Ubi Titer Issue #14
Geometry Over Affinity: Binder Design, Synapse Volume, and Receptor Occupancy
Eight papers, and the through-line is geometry rather than affinity: which residues sit where on a surface, how far apart two binding sites are, which epitope pair a bispecific engages, and, in two pharmacology papers, what volume you assume binding happens in at all.
- antibody design
- protein language models
- VHH / nanobody structures
- TCR-mimetic antibodies
- T-cell engagers
- NK-cell engagers
- cytokine mimetics
- antibody developability
- quantitative systems pharmacology
- receptor occupancy
The field
This Week in Biologics
Paper 1 · mAbs
DyAb: sequence-based antibody design and property prediction in a low-data regime
Training on affinity differences between antibody pairs rather than absolute values turns ~500 labelled points into ~125,000 training pairs, and produced an anti-EGFR variant at 66 pM from roughly 100 starting measurements.
Core finding
DyAb predicts the difference in binding affinity between two antibody sequences rather than the affinity of either one. Sequences are embedded with an antibody-specific protein language model (AntiBERTy or LBSTER), the two embeddings are subtracted to give a relative embedding, and a convolutional network is trained on those differences. Because every pair becomes a training example, a 500-point dataset expands to nearly 125,000 pairs. Across three programmes, 85-89% of designs expressed and bound their target, matching the rate for single point mutants, and most improved on the parent affinity. One anti-EGFR design reached 66 pM against a 3.0 nM lead.
What is novel
Pairwise learning directly attacks the bottleneck in early antibody engineering, where labelled affinity data is scarce and expensive. Crystal structures of the anti-EGFR lead and its top design, solved at 2.4 A and 2.1 A, show the model repeatedly selected a Kabat V97 glycine-to-aspartate substitution that alters the CDR-H3 conformation, with an added proline at V98 that may lock the productive conformation in place. A sequence-only model arrived at a structural solution.
Limitations
Only binding affinity was validated. Aggregation, viscosity, polyreactivity, thermal and chemical stability, and immunogenicity were not assessed, and all of these decide whether a candidate survives development. Performance is expected to degrade for variants far from the training lead, and only three antibody-antigen systems were tested. A linear baseline on absolute embeddings matched DyAb on held-out data from the same distribution, and only fell behind on prospective design rounds.
Why it matters in context
Pairwise and contrastive framings have appeared for small molecules and for protein fitness prediction, and this work carries the approach into therapeutic antibody optimisation with prospective wet-lab validation rather than retrospective benchmarks. It also adds evidence that antibody-specific language models outperform general protein models on antibody tasks, which remains contested in the field.
Paper 2 · Protein Science
AVIDbase: A biologically accurate structural dataset of nanobody-antigen complexes
A manually curated set of 690 non-redundant nanobody-antigen interfaces that corrects biological assembly errors, retains glycans and cofactors, and collapses 72 anti-SARS-CoV-2 RBD complexes into 8 genuinely distinct epitopes.
Core finding
AVIDbase compiles every X-ray nanobody-antigen complex in the PDB as of February 2026 into 1,741 recorded interfaces, 690 of them non-redundant. Where a deposited biological assembly is wrong or incomplete, the correct one is rebuilt, in eight cases directly from crystallographic symmetry mates. Glycans, cofactors and non-canonical residues near the interface are retained and annotated rather than discarded, and missing electron density is flagged by its distance from the interface.
What is novel
The corrections are quantified against every comparable database rather than asserted. Forty-two percent of antigen oligomeric assemblies differ from those in SAbDab, and 20% of entries capture interface residues SAbDab omits entirely; 25% differ from SAAINT-DB and 27% from ANABAG. Rosetta repacking and minimisation measurably improves the correlation between buried polar surface area and interface hydrogen-bond energy, which is direct evidence the cleaned structures better represent the interface in solution rather than in a crystal. Epitope clustering then addresses a bias most datasets carry silently: 72 anti-SARS-CoV-2 RBD complexes reduce to 8 distinct epitopes, 43 at the ACE2 site and 21 at a beta-2 strand site targeted to disrupt trimerisation.
