Ubi Titer Issue #6
Antibody Developability: Bispecifics, Off-Targets, and Clearance
Developability is finally getting measured instead of assumed — bispecific inheritance, proteome-wide off-target screens, and a cheap plate assay that predicts clearance as well as the field's best tools.
- antibody developability
- bispecific antibodies
- off-target screening
- molecular glues
- self-association
- clearance
What this issue covers
- 1.High-throughput machine learning-aided antibody discovery for cell surface antigens
A synthetic antibody library engineered specifically for machine learning readout recovered antibodies against targets that a standard screen missed entirely.
- 2.Decoding Bispecific Antibody Developability: Design Rules and Predictive Models from a 160-Member Library
A systematic study of 160 bispecific antibodies shows that some developability properties inherit predictably from the parent antibodies, while others only emerge once the two arms are combined.
- 3.Degron-independent recruitment of KAT2A expands the target space of CRBN molecular glues
A newly characterized molecular glue degrades a cancer-linked protein by having cereblon grip a surface tyrosine rather than the degron motif every prior cereblon-based glue has relied on.
- 4.Predicting antibody self-association with sequence-structure fusion models: the central role of CSI-BLI in early developability screening
A low-cost, high-throughput plate assay for antibody self-association turns out to predict in vivo clearance about as well as the field's leading specialized assays, and can now be predicted directly from sequence.
- 5.Off-target reactivity in clinical monoclonal antibodies
A proteome-wide screen of 174 approved and clinical-stage antibodies found that more than a quarter bind at least one unintended human protein, and shows these liabilities can sometimes be engineered away.
Paper 1 · Cell Systems
High-throughput machine learning-aided antibody discovery for cell surface antigens
A synthetic antibody library engineered specifically for machine learning readout recovered antibodies against targets that a standard screen missed entirely.
Core finding
Researchers built a synthetic Fab yeast-display library on a single heavy-chain framework paired with four light chains, diversified in the CDRH3 loop using position-specific frequencies drawn from human B-cell repertoires, and screened it against ten human and murine cell-surface targets including PD-L1, TIGIT, and ROBO1, yielding favorable antibodies for all ten.
What is novel
Each antibody's paratope is encoded as a compact, deep-sequenceable nucleotide tag alongside its CDRH3 — an 'antigen recognition module' that turns the library's own screening output into a large machine-learning-ready dataset. Mining the aggregate sequencing data after the primary screen identified additional antibodies against two targets (ROBO2 and PD-L2) that the direct wet-lab screen did not surface, and the team released a public dataset of over 68,000 Fab sequences and 486 characterized antibodies.
Limitations
The library is built on a single heavy-chain germline framework, so its generalizability to other frameworks or antibody formats beyond Fab is not established in this study.
Why it matters in context
This sits within a broader push across the field to design discovery libraries and assays explicitly for machine learning consumption, rather than treating sequencing data as an afterthought to a conventional screen — a shift from post-hoc data mining toward data-generation-first library design.
Paper 2 · bioRxiv
Decoding Bispecific Antibody Developability: Design Rules and Predictive Models from a 160-Member Library
A systematic study of 160 bispecific antibodies shows that some developability properties inherit predictably from the parent antibodies, while others only emerge once the two arms are combined.
Core finding
A team profiled 160 bispecific antibodies built on a uniform knobs-into-holes CrossMab IgG1 scaffold, along with their 65 parental arms, across ten developability assays on a high-throughput biophysical platform, to determine which liabilities transfer predictably from parent to bispecific and which do not.
What is novel
The study separates bispecific developability into distinct inheritance regimes: hydrophobicity and surface charge inherit almost linearly from the parental arms (Spearman correlations of roughly 0.85 to 0.95), meaning parental-level screening alone predicts these outcomes, while self-association and polyreactivity inherit only partially (correlations of roughly 0.60 to 0.88), with emergent failures traced mechanistically to charge complementarity between the two paired variable fragments — a property that does not exist until the arms are combined.
Limitations
All bispecifics in the study share a single format (knobs-into-holes CrossMab IgG1), so the design rules derived here may not transfer directly to other bispecific architectures such as BiTEs or DVD-Ig constructs.
Why it matters in context
This adds a large, systematic dataset to a growing body of work on bispecific developability prediction, including prior public developability competitions that found existing predictive models generalize poorly out of sample — underscoring why format-specific, mechanistically grounded design rules like these are needed rather than one-size-fits-all predictors.
Paper 3 · Science
Degron-independent recruitment of KAT2A expands the target space of CRBN molecular glues
A newly characterized molecular glue degrades a cancer-linked protein by having cereblon grip a surface tyrosine rather than the degron motif every prior cereblon-based glue has relied on.
