The research, led by Tülay Karakulak and Abdullah Kahraman of the Swiss Institute of Bioinformatics and their collaborators, including Holger Moch of the University Hospital Zurich and Christian von Mering of the University of Zurich, drew on an extraordinary wealth of data. The team analysed bulk RNA sequencing from 1,117 cancer patients with metastatic melanoma, ovarian cancer, or acute myeloid leukemia (AML), combining samples from the Tumor Profiler (TuPro) Project, a multi-omics clinical study based in Zurich, with data from The Cancer Genome Atlas (TCGA). To determine what a ‘normal’ splicing pattern looks like, they compared the tumour samples against 818 healthy tissue samples from the Genotype-Tissue Expression (GTEx) project, using skin, ovary, and blood samples as the appropriate reference tissues. For each gene in each sample, the researchers identified the Most Dominant Transcript, the isoform that is expressed at the highest level, and then looked for cases where the dominant isoform in a cancer sample differed from the dominant isoform in healthy tissue. These cancer-specific dominant transcripts, abbreviated cMDTs, are the molecular fingerprints of isoform switching.

The scale of what they found is striking. Across the three cancer types, the team identified 6,114 unique cancer-associated dominant transcripts arising from 4,151 genes. Of these, 1,707 encode transmembrane proteins, a particularly important class because many diagnostic assays and targeted therapies act on proteins embedded in the cell membrane. Most of these switches were rare, patient-specific events: 41 percent of the cMDTs were detected in only a single sample, and the majority appeared in fewer than 10 percent of samples within a cohort. Yet 526 of the switches were found in more than 30 percent of samples in at least one cancer cohort, and 131 of those involved transmembrane proteins. Metastatic epithelial ovarian cancer and metastatic melanoma showed the highest numbers of switching events, while AML showed the fewest, confirming earlier observations that splicing aberrations differ markedly between cancer types. The researchers also found that samples in which a spliceosome gene, encoding a component of the splicing machinery itself, had undergone a dominant-transcript switch carried significantly more cMDTs than other samples, whereas mutations in spliceosome genes without a switch did not produce the same effect. This suggests that functional disruption of the splicing machinery, rather than mutation alone, drives large-scale splicing chaos in tumours.

Among the most frequent switches was one involving the cytoskeletal gene ADD3, or adducin 3. In ovarian cancer, a longer isoform of ADD3 that includes an extra exon was the dominant transcript in 82 percent of metastatic TuPro samples and 91.3 percent of TCGA samples. Previous work in lung cancer has shown that inclusion of this exon, exon 14, is associated with poor prognosis, likely because downregulation of the RNA-binding protein QKI-5, which normally prevents that exon from being included, promotes cell proliferation. Another notable example in ovarian cancer concerned SLC44A2, the gene encoding the choline transporter-like protein CTL2. Healthy ovarian tissue predominantly produces a 704-amino-acid version called CTL2-P1, whereas ovarian tumours frequently switch to a slightly longer variant, CTL2-P2, which uses a different starting exon and carries a distinct N-terminal region. Crucially, only the CTL2-P2 variant has detectable choline transport activity. Because aberrant choline metabolism is known to fuel cancer progression, targeting this isoform switch could offer a therapeutic angle in ovarian cancer, and the team noted that nearly 8 percent of TCGA ovarian samples also carry amplifications of the SLC44A2 gene.

In melanoma, the spotlight fell on PDCD6, a gene involved in programmed cell death. The dominant isoform in healthy skin encodes a protein with five EF-hand domains, structural motifs that bind calcium and mediate interactions with other proteins. The cancer-specific isoform retains only two of these domains, which likely cripples its ability to interact with binding partners such as DAPk1, which promotes apoptosis, and VEGFR-2, through which PDCD6 inhibits angiogenesis. In other words, the tumour appears to be splicing away the very machinery that would help kill it. In AML, the team identified a switch to the longest isoform of USP34, a ubiquitin-specific peptidase whose overexpression has been linked to tumour growth in pancreatic cancer and laryngeal carcinoma. And in a striking ovarian cancer finding, the tumour suppressor TP53 itself showed isoform switching in 20 samples, with 18 samples expressing a 261-amino-acid variant known as Δ133p53α. This isoform lacks both transactivation domains, a proline-rich domain, and part of the DNA-binding domain, and according to the team’s interaction database it cannot bind critical partners including MDM2, MDM4, POLR2E, and EP300, fundamentally altering the behaviour of one of the most studied proteins in all of cancer biology.