Limitations
The dataset covers X-ray structures only, so cryo-EM complexes are absent. Rosetta repacking is stochastic and can depart from the true residue configuration, which is why unminimised structures are also released. Sequence-space analysis shows the fivefold increase over the earlier 123-complex Zavrtanik set mainly improves local sampling rather than reaching genuinely new regions, and large areas of nanobody sequence space remain structurally uncharacterised.
Why it matters in context
Several nanobody structural databases have appeared in close succession, each handling assembly assignment and redundancy differently. This one leans on manual literature review for the edge cases that defeat automated pipelines, and is the only one to rebuild assemblies from symmetry mates. As structure prediction and affinity models for VHH-antigen complexes continue to underperform relative to globular protein complexes, the quality of the underlying training data is an increasingly plausible part of the explanation.
Paper 3 · mAbs
Artificial intelligence-assisted design of MAGE-A4 x CD16 T-cell receptor-like natural killer cell engagers
Docking-angle analysis predicted which TCR-mimetic clones would bind the HLA framework instead of the presented peptide, before any wet-lab testing, and the predicted failure mode appeared exactly as forecast after humanisation.
Core finding
A TCR-mimetic antibody against the MAGE-A4 GVY230-239 peptide presented on HLA-A*02:01 was built into a natural killer cell engager by fusing four anti-CD16a VHH domains to both termini of the IgG light chain, with an N297A mutation silencing Fc effector function. The engager killed MAGE-A4-positive tumour lines in a target- and CD16a-dependent manner, confirmed with purified NK cells and CD16a blockade rather than bulk PBMCs alone.
What is novel
Candidate antibodies were triaged by whether their AlphaFold3-docked pose reproduced the diagonal binding angle natural T-cell receptors use across a peptide-HLA groove. Of 15 modelled clones, 8 were classified canonical, 4 biased and 3 non-specific before any binding assay, and the classification held: the clones predicted to engage the HLA framework rather than the peptide showed the strongest non-specific binding. Clone 5C11 then demonstrated the failure mode precisely. Humanisation added hydrogen bonds between its light chain and the HLA alpha-1 helix while eliminating the CDR3 contact with the peptide, so the molecule gained affinity and lost specificity. Alanine substitution at the structurally predicted peptide contacts G5 and R6 reduced binding as the model anticipated.
Limitations
Cytokine release was not measured, so the premise that NK engagers carry a lower cytokine risk than T-cell engagers is assumed rather than demonstrated here. All efficacy data is in vitro, with no tumour microenvironment or biodistribution assessment. Only the lead antibody was carried into the engager format, so the predictive framework has not been tested by putting a predicted-nonspecific clone through the same cytotoxicity comparison. Specificity was assessed against a limited peptide panel rather than a broad survey of naturally presented human peptides.
Why it matters in context
Peptide-HLA targets open the large intracellular proteome to antibody-based therapy, but the flat, conserved HLA surface makes framework binding a persistent failure mode, and affinity measurements alone do not separate peptide-selective binders from cross-reactive ones. Using predicted binding geometry as a specificity checkpoint before committing to wet-lab characterisation is a practical response to that problem, and the humanisation result shows why the checkpoint matters at more than one stage.
Paper 4 · mAbs
Antibody framework engineering enables augmented cytokine receptor signaling capacities of bispecific single domain antibody-based IL-21 mimetics
Across 81 bispecific VHH pairings, epitope choice governed IL-21 receptor agonism more than affinity, and framework mutations forcing VHH:VHH dimerisation lifted an almost inactive construct from 15% to 82% STAT3 activation.
Core finding
Twenty-seven IL-21R-binding VHHs were combinatorially paired with three IL-2Rgamma binders to give 81 monovalent bispecific antibodies, screened for STAT3 phosphorylation in NK-92 cells. Framework mutations adapted from I-shaped antibody engineering, plus three new sets designed with ProteinMPNN, were then introduced to force non-covalent VHH:VHH dimerisation and rigidify the receptor pairing. In the IgG-like format this raised maximum STAT3 activation from 15.4% to 82.4% and IFN-gamma release from 113 to 1033 pg/mL, approaching wild-type IL-21.
What is novel
Agonism tracked epitope pairing rather than binding strength. Two IL-2Rgamma binders sharing an epitope but differing thirtyfold in affinity produced very different efficacy, while a binder of comparable affinity at a different epitope behaved differently again, indicating that receptor geometry sets the ceiling and affinity modulates within it. The rigidification result cuts both ways, which is what makes it mechanistic: applying the same mutations to a construct whose VHHs engage different epitopes on both subunits reduced activity rather than improving it. Binding affinities were unchanged by the framework mutations, so the effect is geometric rather than a hidden avidity gain.
