²⁰⁶. Clinically validated examples include nusinersen, an FDA-approved splice-switching oligonucleotide therapy for spinal muscular atrophy, which restores production of functional SMN protein by promoting exon inclusion²⁰⁷. A post-translational approach involves the use of proteolysis-targeting chimeras (PROTACs), which co-opt the ubiquitin ligase system to selectively degrade the pathological protein²⁰⁸. Investigational PROTACs such as ARV-110 (ref. 209) and ARV-471 (ref. 210) target androgen receptor and oestrogen receptor proteoforms, respectively, in advanced cancers.

In some diseases, pathology does not arise from a single toxic proteoform but from the disrupted balance between functionally distinct proteoforms, as seen in tauopathies. Changes in the ratio of three-repeat (3R) to four-repeat (4R) tau isoforms are associated with distinct disorders, with 3R predominance associated with Pick’s disease and 4R predominance associated with progressive supranuclear palsy and corticobasal degeneration²¹¹. Both 3R and 4R tau isoforms can contribute to pathology, either independently or through cooperative mechanisms such as co-aggregation, with their relative abundance — rather than absolute presence — being a crucial determinant of disease phenotype²¹². Quantitative proteoform-resolved measurements in patient tissues, including liquid chromatography–mass spectrometry quantification of 3R versus 4R tau isoforms and phosphorylated peptides across tauopathies²¹³, are beginning to define these ratio-dependent associations. Effects of relative proteoform abundance occur similarly in other contexts, such as for myosin heavy chain isoforms in cardiomyopathies²¹⁴.

Edge correction

Therapeutic interventions can also aim to correct undesirable patterns of molecular interaction between proteoforms²¹⁵ (Fig. 5b). Two main strategies are emerging: interaction inhibition and interaction stabilization. Interaction inhibition blocks harmful interactions to prevent the pathological activation of downstream effectors. For example, bemarituzumab, which is currently in phase III clinical trials for gastric cancer, selectively targets the pathological FGFR2b proteoform to block binding of fibroblast growth factor²¹⁶. By contrast, interaction stabilization involves the use of peptides or small molecules to restore physiological interactions that are disrupted by pathological proteoforms. Interaction stabilization is a largely unexplored area; however, paclitaxel, which is one of the oldest and most widely used chemotherapeutics, acts by stabilizing microtubule interactions²¹⁷.

Neighbourhood correction

When direct targeting of a pathological proteoform or its interactions is not feasible, knowledge of its broader network context can be leveraged to identify indirect therapeutic targets. Such approaches may involve drug repurposing to identify FDA-approved compounds capable of influencing disease-associated networks, including the primary or secondary interactors of the pathological proteoform²¹⁸ (Fig. 5c). For example, BET inhibitors have been shown to rewire transcriptional networks in cancer²¹⁹. Furthermore, the functional effects of therapeutically intractable proteoforms could be modulated by degrading crucial primary or secondary interactors using antisense oligonucleotides or PROTACs. Pharmacological bypassing further exemplifies the concept of network-based interventions by targeting compensatory pathways or tertiary nodes to mitigate pathological effects indirectly (Fig. 5d). For example, in preclinical models of Cockayne syndrome B (CSB, also known as ERCC6) mutations, small-molecule BDNF mimetics (such as amitriptyline and 7,8-dihydroxyflavone) successfully restore neuronal differentiation by indirectly reactivating downstream neurotrophic signalling pathways that are disrupted by the primary genetic defect²²⁰. These network-based drug discovery and repurposing approaches could offer rapid clinical translation of proteoform-specific targets.

Biomarkers for diagnosis and prognosis

Proteoform-resolved diagnostics and prognostics should improve patient risk stratification, disease diagnosis and therapeutic decision-making. The integration of proteoform-quantitative trait loci-informed polygenic risk scores²²¹ could enhance traditional genetic risk assessments by incorporating genetic variants that are specifically linked to pathological proteoform production. In addition, RNA transcript biomarkers²²² and proteoform biomarkers could be used as precise indicators of disease states and treatment responses. Examples include clinical trials for use of the AR-V7 isoform as a biomarker of resistance to androgen receptor-targeted therapies in prostate cancer²²³ and detection of soluble VEGFR1 isoforms to predict pre-eclampsia²²⁴; for both of these examples, proteoform-level insights improve clinical outcomes.

Proteoform-resolved biomarkers can also be used for therapeutic targeting. Cell-specific or context-specific proteoforms can be used as selective markers for targeted interventions such as T cell engagers, chimeric antigen receptor (CAR) T cells, antibody–drug conjugates and epitope-directed therapies²²⁵. These proteoform-based targeting strategies can improve precision and minimize off-target effects, particularly in immuno-oncology and tissue-specific diseases.

