Ubi Titer Issue #4
Targeted Protein Model Adaptation Beats Foundation Models
Targeted adaptation beats foundation models: this week's papers show that narrow, local optimization — in sequence design, PLM fine-tuning, and VHH engineering — consistently outperforms general-purpose approaches in biologics.
- antibody design
- protein language models
- VHH
- T-cell engagers
- antibody-drug conjugates
- brain delivery
What this issue covers
- 1.Efficient generation of epitope-targeted antibodies with Germinal
Germinal co-optimizes CDR sequence (IgLM) and structure (AlphaFold-Multimer) to design epitope-targeted nanobodies de novo, achieving 4-22% binding success with 43-101 designs tested per target — no starting binder required.
- 2.TCR-mimic bispecific nanobody-based T cell engager targeting intracellular tumor antigens for cancer immunotherapy
A VHH-VHH bispecific (TCRm Bi-NbTE) simultaneously engages CD3ε and pMHC-I tumor complexes, redirecting T cells to kill tumor cells via intracellular antigens — with in vivo efficacy in CDX and PDX xenografts.
- 3.Efficient inference of non-polyreactive antibody variants dependent on local fine-tuning
Local fine-tuning of a PLM on 240k yeast-display CDR-H3 variants raises polyreactivity prediction success from 0% (base model) to 66.6% — practical demonstration that foundation models need domain adaptation for narrow drug-optimization tasks.
- 4.Discovery and optimization of a pH-responsive ultra-long-acting VHH-based growth hormone mimetic
pH-dependent VHH engineering — 227-fold difference in dissociation rate between neutral and acidic pH — extends GH mimetic activity from 6 days to >15 days in rats, achieving ultra-long-acting PK via FcRn recycling logic in a VHH format.
- 5.A multi-centre, phase 1a/1b dose escalation and expansion study of the HER2-directed antibody-drug conjugate T-Bren (BL-M07D1) in advanced breast cancer and other solid tumours
T-Bren (BL-M07D1, DAR-8, cathepsin B-cleavable, topo I payload) achieves 81.5% ORR in HER2-positive and 69.5% in HER2-low breast cancer across 253 patients; RP2D 4.4 mg/kg Q3W.
- 6.Impact of ASO conjugation and receptor binding affinity on intracellular transport of mono- and bispecific TfR- and CD98-Brainshuttle variants
ASO cargo inverts the TfR1 affinity-transcytosis relationship in Brainshuttles; a bispecific TfR1×CD98hc format rescues BBB delivery by engaging two distinct trafficking pathways simultaneously.
Paper 1 · Nature Biotechnology
Efficient generation of epitope-targeted antibodies with Germinal
Germinal co-optimizes CDR sequence (IgLM) and structure (AlphaFold-Multimer) to design epitope-targeted nanobodies de novo, achieving 4-22% binding success with 43-101 designs tested per target — no starting binder required.
Core finding
A generative pipeline combining AlphaFold-Multimer and IgLM designs functional antibody CDRs from scratch against user-specified epitopes. Tested against four targets (PD-L1, IL-3, IL-20, BHRF1), it achieved 4-22% binding success testing only 43-101 designs per antigen, with nanomolar affinities (140-560 nM) confirmed by BLI. All code and protocols are open source.
What is novel
Prior computational antibody design methods either required existing weak binders as starting points or generated binders without epitope control. Germinal introduces explicit epitope-specification — the user defines where on the target the antibody should bind — and optimizes a Pareto frontier between structural confidence and sequence naturalness without training new networks.
Limitations
Affinities (140-560 nM) are 10-100x weaker than typical clinical standards. No polyreactivity, aggregation, or thermal stability data was reported. Two of four targets showed anomalous binding kinetics or sub-threshold BLI fit quality. Results are nanobody-format only; IgG extension was not demonstrated.
Why it matters in context
Germinal joins a rapidly evolving landscape of computational antibody design tools but is distinctive in pairing structural optimization with an antibody-specific language model and making epitope targeting an explicit constraint. The 2025-2026 period has seen a shift from 'can AI design binders?' to 'can AI design binders at a specified epitope and at therapeutically relevant affinity?' — Germinal advances the epitope-control question while leaving the affinity gap open.
Paper 2 · Signal Transduction and Targeted Therapy
TCR-mimic bispecific nanobody-based T cell engager targeting intracellular tumor antigens for cancer immunotherapy
A VHH-VHH bispecific (TCRm Bi-NbTE) simultaneously engages CD3ε and pMHC-I tumor complexes, redirecting T cells to kill tumor cells via intracellular antigens — with in vivo efficacy in CDX and PDX xenografts.
