Ubi Titer Issue #16
Controls That Decide: Bispecific ADCs, Counter-Selection, and Target-Agnostic Discovery
Eight papers. Several report a control that changes how the main result should be read: a comparator matched on payload and drug load, a counter-selection step during sorting, a validation split that holds out whole sequence clusters.
- antibody-drug conjugates
- bispecific antibodies
- VHH / nanobody engineering
- peptide-HLA targeting
- acute myeloid leukemia
- quantitative systems pharmacology
- developability prediction
The field
This Week in Biologics
Paper 1 · mAbs
Preclinical development of ABL206, a novel bispecific antibody-drug conjugate targeting B7-H3 and ROR1
Against exact-matched monospecific ADCs carrying the same linker-payload at the same DAR, dual B7-H3 and ROR1 targeting produced tumour regression where single targeting produced only stasis, isolating the contribution of bispecific targeting from payload potency.
Core finding
ABL206 is a 2+2 bispecific antibody-drug conjugate, built by fusing an anti-B7-H3 scFv to the C-terminus of an anti-ROR1 IgG1 and conjugating tavatecan, an exatecan-based linker-payload, site-specifically at the N297 glycan to a drug-to-antibody ratio of 4. Compared at equimolar dose against exact-matched monospecific ADCs sharing the identical linker-payload and DAR, the monospecific anti-ROR1 ADC achieved only partial tumour inhibition and the anti-B7-H3 ADC achieved stasis, while ABL206 drove continuous regression. Across 38 patient-derived xenograft models spanning nine tumour types it produced greater than 50 percent tumour growth inhibition in 84 percent of models and regression below baseline in 53 percent.
What is novel
The monospecific comparators carry the same linker-payload at the same drug-to-antibody ratio, so the comparison separates dual targeting from payload potency. Quantitative flow cytometry measured ROR1 surface expression at 25,398 antibody binding sites per cell against 147,800 for B7-H3, a 5.8-fold difference, indicating that the poor internalisation of the anti-ROR1 arm reflects low receptor density rather than slow endocytic kinetics. The bispecific format recruits the abundant antigen to carry the complex to the lysosome. In one model, tumours that had regrown after ifinatamab deruxtecan treatment regressed rapidly on a single subsequent dose.
Limitations
The authors state they cannot exclude that the intrinsic potency of exatecan over DXd contributes to the advantage over ifinatamab deruxtecan, alongside the targeting mechanism and possible pharmacokinetic differences. Binding to cynomolgus ROR1 was not measured directly; it is inferred from complete extracellular domain sequence identity. The proposal that reduced Fc-gamma receptor engagement lowers interstitial lung disease risk is supported only by the absence of pulmonary findings in monkeys, and requires clinical confirmation. Two-dimensional cytotoxicity assays gave attenuated maximum killing, a known limitation for camptothecin-based ADCs.
Why it matters in context
ROR1-directed ADCs have shown activity in haematological malignancies but little in solid tumours, where the most advanced candidate reported a 1 percent overall response rate. B7-H3-directed ADCs have performed better but carry interstitial lung disease as a dose-limiting class effect of deruxtecan-based platforms, reported in over 10 percent of patients. Pairing the two antigens addresses both problems at once, and the approach follows the same thesis as earlier B7-H3 bispecific ADCs while adding controls that separate targeting from payload chemistry.
Paper 2 · mAbs
Discovery and engineering of bispecific TCR-mimic antibodies targeting peptide-HLA complex
Counter-selecting against irrelevant peptide-HLA complexes during B cell sorting, then mapping single-residue binding tolerance across 172 peptide variants, separated eight highly specific anti-WT1 antibodies from 26 candidates where both published benchmarks bound off-targets.
Core finding
Antibodies against the WT1 peptide RMFPNAPYL presented on HLA-A*02:01 were isolated from transgenic mice expressing human antibody variable regions, using both single B cell sorting and immune phage libraries. Specificity was engineered into the selection itself: antigen-specific B cells were sorted for binding to the WT1 complex and simultaneously for absence of binding to a CMV peptide on the same HLA allele, deselecting binders that engage the HLA framework rather than the peptide. Of 26 specific antibodies, 13 bound the off-target peptide M13L and 14 bound PIGQ on peptide-pulsed T2 cells, while 8 bound neither.
