Ubi Titer Issue #13
Beyond Affinity: Charge, Release Chemistry, and Format Transfer
Four of this week's five papers converge on the same limitation from different directions: binding affinity is rarely the property that decides whether a designed molecule works, and the properties that do decide it — surface charge, payload permeability, cooperative unfolding, degradation activity — have to be measured or engineered separately.
- de novo minibinder design
- CAR-T engineering
- isoelectric point
- antibody-drug conjugates
- payload release chemistry
- VHH-Fc developability
- targeted protein degradation
- protein language model alignment
The field
This Week in Biologics
Paper 1 · bioRxiv
Predicting VHH-Fc Developability from Large-Scale IgG Data
Models trained only on 559 IgGs predict VHH-Fc surface properties zero-shot (heparin binding rho 0.82), but thermostability does not transfer because VHH-Fc lacks the CH1-CL interface that drives IgG unfolding.
Core finding
GDPa5 is a 160-member VHH-Fc library profiled across 10 biophysical assays, the first public standardized developability dataset for this format. Models trained exclusively on 559 IgG heavy chains predicted VHH-Fc properties zero-shot: heparin binding at Spearman rho 0.820, hydrophobicity 0.626, and self-association 0.620 at pH 7.4. Thermostability reached only 0.155. Tabular neural networks (TabPFN v2.5, TabICLv2) consistently beat ridge-regression baselines, which sat near zero for most assays.
What is novel
First public VHH-Fc developability dataset and the first use of tabular foundation models for antibody developability prediction. The central claim is that training-set scale matters more than format match: IgG-trained models outperformed VHH-specific models across the panel despite the scaffolds differing structurally. Isoelectric point alone recovered most of the heparin-binding signal (rho 0.795 against a best model of 0.827), indicating that assay is largely a readout of surface charge.
Limitations
The headline comparison places VHH-only models under leave-one-cluster-out cross-validation against IgG-trained models evaluated zero-shot, so evaluation protocol differs alongside training data. Only 100 of 160 members passed the 80% SEC monomer threshold. Polyreactivity failed to reproduce against an independent dataset on the 34 shared sequences, and the authors state they cannot explain the discordance and ask for replication. Adding VHH-Fc data to IgG training gave almost nothing: a 25% spike-in improved a single assay by 0.02.
Why it matters in context
The field has assumed each new therapeutic format needs its own curated developability dataset, a costly requirement as the landscape diversifies beyond IgG. Prior Ginkgo releases established standardized measurement at scale, and the 2025 developability competition showed that current algorithms overfit training distributions and generalize poorly out of distribution. This work suggests surface-driven physicochemical rules are largely shared across scaffolds while cooperative unfolding behaviour is not, and the thermostability failure has a clean mechanistic explanation rather than being a data artifact. IgG-derived clinical thresholds also do not transfer without recalibration.
Work with this on Ubi Biologics
- Ubi Developtool
Profiles sequences across interpretable developability features, the same hydrophobicity, charge and self-association axes this dataset measures experimentally.
- Nanobody Polyreactivitymodel
Predicts the nanobody polyreactivity liability that this study found hardest to transfer across formats and hardest to reproduce across datasets.
- AbLang2model
The antibody-specific language model used here as a featurizer for the best-performing heparin binding and hydrophobicity predictions.
Paper 2 · Nature Methods
Aligning protein-generative models to experimental fitness with ProteinDPO
Direct preference optimization aligns a structure-conditioned protein language model to experimental stability data, then stabilizes H5N1 hemagglutinin by up to 17 C in one round while preserving broadly neutralizing antibody binding.
Core finding
Applying direct preference optimization to ESM-IF1 with protein backbone as prompt, sequence as response and measured stability as preference produces a model that both scores and generates thermostable sequences. Trained on roughly 660,000 Megascale variants across 405 domains, it surpassed the unsupervised base model and supervised fine-tuning on the same data. On H5N1 hemagglutinin, 27 of 45 designs were stabilizing, reaching 17 C melting-temperature improvement in a single design round.
What is novel
First structure-conditioned protein language model preference-optimized against experimental biophysical data. The alignment transfers to regimes absent from training: absolute stability of large proteins, thermal melting of multichain antibodies, and protein-protein binding affinity. Notably it recovered 6 of 8 previously reported stabilizing mutations for H1 and H3 subtypes with no hemagglutinin data in training, and independently identified the three residues proposed to form the pH switch governing transition to the post-fusion state.
