The short answer
Modern all-atom models can represent substantially more chemistry than the original AlphaFold2 or ESMFold protein-sequence workflows, so βAlphaFold cannot handle modificationsβ is now too broad.[1,2,3,4,5] AlphaFold3 and related systems can generate useful hypotheses for modified peptides, and specialized cyclic-peptide methods show measurable progress, including a 2026 benchmark of HighFold4.[3,6,7] But accepting a chemical graph does not demonstrate accurate geometry, correct stereochemistry, a calibrated confidence estimate, or the populations of solution conformations. Published success on folded proteins, covalent modifications, or constrained cyclic peptides does not establish a validated error rate for a long, lipidated, PEG-linked peptide in its free, albumin-bound, receptor-bound, and self-associated states. I would use these models to formulate and discriminate structural hypotheses, not to certify chemical identity, solubility, aggregation behavior, or a lot's quality.
What is genuinely left after the later courses
These questions recur because they combine several distinct problems. BCHM601 asks whether AlphaFold2 or ESMFold can be fine-tuned for heavily modified peptides; BIOL502 asks how successors handle PEG linkers and Aib; BCHM698 asks about C20 fatty-diacid modifications. I preserved those extracted questions verbatim above. The last question's description of AlphaFold2 as a generative model should not obscure the architectural distinction: AlphaFold3 introduced the diffusion-based all-atom structure-generation framework discussed here.[3]
Later coursework already answers the generic caution. BCHM611 distinguishes structure prediction, inverse folding, and backbone generation and says confidence is not folding free energy or binding affinity. BCHM604 warns that absent coordinates can mean unresolved atoms rather than absent chemistry. BCHM700 states that pLDDT is not a directly measured fluctuation amplitude, while BCHM799 explicitly separates predicted structures from solubility, specificity, binding, and function. PDQC815 adds that analytical dilution can change a lipidated peptide's association state.
Those are substantive answers, not gaps. What remains is a current, chemistry-specific capability audit: which inputs are representable, what was actually benchmarked, which metrics can hide failure, and what evidence would justify transferring performance to the modified peptides of interest. I inspected the actual later sections and swept the later parsed corpus, rather than treating a few retrieval hits as proof of absence. The scope sidecar records those distinctions and the relevant quotations.
Define the prediction before choosing the model
βThe structure of a peptideβ is under-specified. It could mean backbone geometry of a receptor-bound segment, orientation of a lipid relative to a membrane, the preferred aqueous conformations of a linker, or an equilibrium distribution of oligomers. Those are different targets, with different experimental observables. A correct bound helix can coexist with an incorrect lipid orientation, and neither result determines how much material stays soluble in a particular formulation.
I would write the question as: given an exact chemical species and an environment, predict a specified structural observable with a specified error measure. The chemical species includes residue identities, stereochemistry, terminal chemistry, attachment positions, linker connectivity, and cyclization. The environment includes interacting partners and relevant physical conditions. The observable might be a set of receptor contacts, a distance distribution, or conformational populationsβnot automatically one coordinate file.
Aib illustrates why replacing a noncanonical residue with a familiar letter is not innocuous. Its two Ξ±-methyl substituents make it chemically different from alanine and remove the Ξ±-stereocenter; its conformational preferences need not match the proxy. A lipidated Lys is not simply ordinary Lys plus an annotation in the figure legend. A short, chemically defined oligoethylene-glycol spacer is also not equivalent to a polydisperse PEG conjugate, for which there is no single chain length to model. These are chemical-identity distinctions before they are machine-learning limitations.
Lau and colleagues' semaglutide discovery study is a useful experimental reminder: the fatty-acid moiety and linker chemistry were key variables affecting albumin affinity and receptor potency.[11] That finding supports treating the modification as part of the molecule, not disposable metadata. It does not establish a universal direction or magnitude of effect for a different peptide, linker, or receptor.
From protein-sequence predictors to chemical graphs
AlphaFold2 demonstrated highly accurate protein structure prediction in CASP14, using a learned architecture informed by sequence alignments and structural data.[1] ESMFold demonstrated direct inference of protein structure from primary sequence using a protein language model and enabled large-scale metagenomic predictions.[2] Their achievements are real. Neither benchmark, by itself, supplies an accuracy estimate for arbitrary PEG chains, unusual covalent attachments, or synthetic stereochemical substitutions.
Fine-tuning is therefore not simply uploading several modified structures and retaining the old interface. The model must distinguish the atoms and bonds whose effects it is meant to learn, output their geometry, and receive an objective that assesses those atoms. An input pipeline that maps different modified molecules to the same canonical sequence has discarded information before learning begins. My inference is that no amount of downstream confidence scoring can recover which chemical graph the user intended from that identical representation alone.
