Ubi Titer Issue #15
Counter-Selection: Off-Target Objectives, Dual-Antigen Engagers, and Sequence Variant Limits
Five papers, and the common thread is counter-selection: each one is about specifying what a molecule or a process must avoid, not only what it must hit, and each shows that naming the off-target changes the outcome.
- de novo binder design
- binder specificity
- tri-specific T-cell engagers
- colorectal cancer
- AAV gene therapy
- physiologically based pharmacokinetics
- sequence variants
- antibody manufacturing quality
- peptide agonist selectivity
The field
This Week in Biologics
Paper 1 · bioRxiv
Specificity-driven protein binder design with Odin-Multi
Optimising one binder sequence against on-targets and off-targets simultaneously raised peptide-MHC discrimination between complexes differing by a single residue from 6.0% to 14.2% of designs, at a cost in overall yield.
Core finding
Odin-Multi extends AlphaFold2-based binder hallucination from a single-complex objective to simultaneous optimisation of one shared sequence against several on-target and off-target complexes, with attractive losses favouring binding to targets and repulsive losses discouraging it at counter-selection targets. Across three pairs of class B1 G protein-coupled receptors, 83.5 to 96.8 percent of jointly optimised designs exceeded an interaction-confidence threshold for both receptors, against 6.8 to 36.3 percent for designs optimised against one receptor at a time.
What is novel
Cross-reactivity and specificity become explicit design objectives rather than properties left to emerge during downstream screening. The most demanding benchmark paired a cancer-testis antigen epitope presented on HLA-A*02:01 against a single-residue variant, sharing 99.6 percent overall identity and a C-alpha RMSD of 0.23 Angstrom; counter-selection raised designs satisfying both the target-confidence criterion and a 2.5 discrimination ratio from 6.0 to 14.2 percent. One counter-selected clone exceeded a previously published minibinder prototype in target avidity while discriminating better, and its predicted complex localises the mechanism to a peptide methionine side chain occupying a pocket with contacts within 4.2 Angstrom of five binder residues.
Limitations
Counter-selection carries a yield cost, with designs exceeding the enrichment threshold falling from 11 of 166 in the target-only condition to 3 of 167 under the most stringent counter-selection. The two counter-selected conditions differed in both off-target identity and weighting, so the comparison does not isolate either factor. The experimental hit rate for the cross-reactive toxin campaign was 1 of 30 designs, and the structural account of discrimination rests on a predicted rather than an experimentally determined complex.
Why it matters in context
Deep-learning binder design now achieves high experimental success rates against single targets, but specificity and breadth have typically been handled after generation, through binding-site selection, structural filtering or laboratory screening. That ordering cannot enrich for an interaction profile the generator never optimised, which matters most when on-targets and off-targets share structural features. Explicit multi-target optimisation addresses a recognised gap for applications from polyspecific antivenoms to peptide-MHC-directed therapeutics, where off-target engagement has produced severe clinical toxicity.
Work with this on Ubi Biologics
- MHC-Difftool
Works with peptide-MHC complexes, the target class where this paper's counter-selection produced discrimination between peptides differing by a single residue.
- BoltzGenmodel
Generative binder design surface, the model class this work extends by adding explicit off-target objectives to the optimisation.
Paper 2 · Antibody Therapeutics
Generation and characterization of a CEACAM5 x CDH17 x CD3 tri-specific T-Cell engager for colorectal cancer therapy
A tri-specific engager against CEACAM5, CDH17 and CD3 reached 94.5% tumour growth inhibition where a clinical-stage CEACAM5 bispecific analogue reached 29.0%, with moderate monovalent affinities and low cytokine release.
Core finding
A 1+1+1 tri-specific T-cell engager targeting CEACAM5, CDH17 and CD3 was built on a human IgG1 framework with knob-into-hole heterodimerisation, pairing a CEACAM5-binding Fab with single-chain fragments against CDH17 and CD3. Monovalent affinities were deliberately kept moderate, at 51.0 nanomolar for human CEACAM5, 42.8 nanomolar for CDH17 and 51.9 nanomolar for CD3 epsilon, leaving avidity from dual tumour-antigen engagement to provide selectivity. Across cell lines spanning the full expression matrix it matched or exceeded every single-target comparator.
What is novel
The dual-antigen premise is tested where it could fail rather than only where it succeeds. In a pancreatic line expressing CEACAM5 but no CDH17, the tri-specific matched a CEACAM5 bivalent engager while a CDH17 bivalent was inactive, showing the CEACAM5 arm carries the molecule when the second antigen is absent. In a colorectal model with high CEACAM5 and low CDH17, it reached 94.5 percent tumour growth inhibition against 74.6 percent for a CDH17 biparatopic, 29.0 percent for an analogue of a clinical-stage CEACAM5 bispecific and 28.1 percent for a CEACAM5 bivalent, while cytokine levels in vivo stayed low across all comparators.
