Ubi Titer Issue #11
Structure-Guided Design: Degrading Fusion Kinases, Oxidation Risk, and Docking Rank
Four papers this week use predicted structure to decide what to build: ternary complex modeling picks PROTAC linkers that hold up against resistance mutations, local electrostatics picks which antibody residues to mutate for oxidation stability, and learned representations decide which peptide conformations and docking poses to trust.
- targeted protein degradation
- PROTAC
- drug resistance
- antibody developability
- tryptophan oxidation
- peptide structural plasticity
- antibody-antigen docking
- protein language models
The field
This Week in Biologics
Paper 1 · Proceedings of the National Academy of Sciences
AI-driven PROTAC design overcomes oncogenic resilience by eliminating the CLIP1-LTK fusion protein
The first degrader built against the CLIP1-LTK fusion clears the whole protein at picomolar potency and keeps working against resistance mutations that raise kinase inhibitor potency by more than a thousandfold.
Core finding
DCL05 is the first reported CLIP1-LTK-targeted PROTAC. It degraded the fusion at DC50 33.8 ± 4.5 pM with Dmax above 99% and showed a proliferation IC50 of 0.16 nM. AlphaFold3 ternary-complex models, RTM Score, and AutoDock Vina helped prioritize six of 17 linker candidates for synthesis. Against the engineered G596R/S600F/G663A triple mutant, proliferation IC50 values were 5,576 nM for lorlatinib, 613.2 nM for gilteritinib, and 30.2 nM for DCL05. In mutant Ba/F3 xenografts, DCL05 suppressed tumor growth while the two inhibitors had no measurable efficacy.
What is novel
Degradation is well matched to a fusion in which CLIP1-driven assembly complements LTK kinase activity. The weaker degradation of NPM-ALK is consistent with a role for ternary-complex geometry in selectivity, but the study does not isolate geometry as the sole causal factor.
Limitations
Resistance substitutions were inferred from ALK and have not been observed clinically in CLIP1-LTK tumors. Efficacy was demonstrated only in engineered Ba/F3 systems and xenografts. DCL05 potency still fell substantially against mutants. Full-length LTK biology and the safety of systemic degradation remain incompletely understood. Xenograft efficacy used n=6 per group and IHC n=3.
Why it matters in context
Published online August 6, 2026, by Chen, Duan, Zhong, Ge, Zhao, Ye, Sun, Li, Kang, Dong, Che, Hou, and Pan. It is useful as a degrader-design methods reference, especially for structure-guided linker prioritization. It is not presently a reusable tool or broad dataset release; proteomics data are deposited in iProX.
Paper 2 · mAbs
Local electrostatics governs tryptophan oxidation and enables rational stability engineering in antibodies
A two-parameter rule combining solvent exposure and local electrostatic potential predicts antibody CDR tryptophan oxidation risk with 79% accuracy, and distal charge-reversing mutations cut oxidation roughly in half without touching the liability residue.
Core finding
An exposure-only rule had 91% sensitivity, 46% specificity, and 55% accuracy for Trp oxidation. Requiring both SASA above 50 Ų and negative local electrostatic potential increased specificity to 87% and accuracy to 79%, while lowering sensitivity to 47%. The two-feature rule correctly classified all 10 CDR Trp sites in a blind panel of eight clinical-stage IgG1 antibodies. A full random-forest model reached 84% accuracy but only 40% sensitivity.
What is novel
Distal charge mutations 10.5–18.6 Å from the Trp reduced oxidation by about 50% on average in four of five antibodies. Binding was retained in only four of eight engineered variants overall. In the anti-CD33 case, one variant reduced oxidation by 27.2% with a 1.4-fold affinity loss; a stronger 49.9% reduction came with a 20.7-fold affinity loss. Direct W96F caused more than a 1,000-fold affinity loss.
Limitations
AAPH forced oxidation may not reproduce real-time storage pathways. The blind-panel structures were predicted with DeepAb, and descriptors were averaged over molecular-dynamics trajectories; the paper does not validate a single-static-structure shortcut. The main 187-antibody dataset is proprietary. The 10-site blind panel is small, and sensitivity remains limited.
