Ubi Titer Issue #10
Conditional Activation: Gated Cytokines, Hapten Adaptors, and Prodrug Degraders
Six papers converge on making activity conditional rather than making molecules stronger: cytokine signaling gated on checkpoint binding, payload release gated on receptor trafficking, T-cell engagement gated on a swappable chemical tag, and degrader function gated on an enzymatic modification.
- conditional activation
- bispecific antibody-drug conjugates
- VHH nanobody engineering
- T-cell engagers
- targeted protein degradation
- molecular glue degraders
- FcRn recycling
- de novo binder design
- carbohydrate binders
The field
This Week in Biologics
Paper 1 · mAbs
JMB2403, a potential best-in-class PD-1-dependent IL2Rβγ-activating tri-specific antibody for safe and potent immunotherapy
Deliberately weak IL-2Rβ/γ agonist nanobodies fused to an anti-PD-1 IgG produce cytokine signaling that switches on only where PD-1 is engaged, avoiding the toxicity that has limited IL-2 therapy.
Core finding
The molecule pairs a bivalent anti-PD-1 IgG with paired single-domain antibodies against IL-2Rβ and IL-2Rγ, selected from a naïve alpaca library specifically for weak agonism. It binds the IL-2Rβγ complex roughly 30-fold more weakly than wild-type IL-2 and does not bind CD25 at all, yet drives STAT5 phosphorylation with an EC50 of 0.2 nM on activated CD8 T cells and minimal signal on resting T cells or regulatory T cells. Pre-blocking PD-1 with the parental antibody reduced signaling approximately 42-fold, establishing that the cytokine activity is genuinely gated on target-cell PD-1 rather than merely concentrated near it.
What is novel
Most attempts to fix IL-2 toxicity engineer the cytokine itself into an attenuated mutein. This inverts the approach: start from intrinsically weak receptor agonists and restore potency through avidity supplied by the targeting arm. The single-domain format is mechanistically necessary rather than convenient, since two 15 kDa domains without light chains can bridge IL-2Rβ and IL-2Rγ into signaling proximity in a way conventional Fab arms are too bulky to achieve. Because IL-2 and IL-15 share the βγ receptor, the design also captures IL-15 biology without requiring IL-15Rα presentation.
Limitations
The cynomolgus study used three animals, one per dose level, under non-GLP conditions, which supports a tolerability signal but cannot establish dose-response. All three developed anti-drug antibodies before the second dose, and the reduced T and NK proliferation seen afterwards cannot be cleanly separated between immunogenicity and receptor desensitization at this sample size. Efficacy models are PBMC-humanized xenografts rather than fully immunocompetent systems.
Why it matters in context
Three strategies currently dominate efforts to make IL-2 tolerable: muteins with abolished CD25 binding, partial agonists biased toward effector cells, and checkpoint-targeted IL-2 fusions. All retain constitutive systemic activity to some degree, and even PEGylated prodrug approaches designed for biased signaling produced capillary leak in clinical testing. The benchmark used throughout this study, a published PD-1/IL-2 fusion, produced no complete responses at equimolar dosing and caused animal deaths, while the nanobody construct achieved complete regression in seven of eight animals. The distinguishing claim is spatial control as a selectivity mechanism separate from affinity tuning.
Paper 2 · Molecular Cancer Therapeutics
Dual Targeting of EGFR and HER2 by a Bispecific Antibody-Drug Conjugate Demonstrates Broad Antitumor Activity in Multiple Solid Tumors
An EGFR×HER2 bispecific ADC outperformed DS-8201 across 22 patient-derived xenografts, with the advantage traced to receptor trafficking rather than broader target coverage.
Core finding
The rationale rests on where receptors go after binding rather than how many are engaged. Antibody-bound HER2 is largely recycled back to the cell surface, while antibody-bound EGFR enters clathrin-dependent endocytosis efficiently, and EGFR-HER2 heterodimers restrict internalization of HER2-directed conjugates. Engaging both receptors simultaneously redirects the complex toward the lysosome, where linker cleavage and payload release occur. The conjugate showed higher maximum binding and more efficient internalization than monospecific controls across every cell line tested regardless of expression pattern, with lysosomal colocalization confirmed, and achieved a 68.2% response rate across 22 patient-derived xenograft models against 27.3% for the benchmark.
