Visit Website →
← Back to blog

6-Step Lab Pipeline for Immunology Peptide Models with COAs

October 1, 2026
6-Step Lab Pipeline for Immunology Peptide Models with COAs

For epitope discovery, combine sequence-based binding predictors, a multimodal immunogenicity model, and targeted mass spectrometry confirmation, then validate with functional T-cell assays. This ensemble approach, anchored by tools like NetMHCpan, NetMHCIIpan, ImmunoStruct, and nf-core/epitopeprediction, filters computational noise before committing lab resources to ELISpot or ICS confirmation.

TL;DR:

  • Sequence-based predictors like NetMHCpan quickly narrow large candidate pools with percentile rank cutoffs, typically below 2% for strong binders.
  • Incorporating immunogenicity models such as DeepImmuno or PanPep after binding prediction helps prioritize peptides with higher chances of T-cell recognition.
  • Structural modeling with tools like ImmunoStruct or AlphaFold is best reserved for top candidates, especially when sequence ambiguity or noncanonical peptides are involved.
  • Confirming peptide presentation with mass spectrometry and validating immunogenicity through functional assays like ELISpot or ICS is essential for reliable epitope identification.
  • Standardized, documented pipelines and verified peptide batch quality are vital for reproducibility and accurate interpretation in immunology research.
Peptasticlabs
Source Peptides With Documented Purity
Peptastic Labs provides research-grade peptides for immunology studies, with independently tested compounds and Certificates of Analysis available on request.
Explore research peptides

Table of Contents

Model classes used for peptide immunogenicity prediction

Immunology peptide models fall into three functional categories, each answering a different question in the discovery pipeline. Sequence-based binding predictors estimate whether a peptide will physically dock into an MHC groove. Immunogenicity and TCR-binding predictors estimate whether, once presented, a T cell will actually recognize the peptide. Structure-aware multimodal models add a third layer: they incorporate three-dimensional conformation to refine predictions where sequence alone is ambiguous.

Three layers of peptide prediction

Sequence-based predictors, the oldest and most widely deployed class, report results in one of three formats: binding affinity (BA, typically an IC50 value in nanomolar), eluted ligand (EL) likelihood, or percentile rank against a reference set of random peptides. BA and EL are not interchangeable. BA estimates raw binding strength in a biochemical assay context, while EL is trained on mass spectrometry-eluted peptides and better reflects what actually gets presented on a cell surface. The IEDB MHC-II prediction help documentation is explicit that its recommended default methods combine both signals, since BA-only predictions miss processing and presentation effects that EL data capture.

Immunogenicity and TCR-binding predictors sit downstream of binding prediction. A peptide can bind MHC with high affinity and still fail to trigger a T-cell response, so these models attempt to close that gap. DeepImmuno was trained on experimentally validated immunogenic and non-immunogenic peptide pairs and is used to reorder binding-predictor shortlists by likely T-cell recognition rather than binding strength alone. PanPep applies meta-learning to TCR-peptide recognition and shows improved generalization in zero-shot settings, meaning it performs better than earlier architectures when scoring peptides never seen during training, a property that matters for novel antigens and emerging pathogens according to the ImmunoStruct paper.

Structure-aware multimodal models represent the newest and most computationally demanding class. Rather than treating a peptide as a string of amino acids, they incorporate predicted or modeled three-dimensional structure, capturing how a peptide sits in the MHC groove and how that surface presents to a T-cell receptor. This matters most for peptides where sequence features alone leave real ambiguity: unusual anchor residues, post-translationally modified peptides, or candidates from less-studied HLA alleles.

Choosing which class to deploy depends on project phase:

  • Early discovery, large candidate pools: start with fast sequence-based predictors (NetMHCpan/NetMHCIIpan) to cut a proteome-scale list down to a manageable shortlist.
  • Mid-stage prioritization: layer in immunogenicity or TCR-binding predictors (DeepImmuno, PanPep) to reorder that shortlist by likely T-cell recognition rather than binding alone.
  • Final-stage candidate selection or ambiguous cases: apply structure-aware multimodal models (ImmunoStruct) before committing to synthesis and functional assay costs.
  • Diagnostic or clinical-adjacent work: treat all in silico outputs as hypothesis generators, since diagnostic applications demand tighter specificity than most predictors were trained to deliver.

