Pose scores, docking ranks and simulated stability are not measurements of binding, specificity or function. Three research articles and one review first published between 11 and 18 September 2026, read as of 30 September 2026, show how far peptide-binder pipelines can go before any occupancy assay: co-folding looks strong on training-overlapping complexes and weaker when they are absent; de novo sequences can be shortlisted entirely in silico; and a peptide used as a simulation probe is not a designed ligand confirmed at the bench.
Pose scores still track training overlap more than wet-lab binding
On 11 September 2026, Fonteyne and colleagues reported a computational pose-prediction benchmark, not a binding study. They compared three co-folding methods—AlphaFold-Multimer (AFM), AlphaFold3 (AF3) and Boltz-2—with two peptide-oriented docking tools, AutoDock-CrankPep (ADCP) and HADDOCK. The PepPro set contains conformationally flexible protein–peptide complexes that are represented in the training sets of all three co-folding methods. Under that condition, co-folding outperformed docking in pose accuracy and ranking. An independent test set classified by redundancy with those training sets reversed the picture: without related complexes in the training data, HADDOCK outperformed every co-folding method evaluated. Boltz-2 was the most accurate method on PepPro but showed reduced generalization on non-redundant complexes. AFM and AF3 were more robust on independent complexes, which the authors treat as better suited to de novo peptide-binder design. Interface predicted template modelling (ipTM) showed strong early enrichment for ranking poses, and peptide-specific predicted local distance difference test (pLDDT) tracked structural accuracy. None of those metrics is an affinity, competition assay or off-target panel.
That gap is the practical warning in an 18 September 2026 review by Riccabona et al.. Generative models have enabled computationally guided high-affinity binders against diverse targets, and experimental success has been reported in that broader literature. The same review argues that the confidence metrics used to filter designs remain optimized for static protein interfaces and can fail on underrepresented or conformationally complex targets. Ensemble-based evaluation is proposed as one route to higher success rates. Fold-switching scaffolds and molecular glues realized through engineered cyclic peptides are flagged as ways to expand function beyond single-target binders. Those are design strategies, not substitutes for measuring occupancy or selectivity.
Design pipelines can stop before any binding experiment
A 16 September 2026 computational study by Ahmed, Ismail and Suleiman shows how far a binder paper can go without a binding assay. The intended target is the ATP-binding region of AbcA, an efflux transporter in Aspergillus fumigatus. After multiple sequence alignment to locate conserved ATP-site sequence, Protein Generator predicted a de novo sequence and structure from the consensus conserved region, and ProteinMPNN designed 60 de novo protein sequences. Cell-penetrating peptide motifs were added and scored with C2Pred; allergenicity was predicted with AllerCatPro and thermal stability with DeepSTABp, yielding ten highly thermally stable designs from 7200 generated binders. Models from OmegaFold were energy-minimized, docked (GRAMM and PDBSum) and simulated; the abcA-2795 complex was the most stable conformation in those molecular dynamics runs. Predicted properties for that design include non-allergenicity, cell-penetration potential and high thermal stability. The abstract states that further experimental validation is required to confirm efficacy and safety. No binding, specificity or antifungal functional assay of the designed sequences is described.
Han, Wray, Blount and Wang (18 September 2026) used a different peptide role: bovine pancreatic trypsin inhibitor (BPTI), an experimentally validated permeant, as a probe in electric-field-steered molecular dynamics of Escherichia coli MscL, in order to capture a partially open channel conformation for later binder design. Structural quality was assessed with QMEAN and energetics with MM/PBSA. Virtual screening then targeted a curated small-molecule antibiotic library from CO-ADD rather than a de novo peptide library. Eight ligands were computationally identified as stable interactors predicted to inhibit channel closure. Four commercially available compounds were evaluated in cell viability assays; the results suggested that compound abJ0y affects bacterial growth, with reduced growth in MscL-expressing cells. That is a cellular phenotype for catalog small molecules, not affinity or specificity data for an AI-designed peptide.
What still has to be shown
The useful split is not computational versus experimental in the abstract, but which claim a result can support.
| Claim you want | Evidence that actually supports it | What these papers supply |
|---|---|---|
| The pose is geometrically plausible | Co-folding or docking ranking, preferably on non-redundant complexes | Fonteyne: PepPro plus a redundancy-classified independent set |
| The design is worth synthesizing | Orthogonal filters (ipTM, peptide pLDDT, MD stability), with known failure modes | Fonteyne metrics; Ahmed in silico shortlist |
| It binds the intended site | Direct binding or competition, not docking energy alone | Not reported for de novo designs here |
| Binding is selective | Related proteins, unbound states, or decoy sites | Not reported |
| Binding does something | Target-linked function with controls | Han: viability for four commercial compounds; Ahmed: none |
When reading a weekly crop of design papers, a short checklist keeps predictions from being promoted to hits:
- Is the target conformation a determined structure or a simulation snapshot?
- Were confidence metrics tested off the method’s training overlap?
- Is there a direct binding measurement, or only docking and MD?
- Was specificity addressed at all?
- If a cellular phenotype is shown, is it for the designed peptide and mapped to the intended target?
One question these abstracts leave open is whether ipTM and peptide pLDDT filters that enrich poses on training-overlapping complexes still enrich peptides that bind in a biochemical assay when the target is conformationally flexible and absent from the training set. Pose ranking on PepPro cannot answer that, and neither can an MD-stable de novo model that has not been tested for occupancy.
This is not a survey of the whole peptide-design field. It is a reading of four mid-September 2026 abstracts, as of 30 September 2026, through the narrow lens of what was predicted versus what was measured.
Frequently Asked Questions
Does a high ipTM mean a designed peptide binds?
No. In the 11 September 2026 pose-prediction benchmark, ipTM showed strong early enrichment for ranking poses, and peptide-specific pLDDT tracked structural accuracy. Those metrics are not experimental binding, affinity or specificity measurements.
Were the AbcA de novo binders tested in the lab?
Not in the retrieved abstract. The 16 September 2026 study ranked designs with docking and molecular dynamics, highlighted abcA-2795 as the most stable simulated complex, and stated that further experimental validation is required to confirm efficacy and safety.
Explore Further
Browse our research peptide catalog and review third-party lab reports & COAs for published batches.
---
Research use only. The information above is provided for educational and laboratory research purposes only. The compounds discussed are not approved for human or veterinary use, diagnosis, treatment, or the prevention of any disease. Nothing here is medical advice.