Limitations
The training signal is stability alone. The authors state that the post-fusion state was not considered during variant generation and that evidence beyond prefusion stabilization and retained native-antibody binding would be needed to establish vaccine efficacy. Sequences generated for the backbone-stabilization experiments were assessed computationally with Rosetta and ESMFold rather than expressed. The model also underperforms a specialized supervised predictor when double mutations are scored as a sum of single-mutation likelihoods.
Why it matters in context
Inverse-folding models show surprising zero-shot ability to guide stability and expression improvements without observing those properties, yet remain behind supervised stability predictors, a gap that motivates borrowing alignment methods from language-model post-training. Supervised fine-tuning is the obvious alternative but maximizes likelihood over positive examples only, and here it overfits visibly: it improves on held-out data from the training distribution while degrading on independent benchmarks and losing the ability to rank absolute stability of natural proteins. Traditional influenza immunogen stabilization relies on proline insertion, disulfide engineering and cavity filling, which often cost expression or perturb folding when applied without whole-protein context.
Work with this on Ubi Biologics
- BoltzGenmodel
Generative design surface for binders and scaffolds, the model class this paper shows can be aligned to measured biophysical properties rather than sequence likelihood alone.
Paper 3 · Nature Communications
AI-enabled discovery and biochemical optimization of minibinders targeting cancer cell-surface proteins
A picomolar CD276 minibinder made a worse CAR than weaker binders. Isoelectric point, not affinity, governed surface trafficking, with a working window of pI 6.5 to 8.5.
Core finding
RFdiffusion minibinders against three B7-family targets showed sharply different design efficiency: PD-L1 yielded a 2.4% binding population and 15 of 17 validated hits down to 2 nM, while CD276 gave few and VTCN1 almost none. Separately, affinity failed to predict function in a CAR format. A published picomolar CD276 binder at pI 9.4 trafficked poorly to the T-cell surface while lower-affinity in-house binders at pI 4.7 to 5.2 expressed well. A genetic algorithm generating roughly 6,000 variants that froze every interface residue and mutated only the rest established pI 6.5 to 8.5 as the window governing surface accessibility, activity and selectivity.
What is novel
Identifies a biochemical determinant entirely outside the binding interface as decisive for whether a designed binder functions in a therapeutic format, and demonstrates a design strategy that tunes it without touching antigen recognition. The work also reports that Chai-1 with language model embeddings predicted experimental binding strength better than the AlphaFold2 interface error score used during the original design pass.
Limitations
No in vivo validation of the minibinder CAR-T cells, and immunogenicity of de novo scaffolds remains untested, which the authors identify as the necessary next step requiring syngeneic models. The VTCN1 campaign essentially failed. The isoelectric point finding derives from variants of a single scaffold family. Detection-reagent performance was inconsistent across targets and cell lines, with one binder showing background on a knockout line in one context but not another.
Why it matters in context
Generative pipelines have made de novo binder design broadly accessible, and prior work established that minibinders can serve as CAR recognition domains. What this adds is a systematic account of what breaks between a validated binder and a working therapeutic. The favourable charge window is consistent with a proteome-wide analysis of extracellular and plasma membrane proteins, suggesting a general constraint of secretory-pathway biology rather than a CAR-specific artifact. The design method itself carries a countervailing bias, producing acidic sequences 2.3 times more often. Highly basic variants also lost target selectivity, killing knockout cells nearly as well as wild-type, most plausibly through nonspecific electrostatic interaction with a negatively charged cell surface.
Work with this on Ubi Biologics
- RFantibodytool
Structure-based de novo design of binders, the design stage this paper shows is necessary but far from sufficient for a functional CAR.
- ipSAEtool
Interface confidence scoring for predicted complexes, directly comparable to the pAE and ipTM metrics benchmarked against experimental binding here.
Paper 4 · Nature
A binding-to-release strategy for targeted anticancer drug delivery
A linker cleaved by a nucleophile already sitting in the target binding pocket releases payload on binding rather than internalization, giving 5.9-fold higher tumour exposure and a 7.5-fold higher tolerated dose.
Core finding
A conjugate linker carrying an electrophilic phosphorus centre releases its payload when target binding positions it next to a nucleophilic residue in the binding pocket, removing the requirement for endocytosis and lysosomal trafficking. For fibroblast activation protein, cryo-electron microscopy and tandem mass spectrometry identified Tyr745 as the reactive residue, and mutating it to phenylalanine reduced payload release by 63%. The resulting conjugate delivered 5.9-fold higher intratumoural payload exposure than its internalization-dependent equivalent and matched antibody-conjugate exposure levels.