No single model class replaces the others. A practical pipeline runs sequence-based filtering first because it is cheap and fast, then reserves multimodal structural scoring for the smaller set of candidates where the added compute is justified by the decision at stake.

Key computational tools and pipelines researchers should know

A short list of tools now accounts for most published immunopeptidomics and epitope-discovery workflows. Knowing what each one outputs, and where it fits in a pipeline, prevents the common mistake of treating any single tool's score as a final answer.

  1. NetMHCpan and NetMHCIIpan: the standard sequence-based predictors for MHC class I and class II binding, respectively. NetMHCIIpan 4.1 in both its EL and BA modes is currently recommended by the IEDB as a default MHC-II method, and researchers typically use percentile rank cutoffs (commonly below 2% for strong binders, below 10% for weak binders) rather than raw IC50 values, since rank is more robust across alleles with different overall binding affinity distributions.
  2. DeepImmuno: a deep-learning model scoring immunogenicity of MHC-peptide pairs rather than raw binding. It is best used as a reranking step after a binding predictor has already narrowed the field, since it was trained specifically to separate immunogenic from non-immunogenic peptides among already-plausible binders.
  3. PanPep: a meta-learning model for TCR-peptide recognition that generalizes comparatively well to peptides absent from its training data, useful when working with novel antigens where transfer performance matters more than raw accuracy on well-studied peptides.
  4. nf-core/epitopeprediction: a community-maintained, standardized pipeline that wraps multiple binding predictors into a reproducible workflow, reducing the version-drift and configuration inconsistency that plague ad hoc scripting across a lab.
  5. Immunolyser 2.0: a locally executable platform for immunopeptidomic analysis that includes MHC-TP, a module for predicting MHC class I and class II haplotypes when direct genetic typing is unavailable, according to Immunolyser's own documentation. It also automates motif clustering for multi-allelic samples, easing deconvolution work that would otherwise require custom scripting.
  6. ImmunoStruct: a multimodal deep-learning model integrating sequence, structural, and biochemical information to predict class I peptide-MHC immunogenicity, trained on a dataset of 26,049 peptide-MHC pairs and reporting improved alignment with in vitro assay results compared with sequence-only methods, per the Nature Machine Intelligence paper.

Pro Tip: Run at least two independent binding predictors on the same candidate list before moving to immunogenicity scoring; disagreement between tools is itself a useful filter for deprioritizing unstable calls.

Combining outputs across these tools requires a consistent strategy rather than an ad hoc average. The IEDB's own guidance treats consensus and recommended methods as a starting point, not a final score, and recommends checking whether a peptide clears binding thresholds under multiple algorithms before treating it as a strong candidate. In practice, that means using NetMHCpan/NetMHCIIpan for the first-pass filter, DeepImmuno or PanPep for immunogenicity reranking within that filtered set, and reserving nf-core/epitopeprediction or Immunolyser 2.0 as the pipeline scaffold that keeps the whole process auditable and repeatable across projects and lab members. A pipeline that cannot be rerun by a second person with the same result is not ready for publication-grade work, which is precisely the gap these standardized tools were built to close.

Structural modeling: adding AlphaFold, TFold, and multimodal predictions

Structure changes candidate ranking most often when sequence alone leaves genuine ambiguity: unusual anchor residues, peptides from less-characterized HLA alleles, or cases where a TCR interface depends on subtle conformational features that a sequence model cannot represent. In those cases, modeling how a peptide actually sits in the MHC groove, and how that surface presents outward to a T-cell receptor, can reorder a shortlist that sequence-based scores alone would rank differently.

ImmunoStruct is the clearest example of this approach in current use. It combines sequence, structural, and biochemical features into a single multimodal model for class I peptide-MHC immunogenicity, and its authors report improved predictive alignment with in vitro assay outcomes relative to sequence-only baselines across both infectious disease epitopes and cancer neoepitopes, according to the ImmunoStruct study. Its training set of 26,049 peptide-MHC pairs gives it broader coverage than many earlier structure-informed efforts, though coverage still concentrates on class I; class II structural modeling remains comparatively underdeveloped.

Structure prediction tools originally built for general protein folding, including AlphaFold and newer variants such as TFold, extend this logic further. Rather than relying solely on a trained immunogenicity model, researchers can generate a structural model of the peptide-MHC-TCR complex directly and inspect it for steric clashes, unusual anchor positioning, or interface geometry that a sequence-only tool would miss. This is computationally heavier than running a binding predictor, so it makes sense as a late-stage check rather than a first-pass filter.