Why it matters in context
Published August 12, 2026, by Lenka, Mehta, Stephens, Seeger, Liang, Watkins, Zarzar, Jafari, Puno, Azumaya, Kelly, Wu, Chiu, Hazen, Wu, Irudayanathan, Alavattam, Kelley, and Izadi. The SASA-plus-electrostatics concept is a promising developability endpoint, but a production implementation should reproduce the published structure and trajectory workflow before testing faster static approximations. The article is CC BY-NC 4.0, so commercial reuse needs license review.
Paper 3 · bioRxiv (preprint)
Peptide structural plasticity is predictable from sequence and environment
ApexFold predicts how a peptide's secondary-structure composition redistributes across water, co-solvent, and membrane-mimetic environments, outperforming static single-conformation structure predictors on that task.
Core finding
ApexFold predicts environment-conditioned helix, beta-like, and unstructured fractions from peptide sequence and medium descriptors. Training used CD measurements from exactly 1,187 peptides in water, methanol/water, TFE/water, and SDS micelles. The two temporal panels contained 206 and 161 peptides. Plasticity Pearson r was 0.92 and 0.87 on those panels; pooled water-to-medium helix-shift correlations were 0.434, 0.472, and 0.546 for methanol, TFE, and SDS.
What is novel
Structural plasticity is a useful design variable distinct from one static structure. AlphaFold3 and Boltz-2 served as environment-free static references, not fully task-matched competitors. High directionality values partly reflect global solvent trends, so shift magnitude and plasticity ranking are more informative.
Limitations
Structure-balanced macro accuracy was 30.1%, with mixed-state peptides hardest. SDS micelles are heterogeneous membrane mimetics. CD deconvolution is uncertain for beta-rich and mixed states, and the three-state target groups several irregular conformations as unstructured. This is a preprint and needs external validation.
Why it matters in context
Posted August 11, 2026, by Torres, Cao, and de la Fuente-Nunez. ApexFold is a strong candidate for a peptide-prioritization wrapper, with the expectation that it ranks sequences and environments for experimental follow-up rather than replacing synthesis or CD. The repository at https://gitlab.com/machine-biology-group-public/apexfold uses the MIT license; the preprint is CC BY-NC-ND 4.0.
Paper 4 · Journal of Chemical Information and Modeling
ARID-sf: A Physics-Informed Deep Learning Scoring Function to Improve Antibody-Antigen Docking Model Ranking
ARID-sf combines classical force-field physics with protein language model embeddings to rerank antibody-antigen docking poses, nearly doubling success rates over raw physics-based scoring while running far faster than existing antibody-specific alternatives.
Core finding
ARID-sf uses OPLS-UA energy components, structural and contact features, and ESM-C embeddings to rerank HADDOCK3 poses. It was trained on more than 1.5 million models and evaluated on 806 cases and more than 700,000 poses. On the 672-case complete set, top-1 success was 0.46 versus 0.39 for HADDOCK scoring; top-10 success was 0.81 versus 0.66. Scoring 132,000 models took about one hour versus about three days for AbEpiScore on the same hardware.
What is novel
The method combines classical force-field physics with learned sequence representations and improves ranking without the runtime of earlier antibody-specific alternatives. Its advantage persists on the difficult unbound set, but the small absolute gain there underscores that reranking cannot recover correct poses that docking never sampled.
Limitations
On the 67-case unbound set, only 39 docking runs generated any correct model. ARID-sf top-1/top-10 success was 0.15/0.20 versus 0.12/0.17 for HADDOCK scoring, leaving absolute performance low. Sampling, especially under conformational change, remains the main bottleneck. CDR-H3-specific performance was not isolated.
Why it matters in context
Published online July 27, 2026, by Grandguillaume, Barroso da Silva, and Etchebest. ARID-sf is a strong MCP candidate because its code and trained weights at https://github.com/DSIMB/ARID-sf are MIT-licensed and designed for HADDOCK outputs. It should be benchmarked against alternatives such as DeepRank-Ab before selecting a default scorer.
Primary papers
- [1] AI-driven PROTAC design overcomes oncogenic resilience by eliminating the CLIP1-LTK fusion protein (Proceedings of the National Academy of Sciences)
- [2] Local electrostatics governs tryptophan oxidation and enables rational stability engineering in antibodies (mAbs)
- [3] Peptide structural plasticity is predictable from sequence and environment (bioRxiv (preprint))
- [4] ARID-sf: A Physics-Informed Deep Learning Scoring Function to Improve Antibody-Antigen Docking Model Ranking (Journal of Chemical Information and Modeling)