What is novel
Converts endocytosis from a hard requirement into a design choice for drug conjugates. The enabling chemistry was developed specifically for this purpose after an existing exchange reaction proved insufficiently reactive toward tyrosine, then tuned across 15 structural variants to balance reactivity against hydrolytic stability. The mechanism is established rather than inferred, with a point mutation at the identified residue serving as the causal control.
Limitations
Release happens outside the cell, so the payload must cross the membrane unaided. Monomethyl auristatin E worked while the membrane-impermeable MMAF was inactive, a constraint the authors state is inherent to the mechanism. The approach also requires a suitably positioned nucleophilic residue within the binding pocket, which bounds the accessible target space. Extracellular release raises off-target exposure questions, addressed by noting the released fraction is a small proportion of administered dose and proposing payload permeability tuning and fractionated dosing as mitigations.
Why it matters in context
Clinical conjugate success has concentrated on a small number of efficiently internalizing antigens, and roughly 180 of more than 2,000 human membrane proteins internalize well, leaving most tumour-specific surface antigens effectively out of reach. Fibroblast activation protein illustrates the gap precisely: it is diagnostically validated across many cancer types by imaging agents, yet a targeted antibody-drug conjugate reached only disease stabilization as a best response, hampered by poor internalization. Other non-internalizing strategies exist through extracellular protease cleavage or two-component systems needing an exogenous trigger; this platform is single-agent and triggered by binding itself. Its generality rests on a survey of nearly 19,000 ligand-protein structures showing 94.4% contain a tyrosine within 10 angstroms of the ligand.
Paper 5 · bioRxiv
A Mammalian High-Throughput Screen for AI-Designed Peptide-Guided Protein Degraders
A pooled human-cell screen selects peptide-guided degraders by degradation activity rather than binding, and works on EWS::FLI1, a disordered fusion oncoprotein with no stable pocket.
Core finding
Libraries of 2,000 language-model-designed 20-mer peptides fused to a truncated CHIP E3 ligase domain are introduced into target reporter cell lines under inducible control. Cells in the lowest 10% of target fluorescence are isolated by cell sorting and enriched guides recovered by sequencing. Validated across four targets, enriched degraders reduced endogenous beta-catenin and Wnt signalling, reduced GFAP and glioblastoma cell viability, and reduced the EWS::FLI1 fusion oncoprotein while suppressing its transcriptional target and increasing apoptosis.
What is novel
Applies pooled CRISPR-style functional selection to targeted protein degradation inside human cells, replacing binding-based selection in non-human display systems. The architecture is modular in the CRISPR sense, with the peptide specifying the target and the ligase domain supplying activity, so retargeting means changing only the guide. It also demonstrates screening against a target expressed from its native genomic locus rather than an overexpressed reporter.
Limitations
Enrichment did not map uniformly onto endogenous degradation. For GATA2, two non-enriched candidates also reduced target abundance, and the authors reframe the screen as a prioritization step requiring downstream validation rather than a clean functional readout. The GATA2 degrader taken forward to a cell-invasion assay did not reach statistical significance against the negative control. Three of four reporters rely on ectopic target expression, which may alter target stoichiometry or epitope accessibility relative to native conditions.
Why it matters in context
Many disease-driving proteins, particularly transcription factors and fusion oncoproteins, lack the defined pockets small molecules require. Peptides and antibody fragments can engage such surfaces, and display technologies discover them from randomized libraries, but those workflows are labour-intensive and select outside the mammalian intracellular context where function ultimately has to hold. Structure-based generative models depend on specified conformations, which complicates design against disordered targets and motivated the sequence-based peptide design models used here. Fusing such guides to a ligase domain is established prior art; the bottleneck addressed is the step after design, rapidly identifying which generated guides actually degrade in cells. The negative controls carry much of the credibility, with non-enriched guides producing no degradation for three of four targets and proteasome inhibition abolishing the effect.
Primary papers
- [1] Predicting VHH-Fc Developability from Large-Scale IgG Data (bioRxiv)
- [2] Aligning protein-generative models to experimental fitness with ProteinDPO (Nature Methods)
- [3] AI-enabled discovery and biochemical optimization of minibinders targeting cancer cell-surface proteins (Nature Communications)
- [4] A binding-to-release strategy for targeted anticancer drug delivery (Nature)
- [5] A Mammalian High-Throughput Screen for AI-Designed Peptide-Guided Protein Degraders (bioRxiv)