Glypican-3, often abbreviated as GPC3, is a member of the glypican family of cell-surface proteins that are anchored to the outer membrane by lipid chains. It has attracted intense interest in oncology because it is overexpressed in a large fraction of hepatocellular carcinomas, and its position on the extracellular surface of tumor cells makes it physically accessible to circulating drugs, imaging agents, and engineered immune cells. Because GPC3 participates in growth signaling pathways, including interactions with Wnt ligands, it is not merely a passive marker but an active participant in tumor biology. Antibodies, peptide radiopharmaceuticals, and antibody-drug conjugates directed against GPC3 are already in various stages of clinical development, which underscores both the value of the target and the appetite for better binding molecules.

The rationale for pursuing peptides rather than full antibodies is practical as much as biological. Peptides are small, chemically definable, relatively inexpensive to manufacture, and can penetrate tumor tissue more readily than large protein therapeutics. They can serve as homing devices that carry radioactive labels for positron emission tomography imaging, deliver cytotoxic payloads directly to cancer cells, or recruit immune effectors to the tumor site. However, discovering peptides that bind a chosen surface patch of a protein with high affinity and specificity has historically depended on screening enormous random libraries, an approach that is costly and frequently yields candidates with disappointing performance. Recent breakthroughs in computational protein design have opened an alternative route, and the new study is a showcase of that route applied to a clinically important cancer target.

The researchers began with a high-resolution three-dimensional structure of human glypican-3 deposited in the Protein Data Bank under the identifier 7ZAW. Rather than treating the protein as a rigid object, they first allowed it to breathe. A 100-nanosecond molecular dynamics simulation of the unbound, or apo, form of GPC3 was run using the GROMACS engine, capturing the natural fluctuations of the protein in solution. The team then clustered the resulting trajectory and selected the centroid of the most populated cluster, which accounted for 35.6 percent of the stable simulation frames, as a representative conformation for all subsequent design work. This step matters because a protein’s shape in solution can differ subtly from its crystal form, and designing against a realistic, dynamically averaged structure can improve the odds that a designed binder actually fits the real target.

Within that representative structure, the team pinpointed three residues that form the core of the intended binding site: tryptophan 260, tyrosine 264, and methionine 269. These aromatic and sulfur-containing side chains create a distinctive hydrophobic and stacking surface that a complementary peptide could grip. Selecting such coordinate-resolved hotspot residues is a common strategy in structure-guided drug design, because anchoring an interaction at energetically favorable points on a protein surface is far more effective than relying on diffuse, nonspecific contacts across a broad interface.

The generative phase of the pipeline relied on two deep-learning tools that have transformed protein engineering in recent years. RFdiffusion, developed in David Baker’s laboratory, uses a diffusion model akin to those behind modern image generators, but it operates in the space of protein backbone geometry, producing brand-new shapes that never existed in nature. The researchers instructed RFdiffusion to generate one hundred peptide backbones positioned around the identified hotspot surface of GPC3. Each backbone was then handed to ProteinMPNN, a companion neural network that designs amino acid sequences predicted to fold stably into a given backbone. The result was a nonredundant library of peptide sequences, each one a candidate molecule with a computationally specified three-dimensional shape tailored to nestle against the chosen patch of glypican-3.

With thousands of theoretical candidates in hand, the team needed to rank them, and they employed a sequence-level prediction model called PepPAP for the first screening pass. The authors are careful to note that the PepPAP output was treated strictly as a relative computational prediction value, not as an experimentally measured binding free energy, an important methodological caveat that distinguishes predicted affinity from physical reality. The top 120 sequences ranked by PepPAP were then subjected to Rosetta FlexPepDock, a physics-based refinement protocol that models the detailed atomic interface between a peptide and its receptor. FlexPepDock generated 600 receptor-peptide poses across the candidate set, and 565 of those poses satisfied the geometry and contact criteria the team had imposed, indicating that the diffusion-generated backbones were largely compatible with realistic binding geometries.

Binding affinity alone does not make a drug candidate, so the pipeline incorporated two additional filters drawn from different corners of bioinformatics. The top 30 docked peptides were evaluated with Protein-Sol, a web tool that predicts the solubility of a protein or peptide from its sequence, and with ToxinPred 3.0, a machine-learning predictor of peptide toxicity. These screens address two of the most common reasons promising molecules fail downstream: poor behavior in aqueous biological environments and unintended harmful interactions with healthy tissue. Twenty-one of the thirty peptides passed both filters, and an integrated ranking across docking energy, predicted affinity, solubility, and toxicity condensed the field to five final priority candidates.