Limitations
Efficacy approached wild-type IL-21 but potency did not. Every mimetic remained substantially right-shifted in dose-response, consistent with the wider pattern for antibody-derived cytokine agonists. One mutation set had to be abandoned because intermolecular association dropped monomer content to 53-62%, and the remaining non-germline framework substitutions will need immunogenicity assessment before they can be considered developable. Structural models used to rationalise the geometry were generated with AlphaFold2, which the authors note is unreliable for antibody-antigen complexes.
Why it matters in context
Antibody-based cytokine mimetics have been reported for IL-2, IL-18, IL-12 and type I interferon receptors, generally with the observation that receptor geometry rather than affinity drives signalling outcome. This work adds IL-21 to that set and contributes a transferable engineering handle, showing that constraining paratope geometry through framework mutations can rescue an inactive format, provided the underlying epitope pair is already compatible with a signalling-competent receptor arrangement.
Paper 5 · mAbs
Generation of a highly versatile TCR bispecific scaffold optimized for targeting tumor-specific peptide-HLA antigens
Twenty-two TCR bispecifics across seven scaffolds and three T-cell recruiters: the most compact format paired with the weakest-binding recruiter gave the widest therapeutic window, with 5 of 6 mice in remission versus 0 of 6.
Core finding
Seven bispecific scaffolds spanning 75 to 200 kDa and 1+1, 2+1 and 2+2 valencies were built around a single affinity-matured PRAME-specific single-chain TCR and crossed against three T-cell-recruiting antibodies. The compact TCER format was most stable under accelerated stress and most potent, and was the only format able to kill tumour cells presenting roughly 50 peptide-HLA copies per cell without also killing antigen-negative controls.
What is novel
Two results run against common design intuition. Added valency did not help: the 2+2 format gained potency but lost specificity, killing PRAME-negative cells, and a bivalent 125 kDa design gained nothing at all, because at approximately 100 copies per cell there is no avidity effect to exploit. More strikingly, the weakest-binding T-cell recruiter performed best. BMhanced bound T cells so weakly it was only detectable above 100 nM, yet drove higher NFAT signalling, more granzyme B, interferon-gamma and IL-2, greater T-cell proliferation, and put five of six mice into remission at 0.025 mg/kg against zero of six for an SP34-similar recruiter at matched in vitro potency. It also required 25-fold higher concentration before inducing target-independent cytokine release in human whole blood. Stability differences across scaffolds were large: the 87 kDa format retained only 14-17% monomer after two weeks at 40 C.
Limitations
The mechanism behind the low-affinity recruiter advantage is not established. Improved biodistribution and more physiological force-mediated receptor triggering are proposed, but no biodistribution study, signalling analysis or structure was performed. The scaffold panel was not designed to isolate binding-site distance as a variable, so the distance-potency relationship is a trend across differing constructs rather than a controlled comparison. Peptide specificity was assessed against a predefined homology-based panel, which the authors note is not a complete safety assessment.
Why it matters in context
Peptide-HLA complexes are presented at far lower density than conventional surface antigens, typically around 100 copies per cell against thousands, which makes potency the binding constraint for this target class and explains why avidity strategies that work elsewhere fail here. The finding that lowering recruiter affinity widens the therapeutic window aligns with a growing body of work on T-cell engager design, and stands in contrast to the majority of clinical engagers, which still use high-affinity CD3 binders.
Paper 6 · Antibody Therapeutics
Developability Improvement of an MMP-9 Inhibitory mAb by CDR-H3 Hydrophobic Residue Substitution
Replacing two hydrophobic CDR-H3 residues extended in vivo half-life from 27.7 to 71.9 hours and tripled the duration of pain relief, while binding affinity and inhibitory potency were essentially preserved.
Core finding
Two surface-exposed hydrophobic residues, Trp102 and Phe104, were identified near the N-terminal end of CDR-H3 in an MMP-9 inhibitory antibody and replaced using a deliberately small alphabet of four substitutions at each position, giving 16 double mutants rather than the 400 an exhaustive scan would require. The lead mutant extended serum half-life from 27.7 to 71.9 hours in mice and lengthened pain relief in a streptozotocin model of diabetic neuropathy from 2 days to 6 after intravenous dosing, and up to 9 days subcutaneously.