AlphaFold3 changed that representational situation. Its paper describes joint prediction of proteins, nucleic acids, small molecules, ions, and modified residues using a diffusion-based architecture.[3] The current local-code input documentation supports modified residues, covalent bondedAtomPairs, and user-defined chemical components through userCCD or userCCDPath.[4] Those facilities make many previously awkward chemical specifications expressible. They are not a promise that every specification is processed identically by every release or hosted interface.
There is a practical distinction between a ligand SMILES and an explicitly bonded component. The documentation says a custom molecule that must bond to another entity needs appropriately named atoms through a chemical-component definition; a SMILES-only representation does not provide the required atom naming for that interface.[4] I would verify the parsed molecular graph and output atom inventory, not merely that JSON was accepted. A disconnected fatty acid placed near a peptide is not a model of covalent lipidation.
RoseTTAFold All-Atom provides another important precedent: its published representation combines residue-level descriptions for standard biopolymers with atom-level descriptions for other groups, including covalent modifications.[5] Its experimental protein-design demonstrations concern defined ligand-binding tasks. They do not validate arbitrary lipidated-peptide solution ensembles. This is why βall-atomβ should be read as a representation capability, not as an all-purpose accuracy guarantee.
ββββββββββββββββββββββββββββββββββββββββββββββββ
β EXACT CHEMICAL GRAPH + SPECIFIED ENVIRONMENT β
ββββββββββββββββββββββββ¬ββββββββββββββββββββββββ
β
REPRESENTATION CHECK
ATOMS / BONDS / CHIRALITY / ATTACHMENTS
β
PREDICTED COORDINATES
+ MODEL CONFIDENCE
β
CHEMICAL VALIDITY CHECK
β
TASK-MATCHED EXPERIMENTAL COMPARISON
β
BOUND POSE? SOLUTION ENSEMBLE?
β β
CONTACT ACCURACY POPULATIONS + CONDITIONS
ββββββββββββ¬ββββββββββ
β
NEITHER ALONE ESTABLISHES SOLUBILITY
What the published benchmarks actually test
The AlphaFold3 paper is stronger than a generic claim that modifications are supported. It reports tests of covalent modifications, including bonded ligands, glycosylation, and modified polymer residues, and defines successful predictions using a pocket-aligned RMSD criterion.[3] It also provides phosphorylation examples where explicitly representing the modification improves predictions compared with substituting the unmodified parent. That is evidence that including chemistry can matter to accuracy, not merely to file syntax.
But its evaluation subsets are not interchangeable. The paper states that bonded-ligand and glycosylation datasets were filtered for experimental model quality and were not filtered by homology to the training dataset; applying a low-homology filter left very small subsets. The modified-residue dataset used low-homology polymer chains.[3] I cannot take a pooled covalent-modification result and relabel it βperformance on unseen lipidated peptides.β The split, chemical class, atom selection, and similarity to training structures determine how far the result transfers.
The paper's PoseBusters analysis used a separately trained model with a 2019 structural-data cutoff, because the usual 2021 cutoff would overlap the evaluation period.[3] This is good evaluation discipline, but it also exposes a common replication mistake: running currently distributed weights on an old benchmark is not necessarily reproducing the historical blind test. I would archive weights, training cutoff, templates, and inference settings alongside the results. A benchmark name without a model version is an incomplete provenance record.
The 2025 study AlphaFold3 for Noncanonical Cyclic Peptide Modeling directly addresses a more relevant class.[6] Its accessible abstract describes hierarchical evaluation of all atoms, CΞ± atoms, and unnatural-residue atoms, for both monomers and complexes, plus confidence metrics and the effects of alignments and templates. I verified that paper and read that abstract, but could not inspect its full publisher text. I therefore do not report numerical performance, an optimal confidence cutoff, or detailed conclusions that the accessible abstract does not provide.
A newer accessible primary study prevents the inquiry from stopping at that limitation. HighFold4, published in 2026, modifies AlphaFold3's handling of user-defined cyclic connectivity without retraining the base model.[7] Its dataset comprises 62 canonical monomers, 32 canonical complexes, 68 noncanonical monomers, and 17 noncanonical complexes. The authors evaluate CΞ±, backbone, all-atom, and unnatural-residue RMSDs and separately examine ring closure. That is a meaningful advance in task-specific assessment.