The energetics of those five finalists are encouraging on paper. Their Rosetta interface energies ranged from −20.459 to −18.255 Rosetta Energy Units, values that reflect substantially favorable computed interactions at the peptide-GPC3 interface. The authors are explicit that these molecules remain computationally prioritized candidates that await experimental validation, and that caution is well placed in a field where previous GPC3-targeting peptides have faltered on close inspection. A recent independent characterization of TJ12P2, a published glypican-3-binding peptide, found a lack of the specificity and potency originally claimed, a sobering reminder that computational or early-stage promise must be confirmed rigorously in the laboratory before a candidate can be taken seriously.

Even so, the significance of the workflow lies in its efficiency and its transparency. The entire study was conducted with publicly available or academically accessible software, including GROMACS, RFdiffusion, ProteinMPNN, PepPAP, Rosetta FlexPepDock, Protein-Sol, and ToxinPred3.0, meaning that other research groups could reproduce the approach against different targets at minimal cost. It consumed no animal subjects, no human materials, and only publicly deposited structural data, which is why the authors report that ethical approval was not required. If experimental binding assays, cellular studies, and ultimately animal imaging experiments confirm that any of the five candidates grips glypican-3 selectively on liver cancer cells, the peptides could be developed into targeted imaging probes, drug-delivery vehicles, or components of theranostic platforms for hepatocellular carcinoma. Even before that, the study adds to a fast-growing body of evidence that diffusion-based generative design, coupled with physics-based refinement and multilayered filtering, can compress years of trial-and-error ligand discovery into a rational, replicable computational procedure, and it offers a template that could be pointed at surface targets in many other cancers in short order.

Subject of Research: Computational de novo design of glypican-3-targeting peptides for hepatocellular carcinoma

Article Title: Structure-based development of Glypican-3-targeting peptide candidates for hepatocellular carcinoma

Article References: Zhang, H., Qiu, J., Cheng, L., Li, C., Xu, J., Qin, Z., & Gao, Y. (2026). Structure-based development of Glypican-3-targeting peptide candidates for hepatocellular carcinoma. 3 Biotech, 16(11), Article 461. https://doi.org/10.1007/s13205-026-05092-4

Image Credits: AI Generated

DOI: 10.1007/s13205-026-05092-4

Keywords: glypican-3, hepatocellular carcinoma, peptide design, RFdiffusion, ProteinMPNN, molecular docking, Rosetta FlexPepDock, molecular dynamics, drug discovery, liver cancer, artificial intelligence, computational biology

The core of the achievement lies in the use of RFdiffusion, a generative deep-learning method that builds novel protein backbones around specified functional motifs. Rather than copying an existing enzyme scaffold, the researchers instructed the diffusion model to create protein structures containing a water-activating zinc-based active site positioned directly adjacent to a peptide substrate binding pocket. The zinc ion, coordinated by catalytic residues in a geometry inspired by natural metalloproteases such as astacin and thermolysin, polarizes a bound water molecule so it can attack the scissile amide bond of a target peptide. This motif-scaffolding approach, refined through multiple rounds of partial diffusion, sequence design and filtering with AlphaFold 3 predictions, allowed the team to generate enzymes in which the catalytic machinery and the substrate recognition surface were designed as an integrated whole.

The experimental results were striking. Of 135 designs tested in the laboratory, 36 percent catalysed cleavage of their cognate peptide substrates at precisely the bond they were designed to target. That success rate is remarkable for any de novo enzyme design campaign, but especially for one targeting amide bond hydrolysis, a reaction class that has historically resisted computational design efforts. The best initial design accelerated peptide bond hydrolysis by more than 100-million-fold compared with the uncatalysed reaction in neutral solution, a benchmark established decades ago by chemists measuring the intrinsic stability of the amide linkage in water.

Even more impressive was what happened next. By screening a library of single-site mutations across 21 positions in and around the active site of one design, and then recombining the most beneficial substitutions, the team identified four changes that boosted catalytic acceleration to more than ten-billion-fold over background. Structural and kinetic analysis indicated that these mutations worked primarily by improving active-site preorganization, holding the catalytic residues and the zinc-bound water in a geometry closer to the transition state of the reaction. In enzymology, preorganization is a key determinant of catalytic power, and the finding that a handful of rationally screened substitutions could push a designed enzyme into the kinetic territory of naturally evolved proteases suggests that the computational pipeline had already captured most of the essential physics of the reaction.

