← All UbiTiter issuesUbiTiter Issue #19
Immunotoxin Selectivity, Nanobody Stability, and Bovine Antibody Engineering
Three papers examine how protein function depends on context: toxin activity relative to delivery, stabilizing residues within a sequence, and binding domains within a scaffold.
- immunotoxins
- receptor selectivity
- nanobody thermostability
- protein language models
- sparse autoencoders
- bovine ultralong CDR-H3
- antibody engineering
The field
This Week in Biologics
Paper 1 · bioRxiv preprint
Attenuating payload potency expands the therapeutic window of immunotoxins
Pairing strong receptor engagement with reduced toxin activity improves selectivity, with different attenuation optima for HER2, HER3 and CD123 constructs.
Core finding
Liang et al. tested whether catalytic attenuation could convert differences in receptor engagement into selective cell killing. In a high-affinity, bivalent HER2 immunotoxin, E148D preserved killing of receptor-high cells while producing an EC50 selectivity ratio of at least 4.6 million; albumin-binding versions also increased the tolerated mouse dose from 0.05 to 20 mg/kg. The preferred attenuation differed across targets, with K51E giving the largest gains for CD123 and HER3.
What's novel
The study varies payload activity alongside binder affinity and valency, showing why greater receptor engagement alone may improve potency without improving selectivity. Attenuation changes the relationship between delivery and cell killing, allowing a binding difference to produce a larger functional separation.
Limitations
The 4.6-million EC50 ratio is a lower bound set by the assay range, so its exact magnitude should not be treated as resolved. The proposed activity threshold also depends on internalization, trafficking and payload stability, which means receptor abundance alone cannot predict the appropriate attenuation.
In context
The work makes catalytic activity part of the targeting problem. Its target-dependent optima explain why maximizing payload potency can be counterproductive when the goal is to preserve killing in cells receiving more toxin while reducing toxicity in cells receiving less.
Paper 2 · Bioinformatics
Mechanistic interpretability of fine-tuned protein language models for nanobody thermostability prediction
Sparse features preserve nanobody thermostability prediction performance and connect model attributions to context-dependent mutation hypotheses.
Core finding
Murakami and colleagues fine-tuned ESM-2 8M for nanobody melting-temperature prediction, increasing test-set Spearman correlation from 0.695 for a frozen-encoder baseline to 0.780. A sparse autoencoder preserved similar predictive information while exposing features associated with VHH tetrad residues and the framework disulfide. Follow-up free-energy calculations agreed with the predicted direction of stability change for eight of nine mutations.
What's novel
The analysis applies sparse decomposition to a representation already adapted to thermostability prediction. It then connects feature activation to local structural hypotheses, including comparisons in which an attributed position was more favorable for a substitution than neighboring sites.
Limitations
A feature attribution explains the model, but its physical interpretation remains dependent on sequence and structural context; the instability of the aggregate positional map under perturbation makes that distinction consequential. Directional agreement with free-energy calculations supports candidate mutations without establishing the experimental size of their predicted Tm changes.
In context
The study adds interpretability to an existing predictive workflow. Its value lies in narrowing a sequence-level prediction to specific, structurally motivated hypotheses, while retaining the distinction between predictive association and a transferable rule for protein stabilization.
Paper 3 · FEBS Letters, review
Becoming the bull: Structural insights and engineering strategies for ultralong CDR-H3 paratopes from cattle
A review connects bovine knob structure with reformatting outcomes, explaining why preserved antigen binding can coexist with altered biological activity.
Core finding
Evers et al. review how bovine ultralong CDR-H3 antibodies concentrate antigen recognition in disulfide-rich knob domains and how those domains can be reused in other architectures. The surveyed studies demonstrate transfers into antibody loops, Fc fusions and isolated peptides, with performance depending on the knob and its new scaffold. Examples of retained binding accompanied by reduced neutralization show that recognition and biological activity can respond differently to reformatting.
What's novel
The review brings structural and engineering results together around the requirements for successful knob transfer. It connects terminal proximity, insertion-loop compatibility and scaffold presentation to the behavior of the resulting molecule, rather than presenting the formats as interchangeable ways to carry the same binder.
Limitations
The cited success fractions come from different targets and workflows, so they cannot provide a head-to-head ranking of formats. Terminal proximity is also a proposed folding explanation whose relevance depends on the individual knob, rather than a universal rule that predicts successful transfer.
In context
The review positions the scaffold as part of a knob domain's functional context. Its synthesis explains why preserving an antigen-binding surface can still leave expression or biological activity altered, and why modular antibody design must account for the geometry of the final molecule.
Primary sources
- [1] Attenuating payload potency expands the therapeutic window of immunotoxins (bioRxiv preprint)
- [2] Mechanistic interpretability of fine-tuned protein language models for nanobody thermostability prediction (Bioinformatics)
- [3] Becoming the bull: Structural insights and engineering strategies for ultralong CDR-H3 paratopes from cattle (FEBS Letters, review)