What is novel
The linker chemistry contributes as much as the bispecific format. A hydrophilic cathepsin B-cleavable peptide linker produced a homogeneous drug-to-antibody ratio of 5.8 and released under 0.1% free payload after 21 days in human, cynomolgus and mouse plasma, a marked stability improvement over the benchmark conjugate. That stability shows up directly in biodistribution, with free payload exposure in tumor tissue roughly 40-fold higher than in plasma. The payload itself, an exatecan derivative, is about four-fold more potent than the comparator payload, and the conjugate also triggered both measured hallmarks of immunogenic cell death.
Limitations
All data are preclinical. The authors acknowledge that these models cannot capture human immune contributions, leaving the immunogenic cell death findings unvalidated clinically. No primate pharmacokinetic or toxicology data are reported, and interspecies differences in conjugate clearance limit extrapolation to human dosing. Because the antibody does not bind rodent EGFR or HER2, the mouse studies cannot assess on-target toxicity, and the broad expression of EGFR in normal epithelium is addressed through argument and precedent rather than direct safety data.
Why it matters in context
HER2-directed conjugates have transformed treatment in both HER2-positive and HER2-low disease but remain limited by low target expression, tumor heterogeneity and acquired resistance. An earlier EGFR×HER2 bispecific conjugate used random conjugation at a lower drug-to-antibody ratio with a microtubule-inhibitor payload, and this work argues that topoisomerase I payloads deliver stronger bystander killing, lower multidrug-resistance susceptibility and a wider therapeutic window. Clinical proof-of-concept for dual targeting within this receptor family comes from an EGFR×HER3 bispecific conjugate that reported a 34% overall response rate in phase 1, with adverse events characteristic of topoisomerase I payloads that proved manageable with standard supportive care.
Paper 3 · Antibody Therapeutics
DOTAM-TCB: A Universal Small Molecule-Guided Hapten- and T Cell-Bispecific Antibodies for Cancer Immunotherapy
A single T-cell engager recognizes a chemical hapten rather than a tumor antigen, so swapping the attached small-molecule ligand redirects it to a new target without re-engineering the antibody.
Core finding
The platform separates tumor targeting from effector engagement entirely. A universal T-cell bispecific antibody binds a calcium-loaded DOTAM cage with equilibrium dissociation constants of 51 to 263 femtomolar, and that cage is conjugated to established small-molecule tumor ligands: a PSMA ligand, acetazolamide for carbonic anhydrase IX, and folate for folate receptor 1. Combining adaptor with antibody produced target-restricted T-cell activation and killing with kinetics comparable to conventional engagers built against the same antigens, while either component alone was inactive. In an ex vivo assay on tumor tissue grown in mice and co-cultured with donor immune cells, the adaptor system exceeded both direct engager controls across nearly all activation markers in both donors.
What is novel
The deliberate shift from antibody-based to small-molecule adaptors was made for pharmacokinetic reasons. The group's earlier platform used antibody adaptors circulating for roughly three weeks, which prolongs on-target toxicity when problems arise. Small molecules distribute faster and clear within hours, turning the dosing schedule into a control mechanism: administer the inert antibody first and use the adaptor as a rapid on-switch, withhold it as an off-switch during adverse events, or pretarget tumor with adaptor before dosing the antibody. This is a structurally different answer to cytokine release syndrome than reducing CD3 binding affinity. A 1.45 Å crystal structure shows the hapten cage buried between the heavy and light chain variable domains, coordinated through its central calcium, explaining the exceptional affinity.
Limitations
No in vivo efficacy data are presented, and the authors state that in vivo work requires optimizing dosing for both components. One of the three adaptors failed outright, showing no binding to cells expressing its target, attributed to suboptimal linker chemistry rather than the platform concept. Antibody format effects remain unexplained, since two formats with identical valency showed different activity profiles. Because small-molecule ligands typically cross-react across species and the chosen carbonic anhydrase ligand binds related enzymes present in kidney, humanized mouse models may report toxicity that does not reflect human biology.