Crystal structures of several designs, deposited in the Protein Data Bank, confirmed that the experimentally determined backbones and active sites matched the computational models with high fidelity. The zinc coordination geometry, the positioning of the substrate binding pocket and the overall fold all closely resembled the design blueprints. Molecular dynamics simulations and machine-learning force-field calculations of the nucleophilic attack step further supported the mechanistic picture, showing that the designed active sites lower the activation energy of the rate-limiting chemical step in a manner analogous to natural astacin-family enzymes.

What elevates this work from an impressive technical demonstration to a potential therapeutic platform is the demonstration that the method can target real human disease proteins. The team designed metalloproteases that selectively cleave TDP-43, the RNA-binding protein whose pathological aggregation is a hallmark of amyotrophic lateral sclerosis and frontotemporal lobar degeneration; amyloid-beta, the peptide central to the amyloid hypothesis of Alzheimer’s disease; and serum amyloid A, the precursor of inflammatory AA amyloidosis. These disease-targeting designs achieved rate enhancements of up to a billion-fold, and specificity experiments showed that each protease preferentially cleaved its intended substrate over the other peptides tested in the panel.

The TDP-43 work illustrates the biological nuance the team brought to the targeting strategy. Full-length degradation of TDP-43 would risk destroying functions essential to the cell, because the protein contains structured N-terminal and RNA-recognition domains needed for its nuclear role. Instead, the designers aimed their protease, called TDPr3, at the intrinsically disordered C-terminal domain that governs liquid-liquid phase separation and aggregation. Cleavage within this region can modulate the phase behaviour and oligomerization of TDP-43 without ablating its essential functions. Experiments with both oligomeric and monomeric forms of full-length human TDP-43, characterized by dynamic light scattering and microscopy, showed that the designed protease progressively cleaved the protein over time, while a catalytically dead knockout variant produced no cleavage products at all.

Beyond degrading disease proteins, the platform opens the door to a powerful new form of biological control: conditional activation of therapeutic molecules. The researchers designed caged cytokines and receptor antagonists that remain inactive until a designed protease clips off a masking sequence, releasing the active biomodulator. Because both the protease and its cleavage site are designed rather than borrowed from nature, the pair can operate bioorthogonally, meaning the activation event does not interfere with, or depend on, any endogenous human protease. Such protease-activatable prodrugs have long been a goal in the field of targeted therapeutics, and the ability to specify both the enzyme and its substrate from first principles makes it possible to build orthogonal activation cascades in which one designed protease unlocks one signal and another unlocks a different one.

The implications extend across therapeutics, biotechnology and bioremediation. Programmable proteases could in principle be deployed to remodel pathological protein deposits, to process biopharmaceuticals during manufacturing, to activate drugs only in diseased tissue, or to break down recalcitrant protein pollutants in the environment. The design code, structural data and experimental protocols have been made publicly available through Zenodo, and provisional patent applications covering aspects of the methods and designs have been filed, signalling serious translational intent. The work was funded by a broad coalition including the Bill and Melinda Gates Foundation, the Defense Threat Reduction Agency, the Howard Hughes Medical Institute, the Audacious Project at the Institute for Protein Design and multiple institutes of the National Institutes of Health.

For the protein design community, the study marks a milestone in a decade-long arc that has taken the field from stable novel folds to functional binders and now to genuinely catalytic enzymes of biomedical relevance. The combination of diffusion-based backbone generation, sequence design, AlphaFold-based filtering and modest experimental optimization proved sufficient to solve a problem, amide bond hydrolysis, that many considered a benchmark of true catalytic sophistication. If the success rate and activity levels reported here can be generalized across arbitrary target sequences, the ability to write a protease for any protein of interest, on demand, would fundamentally change how scientists approach protein-mediated disease, offering a programmable molecular scalpel wherever biology needs one.

Subject of Research: Computational de novo design of zinc-dependent metalloprotease enzymes for targeted protein cleavage

Article Title: Computational design of metalloproteases

Article References: Chen, A., Wu, K., Choi, H., Venkatesh, P., Pellock, S. J., Hanikel, N., shin, H., Evans, D., Didi, K., Schaaf, L. L., Díaz-Perlas, C., Coventry, B., Kim, D., Woodbury, S. M., Ji, P., Honda, S., Bera, A. K., Nguyen, H., Kang, A., … Baker, D. (2026). Computational design of metalloproteases. Nature. https://doi.org/10.1038/s41586-026-11167-2

Image Credits: AI Generated

DOI: 10.1038/s41586-026-11167-2

Keywords: protein design, metalloproteases, RFdiffusion, enzymology, TDP-43, amyloid-beta, serum amyloid A, zinc catalysis, de novo enzymes, protease-activatable prodrugs, David Baker, Institute for Protein Design