Limitations
All efficacy data come from cell-line-derived xenografts in immunodeficient mice reconstituted with human peripheral blood mononuclear cells, with no patient-derived models and no assessment of the tumour microenvironment. Shedding of CEACAM5 into circulation, which the paper notes accompanies loss of epithelial polarity in tumours, is not modelled as a competing sink. No pharmacokinetic data, repeat-dose safety study or formal cytokine release syndrome assessment is presented, and all authors are employees of the developing companies.
Why it matters in context
Single-target immunotherapies against CEACAM5 have repeatedly underperformed in colorectal cancer, with two antibody-drug conjugates discontinued for limited efficacy or ocular toxicity and a bispecific engager reaching only 4 percent overall response as monotherapy. CDH17-directed programmes remain early, with several conjugates and cell therapies in phase 1. Antigen heterogeneity and escape through single-antigen loss are the recognised failure modes that dual tumour-antigen targeting is intended to address, alongside the persistent problem of balancing engager potency against cytokine release.
Work with this on Ubi Biologics
- CD3x5T4 In Vitro Activation and Cytotoxicity QSPmodel
Models T-cell engager activation and cytotoxicity in vitro, the same readouts used here to compare tri-specific against bivalent and biparatopic formats.
Paper 3 · The AAPS Journal
Decoding Nonlinearities in AAV-Based Gene Therapy Using PBPK Modelling
A PBPK model spanning five orders of magnitude of AAV dose shows anti-drug antibody clearance of the expressed product is dose-dependent, negligible at low dose and reaching maximum rate near day 20 at high dose.
Core finding
A whole-body physiologically based pharmacokinetic model covering fifteen tissues, each divided into nine compartments, describes adeno-associated virus vector and the antibody its transgene produces simultaneously across nine dose levels spanning five orders of magnitude in immunocompetent mice. Transgene production is represented as a saturable logistic function of dose, and the effect of anti-drug antibodies on the expressed antibody as a concentration-dependent elimination rate routed through transit compartments to reproduce the delay in onset.
What is novel
Both nonlinearities proved structurally necessary rather than cosmetic. A model carrying only a saturable production coefficient consistently overpredicted plasma concentrations at higher vector doses and failed to reproduce the observed decline at day 21, which is what forced an immunogenicity-driven degradation term into the structure. The resulting dose asymmetry is the transferable result: at the lowest dose examined, degradation of expressed antibody is negligible and begins only after the first week, whereas high doses show both earlier onset and a faster maximum rate. Re-estimating vector dissociation gave a substantially shorter half-life than a previous model weighted toward a different serotype, consistent with tighter binding but faster release after internalisation.
Limitations
The work is confined to mice over a 21-day period, with loss of transgene DNA from the nucleus set to zero beyond that point by assumption. The model overpredicts spleen concentrations at one high dose, attributed to unmodelled macrophage uptake of capsids, and overpredicts transgene product in liver, kidney and heart at high dose because local production is not reduced by the systemic immune response. Validation against an independent dataset reproduced liver and muscle acceptably but systematically overpredicted spleen, kidney and heart.
Why it matters in context
Vector-delivered gene therapy has accumulated clinical setbacks that make first-in-human dose selection acute, including fatal acute liver failure in one muscular dystrophy programme and discontinuation of a second candidate. Re-dosing is generally infeasible because vectors elicit strong anti-capsid responses, while excessive doses carry hepatotoxicity risk, so the first dose must balance efficacy and safety without titration. Empirical scaling approaches do not explicitly account for vector pharmacokinetics or nonlinear transgene expression, and many published models capture only product kinetics while neglecting capsid dynamics.
Work with this on Ubi Biologics
- Target-Gated Cargo QSPmodel
Models delivery-vehicle and payload kinetics together, the same coupled-system problem this PBPK model solves for vector and expressed antibody.
Paper 4 · Antibody Therapeutics
A landscape of sequence variants in 51 commercial antibody therapeutics
Peptide mapping across 51 marketed antibodies found 3,006 sequence variant occurrences, 92.5% at or below 0.1% per site, supporting 0.2% as a pragmatic risk-alert benchmark for commercial material.
Core finding
High-sensitivity peptide mapping by liquid chromatography tandem mass spectrometry was applied across 51 marketed antibody therapeutics, spanning originators and biosimilars, Chinese and multinational manufacturers, eukaryotic and prokaryotic expression systems, and monoclonal, bispecific and antibody-drug conjugate formats. A total of 3,006 sequence variant occurrences were identified, of which 2,780 or 92.5 percent sit at or below 0.1 percent per site, 226 exceed 0.1 percent, 72 exceed 0.2 percent and only 12 exceed 0.5 percent.