Practical steps for incorporating structure without excessive compute cost:

  • Reserve structural modeling for the top tier of candidates that have already cleared sequence-based and immunogenicity filters, not the full candidate pool.
  • Use structural output to flag likely false positives (poor groove fit, clashing side chains) rather than to generate new candidates from scratch.
  • Treat MHC-II structural predictions with more caution than class I, since training data and validated structures for class II remain sparser.
  • Flag noncanonical or post-translationally modified peptides for manual review, since standard structural pipelines are typically trained on canonical sequences and may misrepresent modified residues.

Noncanonical peptides deserve particular attention here. Emerging evidence on the immunopeptidome shows that peptides derived from noncoding RNA and peptides transported via extracellular vesicles both contribute to what actually gets presented on MHC molecules, a source of complexity that standard prediction pipelines were not built to anticipate, according to a 2025 review. For neoantigen discovery work in particular, this means a purely canonical-sequence pipeline may be missing legitimate presented peptides entirely, not just misranking known ones. Structural checks help with ambiguous canonical peptides, but they do not solve the noncanonical discovery problem on their own; that gap is better addressed at the mass spectrometry stage, discussed next.

Experimental validation: confirming predictions in the lab

Computational ranking narrows a candidate list; it does not confirm immunogenicity. That confirmation happens through mass spectrometry immunopeptidomics and functional T-cell assays, and the order in which a lab deploys them has a real effect on cost and turnaround.

Mass spectrometry-based immunopeptidomics identifies peptides actually presented on cell-surface MHC molecules, which is a materially different question from whether a peptide can bind MHC in isolation. According to a 2022 review in Frontiers in Immunology, MS elution data consistently show that many peptides predicted as high-affinity binders are never naturally processed and presented, meaning binding predictions alone systematically overpredict viable candidates. Synthetic overlapping peptide libraries offer an alternative or complementary approach: they do not depend on natural antigen processing, which makes them useful for probing T-cell reactivity against sequences that MS might miss, but they cannot confirm that a peptide is naturally presented in vivo the way MS elution data can.

Naturally presented peptides identified by mass spectrometry are the closest computational surrogate to biological ground truth, and predicted binders that fail to show up in MS elution data warrant lower priority for downstream functional testing, per the same Frontiers review.

Functional assays remain the final word on immunogenicity, since presentation alone does not guarantee a T-cell response. Three assays cover most current use cases:

  • ELISpot: measures cytokine secretion (commonly IFN-gamma) at the single-cell level and remains a standard first functional check for candidate epitopes.
  • Intracellular cytokine staining (ICS): allows multiparameter flow cytometry readouts, useful when a lab needs to characterize which T-cell subsets respond, not just whether a response occurred.
  • Activation-induced marker (AIM) assays: detect antigen-specific T cells by surface marker upregulation rather than cytokine production, useful for capturing responses that ELISpot or ICS panels might not be tuned to detect.

Integrating MS data with expression data and TCR sequencing sharpens prioritization further. A peptide confirmed by MS elution, expressed at meaningful levels in the tissue of interest, and paired with TCR sequencing evidence of clonal expansion in an exposed cohort represents a substantially stronger candidate than one supported by prediction scores alone. Recent work on AI in immunopeptidomics also points to rescoring frameworks, such as MS2 Rescore and AlphaPeptDeep, that combine MS output with computational rescoring to reduce false positives and better capture noncanonical or modified peptides that pure in silico pipelines tend to miss, according to a 2025 review.

A resource-conscious order of operations: run sequence and immunogenicity predictors first, confirm the surviving shortlist against available MS immunopeptidomics data where feasible, and reserve functional assays (ELISpot, then ICS or AIM for finer characterization) for the smaller set of candidates that clear both computational and MS-based filters. For labs building out these protocols in more detail, a dedicated guide to in vitro peptide testing methods covers assay selection considerations beyond what fits here.

A practical end-to-end pipeline for research teams

A reproducible pipeline needs explicit thresholds and documentation checkpoints, not just a sequence of tool names. The following stepwise workflow reflects current practice across the tools already discussed.