Core finding
A modular bispecific VHH-VHH platform simultaneously engages CD3ε on T cells and HLA-A2-restricted peptide-MHC class I complexes presenting tumor antigens WT1 or GPC3. In multiple CDX and PDX xenograft models, the construct suppressed tumor growth, prolonged survival, enhanced T cell infiltration, and showed antigen-restricted T cell activation without treatment-related adverse effects.
What is novel
This is the first reported bispecific VHH-VHH T cell engager targeting intracellular antigens via pMHC-I — extending the addressable antigen space for T cell engagers beyond surface proteins to the full intracellular tumor proteome. The nanobody format avoids the aggregation and structural instability of scFv-based bispecifics used in conventional BiTE-class therapies.
Limitations
All in vivo data are from xenograft mouse models; syngeneic or humanized immune system models would provide a more physiologically relevant readout. HLA-A2 restriction limits potential patient eligibility. No binding affinities are reported for the pMHC-I arm. Manufacturing complexity of the VHH-VHH bispecific format is not characterized.
Why it matters in context
T cell engagers have achieved remarkable efficacy in hematologic malignancies but face fundamental limitations in solid tumors, partly due to restricted surface antigen availability. The TCR-mimic antibody concept has prior precedent with conventional IgG formats. This work's contribution is applying the concept in a VHH bispecific format, which offers manufacturing and stability advantages and makes the modular antigen-switching inherent to the platform.
Paper 3 · mAbs
Efficient inference of non-polyreactive antibody variants dependent on local fine-tuning
Local fine-tuning of a PLM on 240k yeast-display CDR-H3 variants raises polyreactivity prediction success from 0% (base model) to 66.6% — practical demonstration that foundation models need domain adaptation for narrow drug-optimization tasks.
Core finding
Absci fine-tuned protein language models on 240,000 CDR-H3 variants screened for polyreactivity via yeast display with a standardized ELISA assay (inter-experiment Pearson ≥ 0.99). Against base PLMs that achieved 0% success at identifying non-polyreactive variants in relevant sequence space, locally fine-tuned models achieved 66.6% inference success. Out-of-distribution performance on 80 clinical antibodies was preserved.
What is novel
Practical validation — in an active drug optimization campaign, not a benchmark — that local domain adaptation of PLMs is necessary and sufficient to make polyreactivity prediction deployable. The simultaneous development of a reproducible, high-throughput assay standard for polyreactivity represents the enabling infrastructure for this result.
Limitations
Model weights and training data are not publicly released. 66.6% inference success still means one-third of predicted non-polyreactive variants fail experimentally. The approach requires access to large-scale yeast display infrastructure and ELISA screening to generate training data — not trivially accessible to all labs.
Why it matters in context
This paper sits within a 2025-2026 trend of questioning when to use large pretrained biological language models versus when to fine-tune on proprietary domain data. It joins parallel work on developability prediction in establishing that the PLM → local adaptation pipeline is now a standard approach in industrial antibody optimization. The 0% → 67% gap is stark and likely to prompt other groups to report similar local fine-tuning results.
Paper 4 · mAbs
Discovery and optimization of a pH-responsive ultra-long-acting VHH-based growth hormone mimetic
pH-dependent VHH engineering — 227-fold difference in dissociation rate between neutral and acidic pH — extends GH mimetic activity from 6 days to >15 days in rats, achieving ultra-long-acting PK via FcRn recycling logic in a VHH format.
Core finding
GeneScience engineered pH-dependent growth hormone receptor (GHR) binding into a VHH antibody (pH.VHH03) by simultaneously optimizing hinge and variable regions. The result was a 227-fold difference in dissociation rate between neutral (pH 7.4) and acidic (pH 6.0) conditions. In rats, this extended pharmacological activity from 6 days to >15 days, exceeding PEGylated GH and producing a flatter IGF-1 profile at medium-to-high doses.
What is novel
First demonstration of FcRn-recycling-logic PK extension applied directly to a VHH without Fc fusion or albumin conjugation — the pH sensitivity is engineered into the VHH itself. This achieves ultra-long-acting biologics in a small single-domain format that retains the manufacturing simplicity of VHHs.
Limitations
All PK and PD data are in rats only. Species differences in GHR and FcRn biology mean primate studies are required before clinical translation can be projected. Immunogenicity of the engineered hinge region was not evaluated. Ultra-long IGF-1 exposure requires careful safety monitoring.
Why it matters in context
The long-acting GH field has been dominated by PEGylation and albumin conjugation. Engineering the half-life extension mechanism into the binding protein itself — rather than appending a half-life extension moiety — represents a conceptual advance borrowed from pH-dependent antibody engineering established in the IgG field. The translation of this approach to VHH is the novel contribution.