T-cell engagers, often called TCEs or bispecific T-cell redirecting molecules, are designed to do exactly what their name suggests: physically bridge a T cell to a tumor cell. One arm of the molecule binds a tumor-associated antigen on the cancer cell surface, while the other binds CD3, a component of the T-cell receptor complex. When both arms engage their targets, the T cell is forced into close contact with the tumor cell, triggering activation, release of cytotoxic granules, and killing of the malignant cell. The approach has produced striking clinical successes in blood cancers, where antibodies can reach malignant cells circulating in the bloodstream. Solid tumors, however, have proven far more stubborn, and gastric cancer is a textbook example of why.

The obstacle is not usually the lack of a good target. CLDN18.2, a member of the claudin family of tight-junction proteins, is expressed on the surface of gastric cancer cells and is largely absent from most adult tissues outside the stomach mucosa, making it one of the most attractive antigens in the field. The problem is that many gastric tumors are immunologically cold—poorly infiltrated by T cells, lacking the chemokine signals that recruit them, and dominated by immunosuppressive stromal features. A T-cell engager can only redirect T cells that are actually present in or near the tumor. If the tumor bed contains few CD8-positive cytotoxic T cells, even a perfectly designed bridging molecule has little to work with.

The design of the new molecule began in silico. Rather than starting from an existing antibody, the team used an RFdiffusion-assisted workflow—a computational approach that generates novel protein structures by a diffusion-based process, effectively imagining new binding surfaces that can grip a chosen target. Candidate CLDN18.2 binders were screened computationally, and the most promising designs were incorporated into both human and murine CD3-engaging T-cell engager formats, allowing the researchers to test the concept in parallel across human-relevant and mouse-model systems. This design philosophy is part of a broader shift in protein engineering: instead of repurposing natural scaffolds, scientists can now specify a desired binding interaction and let generative models propose molecular architectures that satisfy it.

The selectivity testing focused on a critical safety concern. Claudin proteins are close relatives, and a therapeutic that binds CLDN18.1 as well as CLDN18.2 could damage healthy lung epithelium, where CLDN18.1 is found. In cell-surface binding assays, the selected engager bound CLDN18.2-positive tumor cells and CD3-positive T cells while showing limited detectable binding to cells expressing CLDN18.1. Functional testing followed in increasingly realistic settings: engineered gastric cancer cell lines, patient-derived tumor cells, and patient-derived organoids—three-dimensional cultures grown from actual patient tissue that preserve much of the cellular complexity of the original tumor. Across all of these platforms, the molecule induced CLDN18.2-dependent tumor-cell killing and activated CD8-positive T cells, confirming that the computationally designed binder could perform its biological job, not just bind its target in a test tube.

Then came the in vivo experiments, and with them the study’s central twist. In both syngeneic mouse models—immunocompetent animals with intact immune systems—and humanized models created by reconstituting mice with human peripheral blood mononuclear cells and carrying patient-derived tumor xenografts, the T-cell engager produced early tumor inhibition. But as the experiments progressed, tumor control weakened. The engager was bridging T cells to tumor cells, yet the effect was not sustaining itself. The team examined the tumors and found a consistent pattern: limited accumulation of T cells inside the tumor, increased expression of PD-1, the canonical marker of T-cell exhaustion, sparse populations of CD11c-positive dendritic cells, and weak signaling through CXCL9 and CXCL10, chemokines that normally act as beacons drawing T cells into tissue.

That molecular fingerprint pointed directly at the missing ingredient. Dendritic cells are the professional sentinels and coordinators of the immune system; they present antigen to T cells, secrete inflammatory cytokines such as interleukin-12, and produce the very chemokines—CXCL9 and CXCL10 among them—that recruit effector T cells into tumors. The researchers hypothesized that supplementing the T-cell engager treatment with additional dendritic cells might rebuild the local immune circuit that the cold tumor environment lacked. The experiment confirmed the hypothesis: adding dendritic cells improved TCE-mediated tumor-cell killing in co-culture systems and strengthened tumor control in vivo, converting a waning response into a more durable one.

To understand exactly how the dendritic cells were helping, the team ran mechanistic experiments using blocking antibodies. When they blocked MHC-II, the molecule dendritic cells use to present antigen to helper T cells, the enhancement was only modestly reduced. When they blocked interleukin-12, however, the dendritic-cell-associated boost was consistently diminished. The interpretation is significant for the field: the added dendritic cells were not primarily acting as antigen-presenting machines in this setting, but as cytokine factories, providing the inflammatory support—IL-12 and the downstream CXCL9/CXCL10 chemokine axis—that keeps redirected T cells active, recruited, and functional inside the tumor.