Why it matters in context
Adaptor strategies originated in cell therapy to address antigen escape and the cost of personalized manufacturing, using universal effector cells that bind interchangeable targeting molecules. The lineage runs through fluorescein-based chimeric receptor systems and small molecules that recruit endogenous antibodies to tumor surfaces. Until now, small-molecule tumor ligands have been confined to redirecting engineered cells, recruiting endogenous antibodies, or conjugation to antibody-fragment adaptors, leaving the pairing of engineered antibody effectors with purely small-molecule adaptors largely unexplored. The measured hapten affinity here is orders of magnitude tighter than previously reported anti-hapten binders, which is what makes a stable immune synapse plausible.
Paper 4 · Nature
DCAF11-dependent molecular glue degrader activated by glutathionylation
A screen across seven E3 ligases found a compound that only becomes a molecular glue after cells chemically modify it, revealing a new activation mechanism and a general handle on an underused ligase.
Core finding
An unbiased screening platform immobilized seven E3 ligases simultaneously, exposed them to pooled compounds in whole-cell lysate, and read out induced protein recruitment by mass spectrometry, requiring no prior knowledge of ligase-substrate pairs. Screening 5,000 compounds identified M12, which recruits the RNA helicase DDX18 to the ligase DCAF11 for ubiquitin-mediated degradation. The mechanism proved unusual: purified DCAF11 and DDX18 form no complex with M12 alone, but do after exposure to cell lysate, and boiled lysate fails, establishing that an enzyme is required. Glutathione S-transferase conjugates glutathione onto a reactive chlorine on M12, and only this modified form functions as a glue. Removing or altering that chlorine abolished activity entirely.
What is novel
This is a prodrug molecular glue whose activation depends on ordinary cellular metabolism rather than direct binding chemistry. A cryo-electron microscopy structure at 2.3 Å resolution shows why it works: the glutathione moiety anchors in the central pore of the ligase's propeller domain, the site normally used for recognizing peptide substrates, while the rest of the compound extends outward to remodel the surface and recruit the helicase, burying an interface of 870 square angstroms. The glutathione-binding site is highly conserved and the ligase binds glutathione on its own without any drug, indicating that the compound hijacks an existing physiological sensing function. Treating that warhead as a modular handle, the authors then built four degraders against unrelated targets, covering a bromodomain protein, a chromatin remodeller, a second bromodomain target and a set of kinases, all dependent on the same ligase.
Limitations
The unbiased screen produced one natural glue pairing; the further targets were reached by deliberately linking the warhead to known ligands as PROTACs, which is a different and easier exercise than discovering a glue. Activity depends on cellular glutathione and the transferase that installs it, and while the authors argue this could widen the therapeutic window because many tumours sustain high glutathione while proliferating, they do not measure how potency actually tracks with that status, so the argument cuts both ways for now. Only part of the recruited helicase was resolvable in the structure. The platform covered seven ligases out of roughly six hundred, and the suggestion that comparable mechanisms exist for others is explicitly speculative. No animal efficacy or pharmacokinetic data are reported.
Why it matters in context
Molecular glue degraders trace to thalidomide analogues acting through a single ligase, whose clinical success in blood cancers established the modality. Subsequent work has added a handful more usable ligases through rational screening and cytotoxicity profiling, but only a small subset of human E3 ligases have been successfully co-opted, and the field's stated bottleneck is the absence of scalable discovery methods that do not depend on ligase-specific compound libraries. Several groups have reached this particular ligase through covalent electrophiles and related reactive chemistry; the route described here is mechanistically distinct because cellular enzymes modify the compound rather than the compound reacting directly with the ligase.
Paper 5 · Journal of Pharmaceutical Sciences
Mechanistic modeling of time-dependent antibody clearance in multiple myeloma: A physiologically based pharmacokinetic model of isatuximab
Antibody clearance in myeloma falls by half during treatment because tumor-secreted protein competes for the receptor that recycles antibodies, meaning exposure follows response rather than driving it.