What is novel
An X-scan assay displays a library of 172 single-chain peptide-HLA variants on phage, each differing from WT1 by one amino acid, and reads binding tolerance across every position. That tolerance profile is then scanned against the human proteome by motif matching to identify potential cross-reactive peptides, moving proteome-wide specificity assessment from late safety testing into the discovery phase. Both literature benchmark antibodies, ESK1 and 33H9, bound both off-target peptides, and epitope binning against an HLA-A*02:01-specific antibody resolved the new binders into two distinct structural classes. Binding EC50 values for the leads ranged from 0.181 to 1.02 nanomolar.
Limitations
This is a proof of concept on a single well-studied target, and specificity was assessed using phage-displayed complexes and peptides exogenously loaded onto T2 cells rather than naturally processed and presented peptide repertoires. The proteome scan identifies candidate cross-reactive sequences by motif, which does not establish that those peptides are actually presented on healthy tissue at relevant density. No in vivo efficacy or safety data are reported, and weak binding to two cancer-associated peptides was observed for one lead.
Why it matters in context
Peptide-HLA complexes expose intracellular proteins to antibody-based targeting, reaching the roughly 70 percent of the proteome that does not appear on the cell surface, and clinical validation exists in the form of an approved TCR bispecific for metastatic uveal melanoma. The constraint is specificity: T-cell receptors are naturally cross-reactive across many peptide-HLA pairs, off-target recognition on healthy tissue has caused severe clinical toxicity, and earlier antibodies against this target performed poorly in the clinic for that reason. Counter-selection during discovery removes cross-reactive clones before they are characterised, which is the cheapest point to discard them.
Related tools on Ubi Biologics
- MHC-Difftool
Works with peptide-MHC complexes, the target class this paper screens for single-residue cross-reactivity.
Paper 3 · Journal of Chemical Information and Modeling
Combining AI Structure Prediction and Integrative Modeling for Nanobody-Antigen Complexes
On 40 nanobody-antigen complexes released after every method's training cutoff, AlphaFold3 reached 32.5 percent acceptable top-ranked predictions; pooling ensembles from three predictors and docking with HADDOCK beat that baseline when epitope information was available.
Core finding
A benchmark of 40 nanobody-antigen complexes was assembled from structures released after the September 2021 training cutoff, with no CDR3 sequence homology to nanobodies seen by the predictors, spanning CDR3 lengths from 6 to 24 residues and antigens from 3.48 to 56.39 kilodaltons. AlphaFold2-Multimer achieved 25.0 percent acceptable success on top-ranked predictions and AlphaFold3 32.5 percent, rising to 52.5 percent across the top ten. Pooling conformational ensembles from AlphaFold2, AlphaFold2-Multimer and ImmuneBuilder, clustering on CDR3 conformation, and docking the selected models with HADDOCK exceeded the AlphaFold baseline provided some epitope information was available.
What is novel
The work separates nanobody modelling accuracy from complex prediction accuracy, which are usually reported together. No single predictor dominates the first task: mean CDR3 backbone RMSD was 3.10 Angstrom for AlphaFold2-Multimer, 3.23 for ImmuneBuilder and 3.39 for AlphaFold2 monomer, differences not significant against their standard deviations. Combining ensembles rather than sampling more from one method improved the best achievable model substantially, with median CDR3 RMSD falling from 1.94 Angstrom for AlphaFold2-Multimer alone to 1.23 Angstrom for the merged ensemble. A useful negative result: AlphaFlow, which improves antibody H3 loop sampling, gave no comparable benefit for nanobody CDR3, improving only two of nine difficult cases.
Limitations
The advantage depends on having epitope information from mutagenesis, crosslinking or hydrogen-deuterium exchange, which is often unavailable in the cases where modelling is used, and the margin over the baseline is smaller without it. AlphaFold3 was the strongest single predictor but licensing terms prevented its models being carried into downstream docking, so the best available starting structures remain untested in the full workflow. The benchmark is 40 complexes, and AlphaFold3 was run with 25 models per complex rather than the extensive multi-seed sampling reported elsewhere, which understates its ceiling.