The implications extend well beyond gastric cancer. The study demonstrates that generative protein design tools like RFdiffusion can produce functional, selective therapeutic binders against clinically validated targets, compressing a discovery process that once took years of antibody engineering. Just as importantly, it offers a candid lesson about the limits of that technology: a beautifully designed molecule that bridges T cells to tumor cells cannot, by itself, overcome a tumor microenvironment that lacks the supporting immune infrastructure. The authors’ conclusion—that CLDN18.2-directed T-cell engagers should be tested together with strategies that rebuild local dendritic-cell–T-cell circuits—frames a design principle likely to shape the next generation of solid-tumor immunotherapy combinations.

Caveats remain, as they do in any preclinical translational study. The work rests on engineered cell lines, patient-derived cultures, organoids, and mouse models rather than human clinical trials, and the published version is an early-release, peer-reviewed accepted manuscript that may undergo further editorial refinement. The dendritic-cell supplementation strategy, while mechanistically compelling, would require substantial development before clinical application—questions of cell sourcing, manufacturing, timing, and delivery all remain open. Still, the study offers a clear and testable roadmap: pair AI-designed targeting molecules with interventions that restore the chemokine and cytokine signals cold tumors fail to produce. For patients with CLDN18.2-positive gastric cancer, a disease with limited options in advanced stages, that roadmap represents a rational path from computational design to combination therapy, one that the field will be watching closely as these strategies move toward clinical evaluation.

Subject of Research: RFdiffusion-designed CLDN18.2 T-cell engager combined with dendritic-cell supplementation for gastric cancer immunotherapy

Article Title: RFdiffusion-designed CLDN18.2 T-cell engager coupled with dendritic-cell supplementation enhances antitumor immunity in gastric cancer

Article References: Kan, H., Niu, G., Wan, Z., Zou, Y., Wang, J., Xu, Y., Yang, H., Xu, Z., & Li, B. (2026). RFdiffusion-designed CLDN18.2 T-cell engager coupled with dendritic-cell supplementation enhances antitumor immunity in gastric cancer. Journal of Translational Medicine. https://doi.org/10.1186/s12967-026-08841-4

Image Credits: AI Generated

DOI: 10.1186/s12967-026-08841-4

Keywords: RFdiffusion, T-cell engager, CLDN18.2, gastric cancer, dendritic cells, IL-12, CXCL9, CXCL10, protein design, immunotherapy, organoids, patient-derived xenografts

The competition’s premise deliberately raised the bar beyond standard binder benchmarks. A CAR binding domain must do far more than dock onto its antigen: it must be synthesized as DNA, expressed on the T cell surface within a fixed receptor scaffold, and trigger signaling that is strong enough to activate the cell but not so strong that the T cell becomes exhausted. Teams were told their 80-amino-acid binders would replace the single-chain variable fragment, or scFv, of a second-generation CAR28z construct, a domain that in natural antibodies typically runs 230 to 260 amino acids. Success ultimately depended on CAR-T cell function, not just molecular stickiness. Organizers provided primers on CAR-T biology, the exact scaffold, and assay details, then let competitors loose with whatever AI-driven workflows they preferred, creating an unusually diverse and honest snapshot of how the field actually designs proteins today.

The methodological landscape that emerged was dominated by generative diffusion models. Eighteen teams used RFdiffusion, two used sequence-structure co-design tools such as Chroma, and others turned to hallucination-based approaches like BindCraft, EvoBind, and ColabDesign, or protein language models including PepMLM, ESM-2, and RayGun. Nearly twenty teams employed ProteinMPNN or its soluble variant to design sequences onto fixed backbones, and seventeen used AlphaFold2 to verify that their creations were predicted to fold as intended. Twelve teams essentially followed the default RFdiffusion-ProteinMPNN-AlphaFold2 pipeline, while nine practiced iterative diffusion, trimming low-scoring segments and regenerating until scores improved. The result was extraordinary sequence diversity: more than 83 percent of submissions showed no significant alignment to any sequence from another team, although within-team redundancy was high, with a fifth of sequences nearly identical to a teammate’s design.