Core finding
Researchers identified a cereblon (CRBN)-based molecular glue that selectively degrades the lysine acetyltransferase KAT2A, a protein that cooperates with oncogenes such as c-Myc and KMT2A fusions to sustain malignant programs, and used cryo-electron microscopy to show that CRBN recruits KAT2A independently of the canonical degron motif that virtually all previously characterized CRBN glues depend on.
What is novel
Instead of mimicking a degron, the glue engages a surface-exposed tyrosine on KAT2A in a recognition mode the authors describe as antibody-like, demonstrating that the CRBN-targetable proteome is not limited to proteins bearing degron-like motifs. Selective KAT2A degradation ablated histone H3K9 acetylation, showed antiproliferative effects in acute myeloid leukemia cell lines, and produced efficacy in a patient-derived xenograft model.
Limitations
The work establishes degron-independent recruitment for a single target (KAT2A); whether this recognition mode generalizes to other surface-exposed motifs across the proteome remains untested.
Why it matters in context
This extends an active line of research into the structural basis of CRBN neosubstrate recognition, adding a conceptually distinct recruitment mechanism to a field that has largely focused on cataloging and re-purposing degron-containing targets for molecular glue development.
Paper 4 · mAbs
Predicting antibody self-association with sequence-structure fusion models: the central role of CSI-BLI in early developability screening
A low-cost, high-throughput plate assay for antibody self-association turns out to predict in vivo clearance about as well as the field's leading specialized assays, and can now be predicted directly from sequence.
Core finding
Across a panel of 246 monoclonal antibodies, clone self-interaction biolayer interferometry (CSI-BLI) — a plate-based, low-material assay — showed a moderate positive correlation with high-concentration formulation viscosity and, in a 41-antibody hFcRn transgenic mouse panel, a strong correlation (Spearman ρ = 0.65) with non-target-mediated clearance, comparable to leading nonspecificity assays and clearly outperforming AC-SINS and classical hydrophobicity metrics.
What is novel
The team built an end-to-end model fusing a fine-tuned protein language model (ESM-2) with AlphaFold-derived structural graphs via disentangled multi-stream attention to predict CSI-BLI class directly from sequence, achieving F1 = 0.76 for single-domain VHHs and F1 = 0.57 for full IgGs on edit-distance-controlled holdout sets designed to prevent sequence leakage, with Shapley-value analysis identifying charge, hydrophobicity, and aggregation-propensity patterns as the underlying drivers.
Limitations
Predictive performance is notably weaker for full IgGs than for VHHs, reflecting the added complexity of paired heavy-and-light-chain interactions, and the authors caution that predictions are unreliable for antibodies with unusual CDR lengths, underrepresented germline families, or formats outside the training distribution.
Why it matters in context
This work reinforces a growing recognition that early, inexpensive developability assays can serve as proxies for costly downstream liabilities like clearance and viscosity, and contributes another entry to the expanding set of sequence-structure fusion architectures being applied to antibody developability prediction.
Paper 5 · Structure
Off-target reactivity in clinical monoclonal antibodies
A proteome-wide screen of 174 approved and clinical-stage antibodies found that more than a quarter bind at least one unintended human protein, and shows these liabilities can sometimes be engineered away.
Core finding
Using rapid extracellular antigen profiling (REAP), a yeast-display-based screening method, researchers evaluated 174 FDA-approved and clinical-stage antibodies against 6,172 human extracellular proteins and found that 28% of the antibodies exhibited at least one off-target interaction.
What is novel
Structural and biophysical analysis traced off-target binding to two distinct mechanisms: properties intrinsic to the antibody itself, and genuine epitope mimicry occurring either within related protein families or across proteins with no obvious sequence relationship. As a proof of concept, the team identified previously unrecognized off-target interactions for the clinical antibody tanezumab and engineered its variable domains to eliminate them while preserving both target affinity and developability.
Limitations
Only a small subset of the predicted off-target interactions (nine) were orthogonally validated using an independent binding method, leaving most of the reported hits unconfirmed beyond the primary screening platform.
Why it matters in context
This is among the largest systematic proteome-wide specificity screens of clinical-stage antibodies performed to date, building on an emerging class of cell-based and protein-array methods for polyspecificity testing, and demonstrates that off-target risk in already-approved biologics is more prevalent than routine preclinical characterization typically surfaces.
Primary papers
- [1] High-throughput machine learning-aided antibody discovery for cell surface antigens (Cell Systems)
- [2] Decoding Bispecific Antibody Developability: Design Rules and Predictive Models from a 160-Member Library (bioRxiv)
- [3] Degron-independent recruitment of KAT2A expands the target space of CRBN molecular glues (Science)
- [4] Predicting antibody self-association with sequence-structure fusion models: the central role of CSI-BLI in early developability screening (mAbs)
- [5] Off-target reactivity in clinical monoclonal antibodies (Structure)