Its most useful lesson is not a headline average. A prediction can resemble a ring in coordinate space while lacking the intended chemical closure; RMSD alone can miss this error.[7] HighFold4 reports complete cyclization success in the experiments described, while explicitly acknowledging remaining chirality errors and other abnormalities inherited from the underlying architecture. I would not translate βring closedβ into βall stereocenters and conformational populations correct.β The limitation section is part of the evidence, not an optional footnote.
Transfer to lipidated linear peptides remains unestablished by that study. Its curated structures are cyclic monomers and complexes assembled from PDB-derived sources, not a prospective panel of long acylated peptides with PEG-linker ensembles under varied solution conditions.[7] Sequence deduplication is useful, but it is not the same thing as a demonstrated temporal holdout from the base model's structural training set. I found no basis in the examined dataset description to assign PP a blind success probability from those averages.
What a credible fine-tuning claim would require
I would distinguish representational repair from learning new structural preferences. HighFold4 reports that its tested AlphaFold3 implementation accepted intra-polymer bond declarations without fully carrying that information through the relevant feature pathway; its changes address that gap.[7] That is a version-specific implementation result, not proof that every bond declaration in every current AlphaFold3 release is ignored. It demonstrates why checking documentation, parsed graphs, generated features, and resulting chemistry are separate steps.
For genuine fine-tuning, the training set would need chemically unambiguous structures with usable evidence for the modified atoms, not just canonicalized sequences paired with coordinates of an experimentally resolved core. I would annotate experimental method, environment, unresolved atoms, attachment chemistry, and scaffold relatedness. Missing PEG or lipid density is not a label for one arbitrarily chosen conformation. Simulated structures could provide auxiliary hypotheses, but training on them cannot establish experimental accuracy beyond the assumptions of the simulation.
The loss and evaluation would need to distinguish incorrect stereochemistry or connectivity from conformational disagreement, and distinguish core accuracy from appendage accuracy. Family-held-out tests and an untouched prospective panel would then test generalization; confidence calibration would require its own held-out assessment. Improvements should be compared against a fixed baseline with equal sampling budgets and identical available templates. These are my proposed conditions for a persuasive claim, not evidence that I have trained such a model or that a sufficient public lipidated-peptide dataset already exists.
Boltz is progress, not a shortcut around scope
Boltz-2's primary preprint explicitly extends biomolecular modeling to structure and binding-affinity prediction.[8] Its affinity claims concern small-moleculeβprotein tasks; I should not repeat the older blanket statement that modern structural models never predict affinity. The accurate statement is that a model's particular affinity head must be evaluated for its intended ligand class, separately from coordinate accuracy and separately from solution properties.
The current Boltz documentation allows modified residues, cyclic polymers, and explicit bond constraints, with representation-specific restrictions.[9] It also states that its affinity computation takes a small-molecule ligand chain, with an atom limit and a recommendation against ligands substantially larger than those represented in affinity training. In the documentation inspected, the maximum is 128 atoms under its stated RDKit counting convention, while the training-related recommendation is 56 atoms.[9] These are implementation and applicability limits, not biological boundaries.
A peptide-lipid conjugate does not become an appropriately validated small-molecule affinity task because it can be serialized as a ligand. Renaming an entity changes parser behavior; it does not supply benchmark evidence. Similarly, a graph that loads successfully may still encounter inadequate sampling or misleading confidence. The relevant question is not whether the software crashes, but whether the resulting quantity has been calibrated for this chemical and biological domain.
The unresolved lipid in an experimental structure is the crucial example
The tirzepatide structural study by Sun and colleagues combines cryo-EM, molecular dynamics, and pharmacological experiments.[10] Its methods state that the Lys20-linked fatty-acid chain and C-terminal residues 33β39 were not observed in the cryo-EM structures and were modeled for subsequent simulations. This is precisely the distinction I need when judging a predicted lipid pose: some coordinates in a completed molecular model may be hypotheses added to an experimental core.
The simulations explored multiple lipid-linker conformations and receptor-dependent interactions.[10] Those results are simulation-based evidence under the modeled conditions, not direct images of a unique fatty-acid position. The experimental receptor-bound peptide core can still be informative. What fails is treating an unresolved appendage as an experimentally established reference simply because it appears in a visualization of the completed system.
I would therefore separate three atom sets in any comparison: experimentally supported atoms, modeled missing atoms, and atoms omitted from the reference entirely. Scoring only the peptide core cannot validate an unobserved lipid. Penalizing a flexible linker against one arbitrarily completed reference can be equally misleading. An apparently comprehensive all-atom RMSD can hide both errors if the provenance of the reference coordinates is ignored.