Testing 12,000 designs individually would have been impractical, so the organizers built a pooled screening system with elegant internal bookkeeping. Each codon-optimized binder sequence was cloned into the CAR plasmid library, where the binder itself served as a unique DNA barcode. Using the GeneWeld targeted integration method, the library was knocked into the T cell receptor alpha chain locus of primary human T cells, generating a pool of TRAC-negative, CAR-positive cells. After methotrexate selection using a co-expressed DHFR variant, the cells were split: one arm cultured alone, the other challenged repeatedly with mitotically inactivated CD20-positive Raji tumor cells. Over two weeks, binders that enabled productive antigen recognition drove proliferation of their host cells, and next-generation sequencing of the barcode region revealed which designs expanded. Comparing tumor-challenged cultures against the no-target controls allowed antigen-dependent growth to be quantified robustly for every design simultaneously.

The screen produced a sobering attrition curve. Nearly 98 percent of designs passed DNA synthesis quality control, but only 56.8 percent were recovered at sufficient sequencing depth after growth, implying that many binders failed to yield a viable CAR, likely because of poor expression, folding, or membrane localization. Ultimately, 707 designs, just under 6 percent, showed significant CD20-specific enrichment of at least twofold. Hit rates varied dramatically between teams, from 0.6 percent to 38.4 percent, a spread that itself encodes lessons about which strategies work. The ten best non-redundant designs advanced to individual validation, where each was rebuilt as a standalone CAR construct and subjected to a full battery of functional assays against CD20-positive Raji cells and CD20-negative K562 controls.

Seven of the ten leading designs outperformed untransfected control T cells in proliferation, expansion, and cytokine production, with flow cytometry confirming CAR expression in most of them. Designs 1506, 5300, and 5981 were the strongest cytokine producers, generating as much interleukin-2 and interferon-gamma as the scFv positive control, and design 5300 exceeded it. When it came to the most demanding readout, specific killing of CD20-positive targets, only four designs, 1506, 2383, 3494, and 3718, achieved significant target-specific lysis above background, and two of those matched the scFv control in specificity. Notably, several broadly functional designs showed no measurable binding in isolation, suggesting they may have relied on avidity or multivalent mechanisms within the receptor context. The pooled hit rate of 707 thus appears to be an upper bound on true functional success.

Surface plasmon resonance against detergent-solubilized CD20 added a crucial biophysical dimension. The scFv control bound with a dissociation constant of 1.75 nanomolar, but only three of the top ten designs showed appreciable binding: design 2383 at 643 nanomolar, and designs 1506 and 5981 with weak but detectable signals between roughly 200 and 600 nanomolar. The authors caution that this apparent preference for moderate or weak affinity is consistent with CAR biology, where excessively tight binding can drive trogocytosis and T cell exhaustion, and lower-affinity receptors have sometimes eliminated tumors better. However, the binding assay tested isolated 80-mers produced by bacterial cell-free synthesis rather than full receptors in mammalian cells, so some failures may reflect structural context rather than design quality. The team recommends that future efforts fold candidate binders within the complete CAR sequence, a computationally trivial step that only one team actually performed, and that team produced the top-performing design.

The most consequential findings came from mining the entire dataset for predictors of success and failure. The organizers computed more than 400 features per sequence, from GC content and Shannon entropy to predicted secondary structure and interface energies. Strikingly, 98.9 percent of designs that vanished from the proliferation screen had been created with ProteinMPNN or SolubleMPNN, and the failed designs were almost pure alpha helices, averaging 83 percent helical content, densely packed with lysine and glutamate. The ratio of residues in lysine-plus-glutamate helices was by far the strongest predictor of recovery failure, reaching 0.91 ROC-AUC in a cross-validated logistic model, averaging 35 percent of residues in failed designs versus 15 percent in recovered ones. The likely mechanism is translational interference: glutamate repeats are known to stall ribosomes, and lysine and glutamate codons produce adenosine-rich DNA, with poly-A tracks also disrupting translation. Failed sequences carried nearly four glutamate-glutamate repeats on average versus 1.6 in recovered designs, and adenosine dinucleotide repeats predicted recovery with 0.88 ROC-AUC. An independent phage display study has since confirmed the lysine-glutamate association, suggesting the signal generalizes well beyond CAR-T cells.

Other failure modes emerged with equal clarity. Designs that caused CD20-specific depletion of T cells, possibly reflecting exhaustion, were enriched in cysteines predicted to form intramolecular disulfide bonds, which may have exposed hinge sulfhydryls and amplified tonic signaling. DNA synthesis failures clustered among sequences with high GC content and low entropy. Perhaps most humbling for the structure-prediction community, the standard confidence metrics, including ipTM, pLDDT, ipSAE, LIS, SAP, PDockQ, and language-model likelihoods, showed little predictive power for functional outcomes, and most teams’ own confidence rankings correlated poorly, or even negatively, with experimental success. Simple sequence statistics beat sophisticated structure-based scores. Applying retrospective filters for GC content, entropy, adenosine repeats, glutamate repeats, and lysine-glutamate helix content removed 4,600 sequences and nearly doubled the proportion of designs showing CD20-specific proliferation, from 5.9 to 7.6 percent, rising to 10.6 percent with a cysteine filter that would also have eliminated two of the non-functional top-ten designs.