What is novel
Structural prediction determined which residues were safe to change. The two hydrophobic residues were modelled as sitting at the end of the F beta-strand rather than in the protruding loop responsible for engaging the MMP-9 active site, which is why they could be substituted without losing function, and the biochemical data bore this out: affinity moved from 41 to 74 nM while inhibitory potency slightly improved, from 59 to 38 nM. The developability gains were substantial across every measured axis. Production rose from 66 to 161 mg/L, non-specific binding to cell-surface proteins fell from 5.65% to 0.80%, and monomeric content reached 100%, exceeding trastuzumab measured under identical conditions. The pharmacokinetic improvement and the efficacy improvement scale together, a 2.6-fold half-life extension against a 3-fold longer efficacy window.
Limitations
Chemical and storage stability, immunogenicity, and viscosity at high concentration were not measured, and all three matter for a molecule intended for chronic subcutaneous dosing. The winning substitution introduces an aspartate, which alters overall charge distribution as well as surface hydrophobicity, and the study does not separate which of the two properties drives the half-life gain. Efficacy was assessed in a single rodent model of one indication.
Why it matters in context
Large hydrophobic patches in complementarity-determining regions have been repeatedly implicated in aggregation, polyspecificity and accelerated clearance, and computational developability prediction has grown around that observation. What this study contributes is a complete chain from a structurally identified liability through a small, cheap mutant panel to measured pharmacokinetics and in vivo efficacy in the same molecule, which is the link most developability work leaves to inference.
Paper 7 · CPT: Pharmacometrics & Systems Pharmacology
A Generalization of the Ternary Binding Model to Membrane-Confined Systems With Finite Copy Number
Modelling bispecific engager binding in the actual synapse volume rather than in bulk solution shifts effective antigen concentration by six orders of magnitude, which is the difference between antigen density predicting nothing and predicting a 2.5-fold dose difference.
Core finding
The standard ternary binding model used to interpret bispecific T-cell engager pharmacology assumes three species mixing freely in a shared three-dimensional volume. The productive trimeric complex does not form under those conditions. It forms inside a nanoscale contact zone between two cells, at close-apposition microdomains on microvillus tips, with receptors present in finite integer copy numbers. Replacing the macroscopic volume with an effective reactive contact volume, and replacing the closed-system ligand mass balance with an open-reservoir condition justified by a Damkohler analysis, preserves the original algebra while changing what it can represent.
What is novel
The consequence is structural rather than a numerical correction. Under conventional bulk mapping, effective antigen concentration for two B-cell leukaemia lines sits around 10^-3 nM, far below the 1.49 nM dissociation constant, which places the system in a regime where the antigen term carries no numerical weight and the model predicts identical dose-response for cell lines differing 2.48-fold in CD19 density. Membrane confinement raises effective concentration to roughly 10^3 nM and restores antigen density as a governing variable. The paper also argues the conventional half-maximal formation metric is the wrong therapeutic referent because it normalises each cell line to its own maximum, erasing the density scaling; the proposed replacement, the dose required to form a fixed absolute number of trimers, predicts that the lower-density line needs about 2.48-fold more drug, and that ratio is invariant to the geometric parameters, which all cancel. A stochastic treatment adds a sharper test: bulk mapping predicts essentially zero productive synapses at therapeutic concentrations, which is difficult to reconcile with blinatumomab's picomolar clinical potency.
Limitations
This is a theoretical reframing with a retrospective case study and no prospective experimental validation. Receptor internalisation and recycling, spatial heterogeneity within the synapse, serial killing by a single T cell, and stochastic kinetics at very low copy number are all outside the current formulation. The trimer threshold assumed sufficient for T-cell activation is inferred from native T-cell receptor literature and from blinatumomab's potency rather than measured for any bispecific. The geometric sensitivity figures in the supplement were produced under an earlier formulation the paper explicitly retires, and carry legend language the authors caution against over-reading.