This also explains why membrane, albumin, and free-solution questions cannot be collapsed into one pose. Different partners and physical environments change which conformations are stabilized. A model of a receptor complex may correctly constrain the helical segment while leaving the lipid ensemble uncertain. It is not, merely by being the same chemical sequence, a model of association in a storage vial.
Confidence, sampling, and measured accuracy are different quantities
AlphaFold3's own limitations include chirality errors, clashes, hallucinated structure in disordered regions, and incomplete conformational coverage.[3] Its diffusion procedure generates different outputs, but those samples are not automatically Boltzmann-weighted trajectories of the physical molecule. A distribution of model guesses is a distribution of generated hypotheses, not a measured physical ensemble, unless its relationship to physical populations has been established. Repeated predictions can agree because the model repeatedly favors the same learned pattern, not because the experimental ensemble is narrow.
pLDDT, PAE, pTM, and interface scores answer model-specific confidence questions. They do not become measured accuracy until compared with appropriate reference data. A confidenceβaccuracy relationship learned or tested on well-represented structures may shift for unusual stereochemistry, novel attachments, or flexible linkers. The modified-peptide study's decision to assess confidence separately, and HighFold4's acknowledged chemical failures, reinforce the need to test calibration rather than assume it.[6,7]
A useful evaluation should report at least chemical validity, backbone accuracy, modification-specific accuracy, interface accuracy where relevant, and the fraction of targets that fail each criterion. Report the top-ranked prediction separately from the best of many samples. Selecting the closest structure after seeing the answer measures sampling potential, not the accuracy of a deployable ranking procedure. The AlphaFold3 paper explicitly documents its sampling and ranking choices, which is a standard worth retaining.[3]
Here is a worked uncertainty example, not a published benchmark result. Suppose ten independent, representative lipidated-peptide targets all pass a predefined structural test. Under a binomial model, the one-sided 95% upper confidence bound for the failure probability is 1 β 0.05**(1/n): 25.89% for ten zero-failure trials. With sixty zero-failure trials, the bound is 4.87%. A handful of attractive examples provides much less assurance than β100% successβ suggests.
for n in (10, 60):
upper_failure = 1 - 0.05**(1/n)
print(n, 100*upper_failure)
The independent-target assumption is substantial. Ten nearly identical analogues of one scaffold are not ten independent demonstrations across chemical space. Randomly splitting analogues can leak shared geometry across training and testing. For PP's intended use, I would hold out families by scaffold and modification class where feasible, document sequence and ligand similarity, and report failures rather than discard structures with difficult reference coverage. These are proposed evaluation principles, not a benchmark I have run.
Structure is not an ensemble, and an ensemble is not solubility
Even an accurate low-energy conformation says little about its population unless competing states are represented. For two states with a free-energy difference, their equilibrium population ratio depends exponentially on that difference. A hypothetical 5 kJ/mol error in the relative free energy at 298.15 K changes the population ratio by exp(5000/(RΓ298.15)), approximately 7.52-fold. I executed this arithmetic; it is an illustration of sensitivity, not a claimed error estimate for any named predictor.
Molecular dynamics can investigate environment and conformational alternatives, but it adds force-field, parameterization, protonation, and sampling assumptions. A noncanonical residue or lipid-linker junction needs chemically appropriate parameters, and repeated trajectories can remain trapped in the same basin. Agreement between a neural prediction and a simulation initialized from that prediction is not fully independent confirmation. The tirzepatide study is informative precisely because its simulation and experimental components are identifiable rather than blended into one evidence label.[10]
Solubility requires additional thermodynamic and kinetic information about the solution, associated species, and competing phases. Aggregation may depend on concentration, surfaces, history, and nucleation, not just a monomer's preferred shape. A structural score does not supply those missing variables. PDQC815 already teaches that the analytical procedure itself can perturb association; this inquiry adds that a prediction run also defines a particular modeled system and cannot silently stand in for all of them.
What this means for Peptide Partners
I would maintain a chemical-identity manifest for any modeled material: exact sequence, noncanonical residues, stereochemistry, attachment atom names, linker definition, terminal states, and the corresponding graph after preprocessing. Archive the model version, weights, templates, seeds, partner constructs, and explicit omissions. Label a canonicalized proxy as a proxy. Never let a polished structural image obscure that the lipid or PEG segment was dropped before inference.
For a real research question, choose the smallest defensible prediction claim. A contact hypothesis in a receptor-bound core may be supportable when a free-solution PEG ensemble is not. Compare alternatives and ask which measurement would distinguish them. Appropriate experimental structural constraints and independently established chemical identity should guide interpretation; a high-confidence prediction cannot replace LC-MS characterization, stereochemical evidence, or concentration-dependent assessment of association.