Core finding
Myeloma cells secrete an antibody-like paraprotein at concentrations well above normal circulating immunoglobulin, and that protein competes with therapeutic antibody for the neonatal Fc receptor that rescues antibodies from degradation. As treatment reduces tumor burden the competing protein falls, competition eases, and drug clearance drops. A whole-body model in which drug, endogenous immunoglobulin and paraprotein all compete for endosomal receptor binding, with only five fitted parameters, reproduced the clearance decline reported in published population analyses in both magnitude and timing. Extending the model across response scenarios predicted a 75% clearance reduction and 3.77-fold higher steady-state exposure in complete responders compared with patients whose paraprotein does not fall.
What is novel
This is the first mechanistic physiological model to attribute time-varying antibody clearance to a competing endogenous immunoglobulin, rather than treating the decline as an empirical covariate. Two fitted values are independently informative: the paraprotein binds the recycling receptor roughly twelve-fold more weakly than normal immunoglobulin, consistent with its known folding and modification defects, and the model predicts clearance in patients whose disease secretes a non-competing immunoglobulin subtype closely matching an independent published estimate for a population it was never fitted to. The model was also used to answer a practical question, showing that immunoglobulin replacement therapy costs roughly 29% of drug exposure, likely not enough to warrant dose adjustment.
Limitations
Only average summary data are publicly available, so the model was fitted to digitized curves representing a typical patient rather than individual records. Parameter uncertainty is therefore likely understated, and variability between patients in baseline paraprotein, depth of response and pharmacokinetics is not captured. Longitudinal albumin data were not published for this drug, so a simultaneous contribution from improving cachexia can neither be excluded nor quantified. The model describes how disease affects drug exposure but not the reverse direction.
Why it matters in context
Antibody clearance commonly interacts with disease status, whether through elevated catabolism in advanced disease or through target-mediated elimination, and in both cases responders show falling clearance and rising exposure over time. This creates a well-recognized trap for exposure-response analysis, where the relationship is read as exposure driving response when causality partly runs the other way, and the authors cite a large trastuzumab trial as an example of the costly studies such confounded findings can prompt, its higher-dose arm having produced no benefit. Independent population analyses support the mechanism proposed here, having found the highest clearance in patients whose disease secretes the competing immunoglobulin subtype, correlation with a tumor burden marker, and steeper clearance decline in responders.
Paper 6 · NVIDIA (with ICLR 2026 companion paper)
Latent Generative Search unlocks de novo Design of Untapped Biomolecular Interactions at Scale (Proteína-Complexa)
Pairing a generative model with reward-guided search at inference time produced experimentally validated binders across 127 targets, including picomolar PDGFR binders and the first de novo proteins that bind a free carbohydrate.
Core finding
Binder design has split into generative methods, which sample from models trained on binder-target complexes, and hallucination methods, which skip the generator and optimize sequences against structure-prediction confidence. This work runs both: a flow-matching model that jointly generates sequence and structure in a continuous latent space, then search strategies including beam search and Monte Carlo tree search steering that generation using folding confidence, force-field energy and geometric rewards. The claim was tested at unusual scale. Over one million designed proteins were screened against 127 targets by multiplexed phage display, yielding at least one on-design hit for 86 targets, and in that head-to-head the method's own co-generated sequences reached a 2.45% on-design specific hit rate against 0.76% for the next-best self-generated baseline and 1.81% for the strongest inverse-folding-redesigned baseline.
What is novel
The most consequential experimental finding is that co-generated sequences beat post hoc redesign. Every prior method, generative or hallucination-based, depends on a separate inverse-folding model such as ProteinMPNN to make its backbones expressible; here the model's own sequences (691 on-design hits) outperformed ProteinMPNN redesign of the very same backbones (365 hits), which the authors present as the first large-scale in vitro evidence that end-to-end codesign can remove that step. Five further campaigns extended the range: a 63.5% hit rate against the polar target PDGFR with the best binder at 93.6 pM, nanomolar binders to the muscle-wasting receptor ActRIIA that blocked myostatin signaling in cells with no affinity maturation, 40-50% hit rates for kinase mini-protein and short peptide binders, nanomolar Nipah virus binders from both de novo design and codesign-based rescaffolding of an existing binder, and five binders from 24 designs against a blood group B trisaccharide, which the authors report as the first computationally designed proteins that bind a free carbohydrate.