Why it matters in context
Protein complex prediction uses coevolutionary signal, the statistical trace left when contacting residues in two interacting partners change together over evolutionary time. Antibodies and nanobodies are generated by somatic mutation within an individual rather than co-evolved with their targets, so that signal is absent, and nanobody binding depends heavily on a long CDR3 loop with a wide conformational space. This is why complex prediction accuracy for this format lags other protein-protein systems, and why physics-based information-driven docking remains competitive with end-to-end deep learning here.
Related tools on Ubi Biologics
Paper 4 · bioRxiv
DiffDose: Differentiable Programming for Personalized Dose-Regimen Optimal Control
Automatic differentiation through the solver, dosing events and objective gives gradients on administration times as well as dose amounts, letting one optimised filgrastim dose per cycle cut cumulative neutrophil deficit 86.1 percent below a repeated-daily reference.
Core finding
DiffDose treats the mechanistic model, its dosing events, the numerical solver, quadrature and clinical objective as a single differentiable computational graph, so automatic differentiation propagates objective gradients to dose amounts and administration times without a hand-derived adjoint for each model. Against the three published OptiDose pharmacokinetic and pharmacodynamic benchmarks, forward-mode automatic differentiation matched the hand-derived adjoint optimum, reaching a tracking loss of 4.01 against 4.05, with the shortest time to solution at roughly 1.8 seconds. Finite-difference optimisation stalled at a loss of 82 under matched solver settings.
What is novel
Dose timing becomes an optimisation variable rather than a design choice tested by enumeration, because moving an administration shifts a discontinuity in the trajectory that fixed-schedule methods cannot represent. In a state-dependent delay model of chemotherapy-induced neutropenia across seven CHOP-14 cycles, optimising a single 300 microgram filgrastim administration time per cycle converged on 7.4 to 8.0 days after chemotherapy, reducing cumulative absolute neutrophil count deficit by 93.4 percent against no G-CSF and 86.1 percent against the repeated-daily reference. A fixed day-8 schedule retained most of the benefit, indicating a robust biological window rather than a knife-edge optimum. Scaled to a 250-member virtual population on a 36-state mosunetuzumab model, the optimiser recovered clinically recognisable step-up dosing from the equations and objective alone, with a median 18.8 percent reduction in peak interleukin-6.
Limitations
These are non-convex local optimisation problems, so gradient-based methods exploit accurate derivatives without guaranteeing a global optimum, and multi-start strategies may be needed where clinically distinct basins exist. Every optimised regimen inherits the correctness of the underlying mechanistic model, and legacy implementations, including many in MATLAB, require refactoring into differentiable JAX or Julia stacks. Timing gradients in the delay model required a smoothed administration gate because the direct event implementation did not yield validated derivatives. Additional tumour control was often marginal because 67.1 percent of virtual patients already achieved at least 99 percent tumour reduction under the reference regimen.
Why it matters in context
Regimen selection in practice is largely empirical, using dose-escalation rules early and therapeutic drug monitoring later, and examines few candidate schedules. Pharmacometric optimal control has existed for decades but requires model-specific sensitivity or adjoint derivations that must be redone whenever the model, dosing structure, objective or cohort changes, which has limited adoption. Framing the whole simulation as a differentiable program shifts the effort from deriving sensitivities to validating the differentiated system, and turns a calibrated model from something that scores prespecified regimens into something that proposes them.
Related tools on Ubi Biologics
- Drug-stimulated biomarker turnovermodel
The indirect response benchmark from this paper, runnable on the platform. This is the case the paper reports in the main text, where forward-mode automatic differentiation reached the published optimum fastest.
- Drug injury and delayed tumor lossmodel
The tumour growth inhibition benchmark from this paper, runnable on the platform.
- Bispecific exposure and ternary bindingmodel
The bispecific T-cell engager benchmark from this paper, runnable on the platform.
Paper 5 · mAbs
Predicting non-specific binding of VHHs using machine learning models with cluster-aware validation
Pipeline developability data clusters by project and those clusters track the label, so random cross-validation inflates performance; under leave-one-group-out validation an ensemble model reaches 0.73 AUROC, and 0.79 on a sequence-distant test set.