The authors are careful about the study’s limits: a single CAR scaffold, a single donor isolate, a comparison of Raji-challenged cells against no-target rather than CD20-negative controls, and a design space constrained to 80 amino acids all shape the conclusions. No design is remotely clinic-ready, and all underperformed the scFv control in most readouts. Yet the significance of the work lies less in any single binder than in the demonstration that generative protein design can now be evaluated end-to-end in a therapeutic context, at scale, in the open. The dataset has been released on Zenodo with code on GitHub, and the practical recommendations, avoid lysine-glutamate helix overload, watch for cysteines, fold within the full receptor, and treat moderate affinity as a feature rather than a bug, offer the field its first empirically grounded playbook for designing proteins that do not merely bind, but function inside living cells.

Subject of Research: Experimental validation of AI-designed de novo protein binders as CAR-T cell recognition domains targeting CD20

Article Title: Validation and analysis of 12,000 AI-driven CAR-T designs in the Bits to Binders competition

Article References: Kosonocky, C. W., Abel, A. M., Feller, A. L., Cifuentes Rieffer, A. E., Woolley, P. R., Lála, J., Barth, D. R., Gardner, T., Bits to Binders Competitors, Acosta, D., Angioletti-Uberti, S., Annapure, V., Anthony, N., Barghout, R. A., Beining, M., Bibekar, P., Biton, D., Bryant, P., Bushuiev, A., … Marcotte, E. M. (2026). Validation and analysis of 12,000 AI-driven CAR-T designs in the Bits to Binders competition. Molecular Systems Biology. https://doi.org/10.1038/s44320-026-00246-1

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00246-1

Keywords: CAR-T cells, protein design, artificial intelligence, CD20, ProteinMPNN, RFdiffusion, de novo binders, pooled screening, protein engineering, immunotherapy, translational stalling, benchmarking

The research, summarized in a companion Research Briefing, combines AI-assisted de novo binder design with the assembly of complete CAR constructs and scalable testing platforms in vitro and in vivo. Rather than asking whether a computationally designed protein can bind a target antigen in a dish, the investigators pushed the designs through the full developmental gauntlet that a therapeutic candidate must survive: expression on the surface of primary T cells, signal transduction upon antigen engagement, target cell killing, and durable anti-tumor activity in living models. The resulting dataset links sequence-level and structural-level features of designed binders directly to therapeutic function, providing a design grammar that the field has previously lacked.

The work builds on a remarkable run of advances in computational protein design. In 2023, the introduction of the RFdiffusion framework dramatically raised the success rate of de novo binder campaigns by using diffusion models to generate protein backbones tailored to a desired target surface. A year later, researchers reported targeting overexpressed antigens in glioblastoma using CAR T cells armed with computationally designed high-affinity protein binders, offering early proof that AI-designed recognition domains could function as the business end of a chimeric antigen receptor. More recently, the hallucination-based design pipeline BindCraft described one-shot generation of functional protein binders, and it produced some of the highest-success design campaigns in the new study. Together with ProteinMPNN, a robust deep learning method for assigning amino acid sequences to designed backbones, these tools have made binder generation almost routine. The new work addresses the harder question: which of the many binders that pass computational filters will actually drive T cells to kill cancer?

That question matters because binding is only the beginning. A CAR binding domain operates in a demanding mechanical and biological context. It must fold correctly and traffic to the cell membrane when fused to hinge, spacer, transmembrane and signaling modules. It must bind antigen with an affinity that falls within a productive range; too little affinity produces no signal, while excessive affinity can cause antigen-independent tonic signaling, activation-induced cell death and poor persistence. It must tolerate epitope densities that vary enormously across tumor tissues. And it must be small, stable and non-immunogenic enough to be clinically deployable. Natural antibody fragments often fail several of these criteria simultaneously, which is why the pharmaceutical industry has invested heavily in screening campaigns that yield a single usable binder after months of labor.