Future directions for proteoform medicine

Proteoforms are not just gene products but distinct molecular entities with their own clinical relevance¹³. If twentieth century molecular biology was organized around the gene, the twenty-first century should be focused on the proteome and dissecting relevant proteoforms, which promises transformative advances in medicine. The future of proteoform medicine will depend on the systematic integration of proteoform mapping, functional characterization and predictive modelling, all within an iterative loop that is tightly integrated into a network-modelling framework¹⁷. This vision requires progress in several areas that will increase the translational relevance of studying proteoforms.

Studying human-relevant proteoforms

Proteoform-resolving technologies often require high-quality specimens available in sufficient quantities, which favours the investigation of simpler model systems for feasibility. However, most widely used experimental models — such as cell lines and mice — fail to capture the full spectrum of endogenous human proteoforms, many of which are lineage-specific²²⁶ or species-specific²²⁷. Even when genes are conserved between species, their proteoforms may not be, and the dogma that gene conservation equals conservation of function does not always hold, as has similarly been shown for non-coding RNAs²²⁸. Clinical translation requires systems that faithfully recapitulate human-specific proteoform diversity in disease-relevant contexts. Findings from simpler models must ultimately be validated in more physiologically relevant models, such as organoids, xenografts and humanized mice, or directly in patient-derived samples²²⁹. Equally important is the need to distinguish biologically relevant proteoforms from artefacts introduced by sample handling, aberrant chemical modification or analytical limitations. All proteoform-resolving approaches carry inherent risks of false positivity, particularly when working at the limits of detection. Rigorous quality-control measures, including orthogonal validation, standardized benchmarking datasets and internal standards, will be essential to ensure that candidate proteoforms reflect true biology rather than technical noise.

Visualization and standardization of data

An area that warrants active development to broaden accessibility and integration is the representation and visualization of proteoforms. Proteoform biology occupies the intersection of the disparate fields of genomics, transcriptomics and proteomics, each having unique data standards and disciplinary culture. New, interactive frameworks are needed to visualize and compare the complex groupings of protein variation events. Recent efforts include Biosurfer²³⁰, a tool for multi-hierarchical comparison of proteoforms in genomic context, and the development of community standards for proteoform definition and exchange, such as ProForma²³¹ and PEFF²³².

Standardizing annotations for clinical use

For the increasing knowledge of proteoforms to be clinically actionable, annotation systems must evolve beyond gene-based frameworks to support clinical decision-making. Proteoform data must be translated into standardized, interpretable formats populated with confidence metrics, functional scores and consistent nomenclature. High-quality annotation will require interdisciplinary collaboration to support integration with genomic reporting tools and electronic health records. Frameworks for the annotation of gene variants, such as ClinGen²³³, can be emulated. Importantly, these resources should support dual modes of use: clinician-facing interfaces that emphasize patient risk and treatment relevance, and researcher-facing views that provide mechanistic and structural detail for discovery and validation.

Application across diverse populations

Proteoform profiles vary by ancestry and are shaped by genetic background, environment and lifestyle, which together influence disease risk, therapeutic response and clinical outcome. However, most protein biomarkers and drug targets are derived from datasets that lack ancestral diversity, such as diabetes biomarkers²³⁴ or protein quantitative trait loci from the UK Biobank²³⁵. To increase the generalizability and value of these efforts, proteoform-centric strategies must be integrated into population-scale studies. This will require building ancestry-aware proteoform reference maps and deploying proteoform-resolved assays across global biobanks and clinical trials. Initiatives such as the T2T (ref. 236) and Pan-Genome²³⁷ projects could serve as models. Finally, realizing the full promise of proteoform medicine will require methods and algorithms that integrate proteoform-resolved proteomics with genomics, transcriptomics, metabolomics and phenomics. Cross-omics integration will not only improve the robustness of proteoform identification but also enable a systems-level understanding of how specific proteoforms drive disease biology and therapeutic response.

Supplementary Material

Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41576-025-00915-1.

Related links

3did: https://3did.irbbarcelona.org/

BioPlex Interactome: https://bioplex.hms.harvard.edu/

CanIsoNet: https://www.caniso.net/

CHESS3: https://ccb.jhu.edu/chess/

COSMIC: https://cancer.sanger.ac.uk/cosmic/login

CTDP: https://ctdp.org/

dbSNP: https://www.ncbi.nlm.nih.gov/snp/

Human Proteoform Project: https://proteomics.northwestern.edu/services/human-proteoform-project/

HuRI: https://interactome-atlas.org/

IMEx: https://www.imexconsortium.org/

MassIVE: https://massive.ucsd.edu/ProteoSAFe/static/massive.jsp

OpenCell: https://opencell.sf.czbiohub.org/

PolyPhen: http://genetics.bwh.harvard.edu/pph2/

Protein Data Bank: https://www.rcsb.org/

TDPortal: https://nrtdp.northwestern.edu/tdportal-request/

The Human Protein Atlas: https://www.proteinatlas.org/

TranslatomeDB: http://www.translatomedb.net/

UniProt: https://www.uniprot.org/

VEP: https://www.ensembl.org/info/docs/tools/vep/index.html