Paper 5 · EBioMedicine
A multi-centre, phase 1a/1b dose escalation and expansion study of the HER2-directed antibody-drug conjugate T-Bren (BL-M07D1) in advanced breast cancer and other solid tumours
T-Bren (BL-M07D1, DAR-8, cathepsin B-cleavable, topo I payload) achieves 81.5% ORR in HER2-positive and 69.5% in HER2-low breast cancer across 253 patients; RP2D 4.4 mg/kg Q3W.
Core finding
Phase 1 multicenter study of T-Bren in 253 patients with advanced breast cancer. T-Bren is a HER2-directed ADC with a cathepsin B-cleavable linker, topo I inhibitor payload (Ed-04), and DAR-8. ORR was 81.5% in HER2-positive, 69.5% in HR+/HER2-low, and 58.3% in HR-/HER2-low breast cancer. Median PFS was 18.2, 14.0, and 7.2 months respectively. RP2D established at 4.4 mg/kg Q3W; DLTs were hematologic at 6.2 mg/kg.
What is novel
Clinical validation of a DAR-8 cysteine-conjugated topo I ADC in both HER2-positive and HER2-low disease. DAR-8 payloads have historically been associated with PK challenges and aggregation risk. The efficacy data in HER2-low (especially HR-/HER2-low) add to the growing evidence that topo I ADCs extend benefit well below 3+ HER2 expression.
Limitations
Single-arm Phase 1 without a direct comparator. Patient population is exclusively from China. HER2-low subgroups are modest in size. No PK or pharmacodynamic data reported. Long-term hematologic safety with DAR-8 payload loading requires Phase 3 monitoring.
Why it matters in context
T-DXd established the benchmark in HER2 ADCs with ~79% ORR in HER2-positive breast cancer. T-Bren's Phase 1 data are competitive across all HER2 expression subgroups, supporting the view that the topo I inhibitor class — not specific payload chemistry — is the primary driver of activity in HER2-expressing tumors. Phase 3 (NCT06316531) is registered.
Paper 6 · mAbs
Impact of ASO conjugation and receptor binding affinity on intracellular transport of mono- and bispecific TfR- and CD98-Brainshuttle variants
ASO cargo inverts the TfR1 affinity-transcytosis relationship in Brainshuttles; a bispecific TfR1×CD98hc format rescues BBB delivery by engaging two distinct trafficking pathways simultaneously.
Core finding
ASO conjugation to TfR1 Brainshuttle antibodies increases binding of low-affinity binders (avidity effect) while paradoxically reducing binding of high-affinity variants, and delays BBB transcytosis in the mid-to-high affinity range. A novel bispecific TfR1×CD98hc antibody-ASO conjugate overcomes these limitations by engaging two distinct, non-competing endosomal trafficking pathways, achieving significantly enhanced in vitro transcytosis.
What is novel
Mechanistic demonstration that ASO cargo fundamentally rewires the affinity-transcytosis relationship for TfR1 shuttles — the optimal affinity window shifts after conjugation and cannot be predicted from the naked antibody's behavior. CD98hc is validated as a functional BBB receptor with a distinct trafficking pathway, enabling a bispecific strategy that is synergistic rather than redundant.
Limitations
All data are from an in vitro BBB model. Primate TfR1 expression levels differ from human, making in vivo translation uncertain. No in vivo CNS delivery efficiency or ASO knockdown efficacy data are reported. Bispecific manufacturing complexity and immunogenicity were not characterized.
Why it matters in context
TfR1-mediated Brainshuttle technology is in active clinical development for Alzheimer's disease. The mechanistic finding that ASO cargo inverts the affinity-transcytosis relationship extends prior understanding that intermediate TfR1 affinity is required for transcytosis — and shows that cargo type is an additional variable in this tuning. It is significant for the growing field of antibody-oligonucleotide conjugates for CNS indications.
Primary papers
- [1] Efficient generation of epitope-targeted antibodies with Germinal (Nature Biotechnology)
- [2] TCR-mimic bispecific nanobody-based T cell engager targeting intracellular tumor antigens for cancer immunotherapy (Signal Transduction and Targeted Therapy)
- [3] Efficient inference of non-polyreactive antibody variants dependent on local fine-tuning (mAbs)
- [4] Discovery and optimization of a pH-responsive ultra-long-acting VHH-based growth hormone mimetic (mAbs)
- [5] A multi-centre, phase 1a/1b dose escalation and expansion study of the HER2-directed antibody-drug conjugate T-Bren (BL-M07D1) in advanced breast cancer and other solid tumours (EBioMedicine)
- [6] Impact of ASO conjugation and receptor binding affinity on intracellular transport of mono- and bispecific TfR- and CD98-Brainshuttle variants (mAbs)