By systematically varying binder sequences and structures within the CAR context and evaluating the resulting constructs in standardized cell models, the authors were able to infer the attributes that correlate with efficacy. The study frames a complete framework for designing efficacious CARs, in which binder attributes such as affinity, epitope choice, stability and expression behavior are treated as tunable design parameters rather than accidents of discovery. The significance of this reframing is difficult to overstate. For two decades, CAR engineering has been as much an art as a science, with laboratories borrowing fragments from existing antibodies and adjusting hinges and spacers empirically. A predictive model of what makes a binding domain efficacious turns CAR design into an engineering discipline in which candidate receptors can be specified computationally before a single experiment is run.

The experimental architecture underpinning the study is as important as its findings. The investigators paired binder design with scalable in vitro assays that measure how many of the designed constructs express on T cells, how strongly they signal, and how effectively they eliminate antigen-positive targets. In vivo models then tested whether promising designs retained activity against tumors in a physiological setting, where antigen density, immune suppression and trafficking barriers conspire to defeat otherwise potent receptors. This multi-tier funnel mirrors the path of a therapeutic candidate and ensures that the design rules extracted from the data reflect true clinical relevant properties, not merely binding measurements from immobilized proteins.

The implications extend well beyond one cancer type. Because the pipeline is generative, it is in principle antigen-agnostic: given a target surface, the same design-and-test cycle can produce panels of candidate binding domains against antigens relevant to solid tumors, autoimmune disease, fibrosis and infectious disease. Solid tumors have proven especially refractory to CAR therapy, in part because widely shared tumor-associated antigens are also expressed on essential healthy tissues and because single-antigen targeting invites escape. AI-designed binders, selected with precise affinity windows and epitope specificity, could enable new strategies such as affinity tuning to discriminate between high- and low-expressing tissues, dual-antigen logic gating, and rapidly generated panels against patient-specific neoantigens. The glioblastoma work from 2024 demonstrated that computationally designed binders could target an antigen overexpressed in one of the deadliest solid tumors; the new study supplies the general principles for making such binders reliably efficacious.

There are also cautionary notes that seasoned observers of the CAR field will appreciate. Computational design success rates, even with state-of-the-art tools, remain probabilistic, and the attributes that make a binder effective in a standardized cell line may not transfer directly to the hostile microenvironment of a human tumor. Immunogenicity of non-human-derived protein scaffolds must be assessed rigorously before clinical translation, and the regulatory pathway for wholly synthetic recognition domains is still being defined. Tumor heterogeneity, antigen loss and the immunosuppressive microenvironment remain problems that no binder, however well designed, can solve alone. What the study offers is not a finished therapy but a reproducible methodology for generating and selecting binding domains with predictable properties, which removes one of the largest sources of variability and failure in current CAR programs.

The broader scientific community has taken notice of how quickly the ingredients of this advance came together. RFdiffusion and ProteinMPNN provided the generative backbone and sequence design machinery; AlphaFold-style structure prediction supplied reliable in silico validation of designed conformations; BindCraft demonstrated that hallucination-based pipelines could deliver functional binders in single campaigns; and the earlier glioblastoma CAR study established clinical feasibility. The new research closes the loop by asking what distinguishes the binders that work in a CAR from those that bind beautifully on paper but fail on the cell surface. The answer, encoded in the sequence and structural determinants the authors report, is a practical toolkit for the next generation of synthetic immunology.

If the field can standardize on these design attributes, the consequences could be transformative. Cell therapy developers could move from years of empirical binder discovery to weeks of computational specification followed by targeted validation. Clinicians could obtain CAR constructs tuned precisely to the antigen expression profile of an individual tumor. Academic laboratories with modest resources could design receptors against orphan antigens that no commercial entity would ever fund an antibody campaign for. The convergence of generative AI and cellular immunotherapy has promised exactly this kind of acceleration for several years, and this study provides some of the clearest evidence yet that the promise is becoming an operational reality, one amino acid at a time.

Subject of Research: AI-assisted de novo design of protein binders for constructing efficacious chimeric antigen receptor T cell therapies

Article Title: Defining attributes of effective binders for AI-assisted CAR design

Article References: Defining attributes of effective binders for AI-assisted CAR design. (2026). Nature Biomedical Engineering. https://doi.org/10.1038/s41551-026-01792-7 (Author names are not included in the publisher’s citation export.)

Image Credits: AI Generated

DOI: 10.1038/s41551-026-01792-7

Keywords: CAR T cells, chimeric antigen receptor, AI protein design, de novo binders, RFdiffusion, ProteinMPNN, BindCraft, cancer immunotherapy, protein engineering, Nature Biomedical Engineering, antigen targeting, synthetic immunology