Limitations
Hit rates vary sharply by target and by the interface hotspot residues chosen to condition generation, and the computationally top-ranked hotspot does not reliably give the best experimental result, so multiple hotspot combinations must be tested per target rather than trusting a single prediction. Off-design binding was pervasive, with 126 of 127 targets picking up at least one unintended binder, meaning specificity still requires downstream selection rather than emerging from design. The authors note the margin of codesign over sequence redesign varies across targets and that which epitope properties benefit most from joint optimization remains unresolved. Structure-prediction confidence metrics correlated only imperfectly with measured affinity.
Why it matters in context
Binder design has advanced quickly through diffusion generators and structure-predictor-driven optimization, but published campaigns have typically validated a modest number of designs against a handful of targets, leaving method comparison largely computational. Screening over a million proteins across 127 targets in one all-to-all format is a different kind of evidence, and it is the first wet-lab head-to-head of several contemporary open methods under identical conditions. Two long-standing hard cases move here: densely polar interfaces, where success has historically tracked epitope hydrophobicity because inverse-folding models struggle with polar residues, and carbohydrates, which are small, hydroxyl-rich and carry a desolvation penalty that had kept them out of reach. Code, weights and the synthetic training dataset are released, which contrasts with several competing systems that remain proprietary.
Our work
This Week at UniBio Intelligence
Material additions to our data, models, tools, and research platform.
platform
Therapeutic Antibody Explorer covering 1,362 clinical and preclinical antibody therapeutics
A searchable roster of antibody programmes filterable by target, indication, clinical phase, molecular format and target class, with flags showing which records have sequence, structure or released model data available. Useful for scoping a target landscape before committing to a programme.
tool
OpenDDE released for all-atom biomolecular complex prediction
Predicts an all-atom complex from supplied molecular entities and reports confidence, entity conservation and geometry checks together, with a dedicated antibody-antigen mode accepting a heavy or VHH chain, optional light chain and one antigen.
model
Two QSP models published: target-gated cargo delivery and AAV-delivered antibody in brain
The target-gated cargo model covers construct opening and local versus systemic receptor occupancy for conditionally activated formats, directly relevant to this issue's conditional-activation papers. The 725-state rat model couples AAV biodistribution, transduction, transgene-derived antibody, FcRn-aware distribution and brain and CSF transport for gene-therapy-delivered antibodies.
data
VIDRA genetic dose-response data queryable through UbiMCP
Variant-Informed Dose-Response Analysis models gene-phenotype dose-response relationships from more than 1.6 million trait-associated germline variants, spanning common trait, rare disease and gene burden associations. Useful for asking whether partial versus complete target inhibition is likely to be tolerated before committing to a modality.
Primary papers
- [1] JMB2403, a potential best-in-class PD-1-dependent IL2Rβγ-activating tri-specific antibody for safe and potent immunotherapy (mAbs)
- [2] Dual Targeting of EGFR and HER2 by a Bispecific Antibody-Drug Conjugate Demonstrates Broad Antitumor Activity in Multiple Solid Tumors (Molecular Cancer Therapeutics)
- [3] DOTAM-TCB: A Universal Small Molecule-Guided Hapten- and T Cell-Bispecific Antibodies for Cancer Immunotherapy (Antibody Therapeutics)
- [4] DCAF11-dependent molecular glue degrader activated by glutathionylation (Nature)
- [5] Mechanistic modeling of time-dependent antibody clearance in multiple myeloma: A physiologically based pharmacokinetic model of isatuximab (Journal of Pharmaceutical Sciences)
- [6] Latent Generative Search unlocks de novo Design of Untapped Biomolecular Interactions at Scale (Proteína-Complexa) (NVIDIA (with ICLR 2026 companion paper))