Core finding
A dataset of 1,563 VHHs with baculovirus particle binding measurements, collated across more than a dozen discovery projects, splits almost evenly between acceptable and problematic non-specific binding. Sequence-space projection shows the data forms clusters of closely related molecules, and critically the binding classes follow the cluster structure, so cluster membership acts as a confounder a model can exploit without learning anything about the drivers of polyreactivity. Under a cluster-aware leave-one-group-out scheme across seven true clusters and ten groups of unaffiliated sequences, a soft-voting ensemble of logistic regression, random forest and LightGBM reached 0.73 AUROC, and 0.79 AUROC with 0.46 Matthews correlation on a held-out set of sequences at least 10 Levenshtein distance from the rest.
What is novel
The contribution is a validation protocol rather than an architecture. Sequences within 10 edit distance of any test molecule are purged from training in each cross-validation round, and feature selection is performed inside the cross-validation loop so its variability enters the performance estimate. Reporting performance separately for clustered and diverse sequences exposes a pattern that a single aggregate number hides: most models performed worse on closely related sequences, the regime that matters during lead optimisation, than on diverse ones. A logistic regression on mean-pooled ESM2 embeddings was the exception, scoring 0.73 and 0.70 across the two regimes, while structure-derived physicochemical descriptors offered interpretability at comparable accuracy.
Limitations
Absolute performance is modest and suited to triage rather than decision-making, and the authors present it as such. Binarising a continuous, heteroscedastic assay readout at a threshold discards information, though it reduces the effect of experimental noise at high values. The descriptors depend on AlphaFold2 models and commercial software packages, which constrains reproduction. Results are specific to one non-specific binding assay in one organisation's molecule population, and a single assay is an imperfect proxy for in vivo clearance behaviour.
Why it matters in context
In silico developability filters are increasingly used to triage candidates before material is produced, and the standard practice is to train them on accumulated historical pipeline data because no other data exists at scale. That data is structurally unlike the benchmarks machine learning methods are usually validated on: small, low in diversity, and composed of tight families of close relatives from a handful of campaigns. Single biophysical descriptors previously associated with polyreactivity, including isoelectric point, hydrophobicity, CDR3 length and arginine content, show only weak correlations here, with significance often driven by a few outliers, which is the argument for multivariate models and for validating them honestly.
Related tools on Ubi Biologics
- Nanobody Polyreactivitytool
Predicts nanobody polyreactivity. This paper models a related liability using a different assay, and its cluster-aware validation scheme applies to how any such predictor should be benchmarked.
- Ubi Developtool
Developability assessment, the setting in which a triage-grade polyreactivity score would be used.
Paper 6 · mAbs
A framework for evaluating unexpected reactions in antibody-drug conjugates during clinical development: a case study of a promiscuous payload
A duocarmycin linker-drug formed an intramolecular cross-link to glutamic acid 168 in its own antibody light chain, abolishing cytotoxicity at roughly 15 percent after three months and 39 percent after six, while routine DAR and purity assays stayed within specification.
Core finding
During stability testing of the clinical-stage ADC SYD1875, potency fell after three months at 25 degrees Celsius while average drug-to-antibody ratio, reversed-phase purity and capillary electrophoresis results gave no explanation. Peptide mapping by liquid chromatography and tandem mass spectrometry identified a new species 18 daltons lighter than the expected conjugated peptide and less hydrophobic, whose fragmentation indicated hydroxylation of the DNA-alkylating moiety. Further unbiased searching showed the underlying event: the linker-drug had formed a covalent intramolecular cross-link joining heavy chain residues 39 to 43 to light chain residues 153 to 172, at glutamic acid 168.
What is novel
The paper sets out a generalisable sequence from detection through identification to clinical risk assessment, and then executes each step. Because a cross-linked variant cannot practically be synthesised, hydroxylated versions of the payload and linker-drug were prepared chemically as mechanistic surrogates, conjugated, and shown to lose cell-killing activity, confirming causation rather than correlation. Cross-linking reached approximately 15 percent at three months and 39 percent at six months at the stressed condition. Incubation in mouse, cynomolgus monkey and human plasma showed no difference in rate or extent, establishing the reaction as temperature-driven chemistry rather than an enzymatic process, which makes preclinical species predictive for this attribute.