Why it matters in context
Ternary complex formation underpins dose selection reasoning for bispecific engagers, chimeric antigen receptor constructs and, in a related form, bifunctional degraders. The claim here is not that the established algebra is wrong but that the volume it is applied to determines which asymptotic regime the system occupies, and therefore whether target expression has any predictive role in a dose-response model. The framework applies to any bridging therapeutic acting at a cell-cell interface, and yields falsifiable predictions about how dose ratios should track copy-number ratios when contact geometry is shared.
Paper 8 · CPT: Pharmacometrics & Systems Pharmacology
Semi-Mechanistic Modeling of S-531011, a Humanized Anti-CCR8 Monoclonal Antibody, for Prediction of CCR8 Receptor Occupancy in Human Tumor Tissues
A three-compartment receptor occupancy model predicts tumour-tissue target engagement that the ongoing trial cannot measure, after calibration against mouse data moved the in vitro dissociation constant thirtyfold.
Core finding
A pharmacokinetic and receptor occupancy model for an anti-CCR8 antibody tracks four species in each of three compartments, with the bivalent antibody binding target sequentially in 1:1 and then 1:2 stoichiometry, and target continuously synthesised and degraded. The purpose is to estimate receptor occupancy in tumour tissue, which the ongoing first-in-human study cannot sample. Simulations indicate 80 to 800 mg every three weeks maintains over 90% tumour receptor occupancy at trough.
What is novel
The model is calibrated against observed mouse tumour receptor occupancy rather than assembled from in vitro parameters and trusted, and that calibration is where the useful result sits. The dissociation constant measured in a kinetic exclusion assay was 18.62 pmol/L; fitting the observed animal data required 558.6 pmol/L, a thirtyfold shift, which the authors attribute to the difference between a binding constant measured against cells in buffer and one operating against a native membrane receptor subject to turnover and permeability limits. Global sensitivity analysis identified the internalisation half-life of the antibody-target complex as by far the dominant parameter, well ahead of clearance or central volume. The tumour penetration assumption was then stress-tested by reducing the perfusion coefficient to a third and a tenth of its fitted value, under which 240 mg still maintains over 90% occupancy.
Limitations
No human receptor occupancy data exists in either blood or tumour, so the human prediction rests entirely on a model refined against mouse observations and cannot be checked against truth. Tumour volume is held constant throughout treatment, which is inconsistent with effective target-cell depletion and would alter transfer rates and target amounts if it changed. Day 7 animal occupancy data was excluded from model evaluation because regulatory T-cell depletion confounds it, which means the quantity modelled is receptor binding rather than pharmacological depletion. The degradation rate for unbound target was assumed equal to that of the complexes, and the authors' own sensitivity analysis shows that assumption can materially affect predicted tumour occupancy.
Why it matters in context
Dose selection for antibodies in oncology has moved away from maximum tolerated dose toward integrated assessment of target engagement, and receptor occupancy models have become a standard instrument for that argument where the tissue of interest cannot be sampled. The value here is less the specific dose recommendation than the demonstration that an in vitro binding constant can be wrong by more than an order of magnitude in the direction that matters, and that calibrating against in vivo occupancy is worth the effort before a model informs dose.
Primary papers
- [1] DyAb: sequence-based antibody design and property prediction in a low-data regime (mAbs)
- [2] AVIDbase: A biologically accurate structural dataset of nanobody-antigen complexes (Protein Science)
- [3] Artificial intelligence-assisted design of MAGE-A4 x CD16 T-cell receptor-like natural killer cell engagers (mAbs)
- [4] Antibody framework engineering enables augmented cytokine receptor signaling capacities of bispecific single domain antibody-based IL-21 mimetics (mAbs)
- [5] Generation of a highly versatile TCR bispecific scaffold optimized for targeting tumor-specific peptide-HLA antigens (mAbs)
- [6] Developability Improvement of an MMP-9 Inhibitory mAb by CDR-H3 Hydrophobic Residue Substitution (Antibody Therapeutics)
- [7] A Generalization of the Ternary Binding Model to Membrane-Confined Systems With Finite Copy Number (CPT: Pharmacometrics & Systems Pharmacology)
- [8] Semi-Mechanistic Modeling of S-531011, a Humanized Anti-CCR8 Monoclonal Antibody, for Prediction of CCR8 Receptor Occupancy in Human Tumor Tissues (CPT: Pharmacometrics & Systems Pharmacology)