Identifying these switches at the RNA level is one thing; proving that the altered proteins are actually made is another. To do this, the researchers turned to data-independent acquisition mass spectrometry from matched TuPro protein samples, developing a three-tiered evidence system. Class I evidence comes from peptides unique to the cancer-specific isoform, the strongest possible proof that the switched protein is translated. Class II evidence uses peptides found on the cancer isoform and other non-canonical isoforms but absent from the normal dominant transcript. Class III evidence, the weakest tier, relies on differential expression patterns of multiple peptides between samples with and without the switch. Together, these approaches validated 12.1 percent of the more than 33,000 identified cMDTs, one of the highest detection rates reported to date. The validated switches included the ADD3 long isoform, detected with a specific peptide in 45 percent of matched ovarian cancer proteomics samples, and a long isoform of ATP2B4, a calcium-transporting enzyme implicated in apoptotic resistance in pancreatic cancer, found in 77.5 percent of matched samples. Importantly, a search of normal ovary proteomics data in the PaxDB database found no trace of the ADD3 cancer-specific peptide, underscoring its potential as a tumour-specific marker.

Perhaps the most consequential part of the study concerns what these switched proteins can no longer do. The team built an updated version of their CanIsoNet database, mapping nearly 100,000 protein isoforms onto the latest STRING protein interaction network and the 3did database of domain-domain interactions. Among 1,652 cMDTs with domain interaction information, 906 had lost all interactions with their protein partners, while only 718 retained all of them. A vivid example is TRAPPC5, a subunit of a vesicle-transporting complex that shuttles cargo from the endoplasmic reticulum to the Golgi. In melanoma, a truncated 121-amino-acid isoform of TRAPPC5 was the dominant transcript in up to two-thirds of samples, compared with a 188-amino-acid version in healthy tissue, and the shortened protein loses every one of its native interaction partners, including clinically annotated proteins. High expression of this truncated isoform was associated with lower survival in TCGA data, whereas expression of the normal isoforms had no survival impact, marking it as a candidate biomarker for melanoma.

The implications for drug treatment may be even more significant. The researchers constructed a new database, called IsoDrug, that links protein isoforms to the small molecules and drugs that bind them, drawing on structural data from more than 8,600 entries in the Protein Data Bank. Of the 641 cMDTs with drug interaction information, whose canonical counterparts are targets of 166 drugs, 56 percent had lost the binding residues needed to interact with their potential drug targets. Among tyrosine kinase inhibitors, the workhorse drugs of precision oncology, the team found 15 cancer-specific isoforms whose canonical versions bind one of 11 TKI compounds. The most frequent was a 154-amino-acid isoform of PDGFRA, expressed predominantly in melanoma samples, which lacks the entire tyrosine kinase domain that the drug sunitinib targets. A similar case is EPHB4, a receptor overexpressed in ovarian cancer, whose cancer-specific isoform retains only the ligand-binding domain and has lost the complete kinase domain targeted by inhibitors. Patients whose tumours predominantly express these truncated isoforms would likely derive no benefit from the corresponding drugs, even if the gene itself appears amplified or mutated on standard genomic tests.

The authors are careful to note the limitations of their work. Many of the single-sample switches could reflect technical variability rather than true biology, and the short-read sequencing they used cannot capture full-length transcripts as precisely as emerging long-read methods would. They also defined cancer-specific switches strictly within each tissue context, meaning some of these isoforms may exist in other healthy tissues. Still, the path forward is clear: the team highlights isoform-specific peptides from ADD3, ATP2B4, and PARP14 as prime candidates for liquid biopsy tests that could detect cancer or guide therapy decisions from a simple blood sample, and they call for drug-screening experiments in patient-derived cell cultures to confirm the predicted losses of drug binding. What emerges from this study is a compelling message for oncology: reading a tumour’s genome is no longer enough. To understand what a cancer is truly capable of, and which drugs it will resist, clinicians and researchers must also read the tumour’s RNA, exon by exon, and ask not just which genes are active, but which versions of their proteins the cancer has chosen to build.