Limitations
The findings concern one ADC with a duocarmycin payload, and the specific cross-linking chemistry depends on a reactive DNA-alkylating warhead positioned near an accessible glutamic acid, so the particular reaction may not generalise even to other ADCs using this linker-drug class. The quantitative extent of cross-linking is a relative estimate from peptide mapping, with the six-month figure reported as data not shown. Because the molecule carries multiple linker-drugs, a species with one cross-linked payload retains others that remain active, which complicates the relationship between cross-link level and potency loss.
Why it matters in context
ADC quality control inherits methods from monoclonal antibodies plus attributes specific to the linker-drug, but regulatory guidance specific to ADCs remains limited and no structured approach to investigating unexpected attributes has been well established. The case illustrates a general hazard of the format: a chemically reactive small molecule is held for years in close proximity to a protein dense in nucleophilic side chains. The consequence that travels furthest is bioanalytical, since the original ELISA could not distinguish cross-linked from intact conjugate and would have overestimated active drug, requiring a drug-load-sensitive assay before pharmacokinetic data could be interpreted correctly.
Paper 7 · Cell Reports Methods
Development and application of nbLIBRA-seq for high-throughput discovery of antigen-specific nanobodies
Adapting antigen-barcoded single-cell sequencing to immunised alpacas recovered 362 antigen-specific heavy-chain B cells against TfR1 from one blood sample, and 1,125 across two related viral glycoproteins screened in a single multiplexed library.
Core finding
LIBRA-seq links B cell receptor sequence to antigen specificity by conjugating each antigen to a unique DNA barcode, so single-cell sequencing returns the receptor sequence alongside counts of which antigens that cell bound. Adapting it to camelids required a flow cytometry panel gating on IgG subclasses 2 and 3 together with VHH to capture heavy-chain-only antibody producers, and alpaca-specific primers to recover variable and hinge regions. From a single blood sample from an alpaca immunised seven times with the extracellular domain of human CD71, 362 antigen-specific B cells were recovered after strict filtering, using five distinct inferred germline genes with CDR3 lengths from 5 to 23 residues.
What is novel
Multiple antigens can be screened within one library, which conventional panning cannot do. Including the fusion glycoproteins of respiratory syncytial virus and human metapneumovirus in a single experiment recovered 1,125 antigen-specific heavy-chain-expressing B cells and produced cross-reactivity information between the two related antigens as a direct output rather than a follow-up study. Recovering the hinge region alongside the variable domain preserves isotype information. Lead candidates selected on barcode score were expressed as heavy-chain antibodies and assessed for internalisation in Jurkat T cells, confirming that predicted binders behave as expected.
Limitations
Barcode-derived scores predict specificity rather than measure affinity, and candidates still require expression and binding confirmation. Sixty-nine B cells scored above the positive threshold against the influenza hemagglutinin control antigen, which the authors report and which sets a realistic expectation for background in this readout. The work validates the method on a small number of leads per campaign rather than characterising the full recovered repertoire, and depends on a successful immunisation, so it accelerates screening without removing the animal step or the response-quality dependency.
Why it matters in context
Nanobody discovery has been dominated by phage display panning, which requires several rounds of selection and amplification, handles one antigen per campaign, and enriches for molecules that tolerate the selection conditions rather than necessarily the best binders. Single-cell approaches that pair receptor sequence with antigen specificity have reshaped conventional antibody discovery but had not been established for camelid heavy-chain antibodies, where the relevant B cells are defined by different isotypes and the constant region primers differ. The gain is in the breadth of starting material available per animal and per experiment.
Related tools on Ubi Biologics
Paper 8 · Antibody Therapeutics
Discovery of candidate antibodies and antigens for acute myeloid leukemia therapy by combining whole-cell phage display selection and CRISPR-Cas9 library screening
Whole-cell phage display against intact AML cells, with no target chosen in advance, followed by CRISPR-Cas9 knockout deconvolution, identified PTPRG, Nectin-2 and CD105; Nectin-2 CAR-T matched CD123 CAR-T against primary patient blasts.
Core finding
A chimeric rabbit-human Fab-phage library was panned against intact AML cells rather than a purified antigen, so selection was target-agnostic. Fabs carried a Sortase A recognition site allowing only Fab-displaying phage to be biotinylated and captured, which removes phage that adsorb non-specifically to the cell surface. Panning against AML cells and healthy donor mononuclear cells in parallel, with heavy chain CDR3 as the clone identifier read by sequencing, converted enrichment into a differential measurement and yielded 192 candidate identifiers, of which 42 expressed Fabs bound the AML line.