  1. Define the experiment scope: specify the target antigen or proteome, relevant HLA alleles or supertypes, and whether the goal is vaccine design, diagnostic development, or basic immune response characterization, since each downstream step differs by objective.
  2. Run the in silico ensemble: apply NetMHCpan/NetMHCIIpan first, using percentile rank cutoffs (commonly below 2% for strong binders, below 10% for weak binders) to generate an initial shortlist, then rerank with DeepImmuno or PanPep for immunogenicity likelihood.
  3. Apply structural checks selectively: for the top tier of candidates, or for ambiguous cases involving unusual alleles or modified residues, run ImmunoStruct or structure-prediction tools like AlphaFold or TFold before finalizing the list.
  4. Confirm presentation with MS immunopeptidomics where feasible, treating peptides absent from elution data as lower priority regardless of their predicted binding score.
  5. Validate functionally: run ELISpot as a first pass, then ICS or AIM assays on confirmed responders to characterize the responding T-cell population.
  6. Iterate: feed functional assay results back into threshold selection for the next round, tightening cutoffs where false positive rates were high and relaxing them where confirmed epitopes fell outside the initial computational shortlist.

Pro Tip: Log every threshold change with a reason and a date; a pipeline whose cutoffs drift undocumented across projects becomes impossible to compare or reproduce a year later.

Numeric thresholds should flex with project goals. A vaccine-design project casting a wide discovery net can tolerate a looser percentile cutoff to avoid missing true positives, while a diagnostic application, which the immunoinformatics literature notes requires stricter specificity than most predictors were originally trained to deliver, should tighten thresholds and weight functional confirmation more heavily before any candidate moves forward.

Documentation is not optional overhead in this workflow, it is what makes results usable by anyone other than the person who generated them. A minimum checklist includes:

  • Certificates of Analysis for every peptide batch used in functional assays, confirming identity and purity before results are interpreted.
  • Exact software versions and parameter settings for every predictor and pipeline run (NetMHCpan version, nf-core/epitopeprediction release, threshold values).
  • Raw MS and assay data deposited in an accessible repository, not summarized figures alone.
  • Metadata on HLA typing method, donor or sample source, and assay conditions (temperature, incubation time, reagent lots).

Peptide quality itself affects reproducibility more than researchers sometimes assume. Post-translational modifications, degradation during storage, and reconstitution errors can all shift assay results independent of any biological signal, which is why batch documentation and independently verified purity matter as much as the computational side of the pipeline. A peptide sequencing methods overview covers additional detail on confirming peptide identity before it enters an assay.

Where predictions go wrong: limitations and best practices

High predicted binding affinity is necessary but not sufficient for immunogenicity, and treating a strong binding score as a confirmed epitope is the single most common overinterpretation in this field. MS elution studies repeatedly show that many predicted high-affinity binders are never naturally processed and presented, and even among peptides that are presented, only a subset triggers a measurable T-cell response, per the Frontiers review on immunopeptidomics validation.

Dataset bias compounds this problem. Training data for most binding and immunogenicity predictors skew toward well-studied HLA alleles, meaning performance on underrepresented alleles is less reliable even when a tool reports strong aggregate accuracy. High-throughput datasets used for training also carry their own noise, and rescoring frameworks like MS2 Rescore exist specifically to filter that noise out before it propagates into downstream predictions.

MHC class II prediction remains a harder problem than class I across the board. The binding groove is open-ended rather than closed, which makes exact anchor positions less predictable, and a 2025 review of AI in immunopeptidomics notes that combining cleavage signatures, expression data, and allele-specific context improves class II prediction more than relying on sequence-based binding scores alone. Post-translationally modified and noncanonical peptides, including those derived from noncoding RNA, add further complexity that standard canonical-sequence pipelines were not built to capture.

A few practices reduce the risk of these pitfalls compounding across a project:

  • Treat any single tool's output as a hypothesis, not a confirmed result, regardless of how confident its reported score appears.
  • Cross-check binding predictions against MS elution data before committing functional assay resources to a candidate.
  • Flag results generated for underrepresented HLA alleles as lower confidence pending additional validation.
  • Handle donor HLA typing and functional assay data with the same privacy and biosafety controls applied to other identifiable human-subject data.
  • Use standardized, versioned pipelines (nf-core/epitopeprediction, Immunolyser 2.0) and deposit raw data to support reproducibility across labs and over time.