Subject of Research: Alternative splicing isoform switches and their effects on protein and drug interactions in melanoma, ovarian cancer, and acute myeloid leukemia

Article Title: Switches in alternatively spliced isoforms impact protein and drug interactions in melanoma, ovarian cancer, and acute myeloid leukemia

Article References: Karakulak, T., Tumor Profiler Consortium, Aebersold, R., Ak, M., Al-Quaddoomi, F. S., Albert, S. I., Albinus, J., Alborelli, I., Andani, S., Attinger, P.-O., Bacac, M., Baciu-Drăgan, M.-A., Baumhoer, D., Beck-Schimmer, B., Beerenwinkel, N., Beisel, C., Bernasconi, L., Bertolini, A., Bodenmiller, B., … Kahraman, A. (2026). Switches in alternatively spliced isoforms impact protein and drug interactions in melanoma, ovarian cancer, and acute myeloid leukemia. British Journal of Cancer. https://doi.org/10.1038/s41416-026-03638-0

Image Credits: AI Generated

DOI: 10.1038/s41416-026-03638-0

Keywords: alternative splicing, isoform switching, melanoma, ovarian cancer, acute myeloid leukemia, biomarkers, drug resistance, proteomics, spliceosome, TP53, PDGFRA, transmembrane proteins

TLR4 sits at the front line of the body’s defense against bacteria. When the portion of the receptor exposed outside the cell detects a bacterial molecule, the receptor undergoes changes that prompt two TLR4 proteins to come together as dimers, and those dimers can then trigger cascades inside the cell, including the NF-κB pathway, one of the principal drivers of inflammation. The receptor can also respond to certain non-bacterial molecules associated with tissue damage, which helps explain why its misfiring is tied to such a broad range of inflammatory conditions. Despite decades of interest, no FDA-approved drug specifically blocks TLR4, in part because the receptor’s most druggable-looking surfaces are also the hardest to engage with precision. The Scripps team’s strategy was to aim not at the well-studied extracellular region but at the stretch of the protein buried inside the membrane, a site long considered beyond the reach of rational design.

The biological rationale for this approach emerged from a simple question posed by first author Colleen Maillie, a research project analyst at Scripps Research. Scientists had generally assumed that the regions of TLR4 exposed outside and inside the cell were the main signaling drivers, with the membrane-spanning segment serving a purely structural role. To test that assumption, the researchers introduced a fragment of TLR4, comprising the membrane-spanning region plus a small neighboring section, into human cells grown in the laboratory. The fragments readily associated with full-length TLR4 within the membrane, an interaction the team detected using a screening method recently developed in the Mravic lab that emits light when tagged proteins come into close proximity. Crucially, the presence of these fragments reduced overall NF-κB signaling responses, providing the first concrete evidence that the transmembrane region is not a passive tether but an active determinant of the receptor’s inflammatory output.

That screening assay matters as much as the molecules it helped find. Measuring whether two proteins interact inside a lipid bilayer is far harder than doing so in water, because most conventional biochemical tools are optimized for aqueous environments. The Mravic lab’s light-up approach was built specifically to identify which synthetic proteins target membrane proteins, and the TLR4 fragments provided an early demonstration that the method works in practice. With that tool in hand, the team could move from observation to engineering: if simply presenting the receptor with its own transmembrane sequence could dampen signaling, then a purpose-built synthetic protein designed to bind that region more tightly might dial the response down even further.