What is novel
This supplies an empirical answer to a question the field had been addressing through theory and development-stage material. A widely referenced per-site control limit of 0.1 percent was proposed on the basis of process capability and a small number of molecules; this survey shows marketed products cluster well below it and proposes a higher value as a pragmatic risk-alert level. Method sensitivity is load-bearing for that conclusion: changing column chemistry and increasing sample load raised confidently identified variant sites in a reference antibody from 13 to 18 to 43, exceeding prior interlaboratory counts, so the higher observed proportion above 0.1 percent reflects a sharper instrument rather than poorer products. One biosimilar carried dozens of co-localised tyrosine-to-phenylalanine occurrences, a recognised signature of amino acid depletion in culture rather than random misincorporation.
Limitations
The study is cross-sectional and observational, establishing which variants are present without addressing clinical consequence; the impact of variants above the proposed alert level is explicitly not assessed and is acknowledged as product-specific. No immunogenicity prediction, structural modelling or forced degradation work accompanies the survey, and underlying data are available on request rather than deposited.
Why it matters in context
Sequence variants arise from errors in replication, transcription or translation and are a recognised quality attribute because substitutions can alter structure, function, stability or immunogenicity. Prior discussion of control limits has relied on theoretical error rates, limited datasets or development-stage molecules, leaving a gap about what approved products actually carry. Because this survey originates from a national regulatory control laboratory rather than a single manufacturer, it is positioned to inform regulatory expectations as well as internal specifications.
Work with this on Ubi Biologics
- Therapeutic Antibody Explorerexplorer
Covers clinical and preclinical antibody therapeutics, the same marketed-product landscape this survey profiles for manufacturing quality attributes.
- Ubi Developtool
Profiles sequences across developability features, complementary to the empirical manufacturing-quality benchmarks established here.
Paper 5 · bioRxiv
Structure-Guided Design of C5aR1-Selective Peptide Agonists
Empirical force field scanning identified one peptide position where hydrophobic substitution gains potency at C5aR1 without a matching gain at C3aR, and the prediction transferred across two scaffolds.
Core finding
Mutational scanning with an empirical force field over homology models refined by molecular dynamics identified a single position in two complement-derived peptide agonist scaffolds where hydrophobic substitution was predicted to improve engagement of one anaphylatoxin receptor without a corresponding gain at a closely related receptor. Synthesis and profiling across three receptors confirmed the direction, with the methionine substitution in the first scaffold producing more than a log unit gain in potency at the intended receptor relative to parent.
What is novel
The prediction was receptor-specific and directional rather than simply affinity-improving, and it transferred to a structurally distinct second scaffold where analogous substitutions also raised potency at the intended receptor. The study additionally documents a pathway-dependent trap: adding a second substitution at an adjacent position reduced potency at the target receptor but reduced it more at the off-target receptor, so a selectivity metric based on potency difference improved while absolute target potency fell. Reading selectivity in isolation would have scored that change as progress.
Limitations
All pharmacology is in heterologous overexpression systems rather than native cells, with no in vivo work. Discrimination against the second complement receptor was not achieved by any analogue in the second series, none exceeding tenfold, and the analogue with the most favourable serum stability showed essentially no receptor preference. Serum stability was assessed for only one of the two series, and models were built from structures of related complexes rather than of the peptides studied.
Why it matters in context
Complement anaphylatoxin signalling is implicated across acute and chronic inflammatory and neurodegenerative conditions, and receptor inhibition has been the dominant therapeutic strategy, leaving selective agonists valuable mainly as pharmacological probes. Short peptides derived from the endogenous ligand retain the minimal activation motif but lose the receptor discrimination conferred by the full scaffold, and several established peptide agonists show promiscuity across related receptors. Because individual ligands stabilise distinct receptor conformations, potency, subtype selectivity and signalling bias must be assessed together rather than optimised in sequence.
Primary papers
- [1] Specificity-driven protein binder design with Odin-Multi (bioRxiv)
- [2] Generation and characterization of a CEACAM5 x CDH17 x CD3 tri-specific T-Cell engager for colorectal cancer therapy (Antibody Therapeutics)
- [3] Decoding Nonlinearities in AAV-Based Gene Therapy Using PBPK Modelling (The AAPS Journal)
- [4] A landscape of sequence variants in 51 commercial antibody therapeutics (Antibody Therapeutics)
- [5] Structure-Guided Design of C5aR1-Selective Peptide Agonists (bioRxiv)