What is novel
Target deconvolution used a genome-wide CRISPR-Cas9 knockout library instead of immunoprecipitation and mass spectrometry. Cells lose antibody staining when the gene encoding the cognate antigen is knocked out, so sorting the unstained population and sequencing its guide RNAs names the antigen directly from the genome. Two rounds of staining and sorting enriched antigen-negative cells from 4.1 percent to 69.9 percent, and analysis ranked PTPRG first, with four of its five guides significantly enriched. The same approach identified Nectin-2 and CD105 for other antibody categories, and knockout lines confirmed antigen dependence of binding in each case.
Limitations
None of the three antigens shows broad cell-intrinsic dependency in public CRISPR dependency data, though that does not preclude targeting them with effector-dependent modalities where antigen presence rather than essentiality is what matters. Selection appears biased toward antigens with large ectodomains, which the authors propose as the reason established AML targets such as CD33, CD123 and CD70 did not emerge. Only 42 of 187 recovered phagemids produced Fabs that bound, indicating substantial attrition in CDR3-mediated clone recovery. The primary-sample cohort is small, and because antibodies were selected on intact cells some may recognise conformational or glycosylation-dependent epitopes that were not characterised.
Why it matters in context
Antibody and cell therapy for acute myeloid leukemia has been constrained less by binder generation than by the shortage of surface antigens present on leukaemic cells and absent from normal haematopoietic tissue, with CD33 and CD123 both carrying expression on healthy cells. Conventional discovery selects a nominated target and raises binders against it, which cannot surface antigens nobody proposed. Running discovery against whole cells inverts that order, and pairing it with genetic deconvolution addresses the historical weakness of the approach, namely that it produces useful antibodies whose targets remain unknown.
Since we last wrote
Follow-Ups
Updates on work covered in earlier issues.
Proteina-Complexa is now a full preprint, with a crystal structure
Issue #10 covered this binder-design campaign from NVIDIA's laboratory release: over one million designs against 127 targets, codesigned sequences returning 691 on-design hits against 365 for inverse-folding redesign of the same backbones, and picomolar PDGFR binders. The preprint adds physical validation, including a 2.95 Angstrom crystal structure of an 800-residue design that resolved 794 residues and matched its computational model at 2.08 Angstrom backbone deviation, a length at which most design methods yield nothing usable. It also clarifies the carbohydrate result: the design that discriminates the A blood-group antigen from B places a tyrosine to hydrogen-bond the N-acetyl group unique to A, while a separate design binds a fucose residue common to A, B and O, an off-target interaction that compresses the dynamic range of the cell assay.
Our work
This Week at UniBio Intelligence
Material additions to our data, models, tools, and research platform.
tool
Immunogenicity prediction is now available in Ubi Biologics
You can assess HLA class II presentation, population-level DRB1 risk, and MHC class I presentation through the HLAIIPred, ImmunoGeNN and MHCFlurry workflows. Immunogenicity risk is one of the developability liabilities that decides whether a sequence is worth carrying forward, alongside the non-specific binding covered in this issue.
Primary papers
- [1] Preclinical development of ABL206, a novel bispecific antibody-drug conjugate targeting B7-H3 and ROR1 (mAbs)
- [2] Discovery and engineering of bispecific TCR-mimic antibodies targeting peptide-HLA complex (mAbs)
- [3] Combining AI Structure Prediction and Integrative Modeling for Nanobody-Antigen Complexes (Journal of Chemical Information and Modeling)
- [4] DiffDose: Differentiable Programming for Personalized Dose-Regimen Optimal Control (bioRxiv)
- [5] Predicting non-specific binding of VHHs using machine learning models with cluster-aware validation (mAbs)
- [6] A framework for evaluating unexpected reactions in antibody-drug conjugates during clinical development: a case study of a promiscuous payload (mAbs)
- [7] Development and application of nbLIBRA-seq for high-throughput discovery of antigen-specific nanobodies (Cell Reports Methods)
- [8] Discovery of candidate antibodies and antigens for acute myeloid leukemia therapy by combining whole-cell phage display selection and CRISPR-Cas9 library screening (Antibody Therapeutics)