A useful PP report would say βthis model proposes a contact involving these experimentally unresolved atomsβ rather than βAI confirms the structure.β If a supplier invokes prediction to support quality, ask which chemical features were represented and which experimental data validate the claim. Neither a coordinate file nor an affinity score certifies a physical lot. These are research-use scientific judgments, without human administration recommendations or clinical efficacy claims.
Questions to bring to the lab
- Which specific observable do we need: receptor-bound backbone, modification orientation, solution-state distances, association, or something else?
- Does the parsed input preserve every stereocenter, linker bond, terminal group, and attachment position, and how will we audit the output graph?
- Which reference coordinates are supported by experimental density or restraints, and which lipid or PEG atoms were modeled into missing regions?
- What chemically matched, genuinely held-out targets would establish our top-ranked prediction accuracy rather than best-of-many sampling potential?
- Can confidence discriminate failure for our modification classes, including incorrect chirality or bond topology, rather than merely rank plausible-looking cores?
- Which orthogonal measurements could distinguish competing conformations without assuming that a single structure or diluted sample represents the original solution?
- What result would make us abandon the structural hypothesis, and are our proposed validation data independent of the information used to generate it?
What I still don't know
I do not have a prospectively validated error rate for long lipidated, PEG-linked peptides matching PP's materials. The primary studies inspected establish progress in related structural tasks, not a universal predictor of these molecules' environmental ensembles. My overall confidence is medium because the capability landscape is changing quickly and the relevant chemical space remains sparsely benchmarked, even though the distinctions between representation, confidence, and validation are firm.
The 2025 noncanonical-cyclic benchmark was accessible only at abstract level in this audit; I have not invented its numerical findings. A 2026 author-page announcement of a broader modified-peptide benchmark lacked a linked paper and complete verifiable methods, so I excluded it from the evidentiary bibliography. Current repository documentation is archived with revision identifiers; local-code support should not be projected onto a hosted service without checking that service separately. These access and scope limits remain visible rather than being converted into claims that no relevant work exists.
References
-
Jumper John, Evans Richard, Pritzel Alexander et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596(7873):583-589. doi:10.1038/s41586-021-03819-2
-
Lin Zeming, Akin Halil, Rao Roshan et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science. 2023;379(6637):1123-1130. doi:10.1126/science.ade2574
-
Abramson Josh, Adler Jonas, Dunger Jack et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024;630(8016):493-500. doi:10.1038/s41586-024-07487-w
-
google-deepmind/alphafold3 contributors. AlphaFold 3 Input. Official repository documentation; revision
9b30e9aa4b232c32a40fe2f3f22ee311f2be455c. Source. Accessed 2026-10-15. -
Krishna Rohith, Wang Jue, Ahern Woody et al. Generalized biomolecular modeling and design with RoseTTAFold All-Atom. Science. 2024;384(6693):eadl2528. doi:10.1126/science.adl2528
-
Zhang Chengyun, Wang Wentong, Zhu Ning et al. AlphaFold3 for Noncanonical Cyclic Peptide Modeling: Hierarchical Benchmarking Reveals Accuracy and Practical Guidelines. Journal of Chemical Information and Modeling. 2025;65(18):9777-9789. doi:10.1021/acs.jcim.5c01393
-
Zhang Chengyun, Wang Wentong, Zhu Renjie et al. HighFold4: extending AlphaFold3 to accurate cyclic peptide conformation prediction via custom chemical connectivity. Briefings in Bioinformatics. 2026;27(5):bbag505. doi:10.1093/bib/bbag505
-
Passaro Saro, Corso Gabriele, Wohlwend Jeremy et al. Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction. bioRxiv (preprint). 2025. doi:10.1101/2025.06.14.659707
-
jwohlwend/boltz contributors. Prediction (Boltz documentation). Official repository documentation; revision
296ec7e93f3e098b0a120f0f2b966f6053bf3135. Source. Accessed 2026-10-15. -
Sun Bingfa, Willard Francis S., Feng Dan et al. Structural determinants of dual incretin receptor agonism by tirzepatide. Proceedings of the National Academy of Sciences. 2022;119(13):e2116506119. doi:10.1073/pnas.2116506119
-
Lau Jesper, Bloch Paw, SchΓ€ffer Lauge et al. Discovery of the Once-Weekly Glucagon-Like Peptide-1 (GLP-1) Analogue Semaglutide. Journal of Medicinal Chemistry. 2015;58(18):7370-7380. doi:10.1021/acs.jmedchem.5b00726