Designing such a molecule required confronting a stubborn gap in computational biology. Protein design software has matured enormously for soluble proteins, whose folding rules in water are increasingly well captured by physics-based equations and machine learning models. Membranes are a different story. The cell membrane consists of two compact layers of oily molecules, an environment with biochemical properties radically different from the water-based settings where most proteins reside, and many of the rules governing how proteins fold and function within those greasy layers remain poorly understood. As Mravic explained, models for protein interactions in water have become increasingly accurate, but for membrane proteins they are not, because there are unique atomic details underlying molecular biophysics in lipid bilayers that current equations and AI models do not accurately capture.

Unable to rely on the software’s raw predictions, the team treated the computer-generated three-dimensional blueprints as starting points rather than finished designs. They then applied custom design criteria developed in the Mravic lab to optimize what the researchers call apolar packing, the tightness with which the chemical structures of the designed proteins fit together with their TLR4 target. The underlying theory, which the team encoded into software, holds that maximizing apolar packing produces more stable protein interactions within the hydrophobic membrane environment. From the many structural options generated, the researchers narrowed the field to the nine candidates with the best predicted biophysical and chemical features and took those forward into tests in living cells.

The results validated the strategy with striking efficiency. Using the same light-up screening approach, the team found that eight of the nine synthetic proteins showed signs of associating with TLR4. Three emerged as the strongest candidates, and one, designated Design-6, showed the most robust evidence of interaction along with several other appealing qualities. Unlike the naturally occurring transmembrane fragments tested earlier, Design-6 did not aggregate as much, a critical property for any molecule that might eventually be developed as a biologic. Most importantly, Design-6 substantially reduced NF-κB inflammatory signaling, demonstrating that a rationally designed transmembrane protein could not merely bind to its target but measurably alter the receptor’s biological function.

The implications extend well beyond TLR4 itself. The study establishes that the membrane-embedded regions of immune receptors can be treated as legitimate, programmable drug sites, opening what the researchers describe as a potential launchpad for a new class of biologics that operate from within the membrane. It also delivers a set of computational tools, built around the apolar packing design criteria, that other laboratories can apply to proteins embedded in membranes, a category that includes many of the most sought-after targets in modern pharmacology. Because roughly a quarter of known drug targets are membrane proteins, design methods that work in lipid bilayers rather than despite them could reshape how researchers approach targets that conventional antibodies and small molecules have failed to reach.

Considerable work remains before any of this reaches patients. The experiments were conducted in human embryonic kidney cells, which are convenient for laboratory study but not especially relevant to inflammation-driven disease, so the team must confirm that the approach holds in more disease-relevant cell types such as liver cells and immune cells. Delivery poses another obstacle: these are oily synthetic proteins that must reach and insert into cell membranes, and no established clinical method yet exists for getting them there. Maillie acknowledged the challenge candidly, noting that this is an innovative space that carries a lot of risk and a long roadmap to the clinic, but emphasizing that the team has shown that with clever design and biophysics, they are getting closer. Maillie, Ward and Mravic are inventors on a provisional patent application titled Compositions and Methods for Inhibiting Toll-Like Receptor 4 Mediated Inflammation, signaling that the group intends to pursue the therapeutic potential of the platform.

What makes the study resonate beyond its immediate findings is the way it reframes a familiar target. TLR4 has long been described as a key sensor of bacteria that activates and mobilizes immune cells to fight infection, and, in Ward’s framing, as both a sensor and a dial to tune innate immunity, one that vaccine adjuvants can turn up and inhibitors can turn down. The Scripps work shows that the dial’s most sensitive setting may lie in the part of the receptor nobody thought to touch. By proving that a designed transmembrane protein can find its target, bind it, and quiet the inflammatory signal it produces, the study turns a long-standing blind spot in protein design into a working blueprint, and hands the field both a molecule and a method for building the next ones.

Subject of Research: De novo computational design of transmembrane proteins that inhibit the innate immune receptor TLR4

Article Title: Computer-designed protein targets immune receptor linked to inflammation at a new “undruggable” site

Article References: Computer-designed protein targets immune receptor linked to inflammation at a new “undruggable” site. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: TLR4, protein design, transmembrane proteins, innate immunity, inflammation, NF-κB signaling, Scripps Research, membrane biophysics, apolar packing, drug design, computational biology, PNAS