How documented peptide quality supports reproducible research

Reproducibility in immunology peptide modeling depends on more than the computational pipeline. The peptides used to validate a prediction, whether in ELISpot, ICS, or AIM assays, need to match their stated identity and purity, or the resulting data cannot be trusted regardless of how sound the upstream modeling was.

Its stated approach layers in-house checks with third-party verification and full batch documentation, giving researchers a Certificate of Analysis on request rather than relying on supplier claims alone. For a validation pipeline that already depends on tight thresholds and careful documentation at every computational step, as outlined above, the same rigor applied to the physical peptide inputs closes a gap that predictors and pipelines cannot address on their own.

Variance introduced by uncertain peptide sourcing, degraded stock, inconsistent reconstitution, or unverified identity, can look identical in an assay readout to a genuine biological signal, making a confirmed epitope indistinguishable from an artifact. Documented purity and traceable batch records do not replace functional validation, but they remove one avoidable source of noise from an already resource-intensive process. For researchers building out the kind of pipeline described in this guide, treating peptide sourcing with the same documentation discipline applied to software versions and assay metadata is a small step that protects the value of everything upstream.

Sourcing research-grade peptides for immunology work

Researchers running the pipeline described above need peptides that arrive with the same documentation rigor as their computational pipeline. Peptastic Labs' catalogue covers eight product lines, including Metabolic, Cognitive & Neuro, Tissue & Repair, Longevity, Cosmetic Science, Blends, Hormone & Reproductive, and Ancillaries & Reagents, all independently HPLC-verified to at least 99% purity.

Peptasticlabs

Batches typically ship with documentation, and a Certificate of Analysis may be available on request to provide a verifiable record for reproducibility. Bulk and wholesale ordering options can be offered for teams conducting larger validation studies. The research page collects application notes relevant to immunology use cases for researchers scoping a new project.

Browse the catalogue to check current stock and documentation for a specific peptide, or reach out through the research page with questions about batch specifications before placing an order.

Where peptide immunogenicity modeling is headed

Multimodal models will keep expanding, but the more consequential shift is upstream: better MS-trained datasets, not just better architectures, are what will close the gap between predicted and actual immunogenicity. ImmunoStruct's reliance on a 26,049-pair training set is a step in that direction, but it also shows how far current data still lags what full generalization would require.

Diagnostics and vaccine design need different validation standards, and pipelines built for one should not be assumed to transfer cleanly to the other without rechecking specificity requirements. What the field lacks most is not another model architecture but shared benchmarks and open data-sharing practices that let labs measure generalization to genuinely unseen peptides rather than to held-out slices of the same training distribution. Standardized pipelines like nf-core/epitopeprediction point toward that future, but the underlying datasets need the same community investment the tools have already received.

— Tintastic

This article is general information, not a substitute for advice from a qualified doctor. Consult a qualified healthcare professional about your own circumstances before acting on anything here.

Sources

FAQ

What are the four main types of peptides?

Peptides are commonly grouped by function into signaling peptides (hormones and neurotransmitters), structural peptides (collagen fragments and related repair peptides), transport peptides, and defense or immune-related peptides such as antimicrobial and immunomodulatory peptides. In immunology research specifically, the more relevant grouping is by role in the discovery pipeline: binding peptides, immunogenic epitopes, structural (multimodal) candidates, and validated functional epitopes.

Which peptide is best for the immune system?

There is no single best peptide for immune support, since immunology research relies on identifying specific epitopes matched to a target antigen and the reader's own HLA context rather than a universal candidate. Research-grade peptides used in immunology studies span product categories including immunology-specific peptides, blends, and ancillaries, selected based on the experimental design and target pathway under study.

What are immune system models?

Immune system models are computational or experimental frameworks used to predict or study immune responses, ranging from sequence-based binding predictors like NetMHCpan to structure-aware multimodal models like ImmunoStruct and TCR-binding predictors like PanPep. They are typically paired with experimental validation methods, including mass spectrometry immunopeptidomics and functional T-cell assays such as ELISpot, to confirm that a computational prediction reflects an actual biological response.

Is Ozempic considered a peptide?

Ozempic's active ingredient, semaglutide, is a peptide-based GLP-1 receptor agonist, though it falls under metabolic and diabetes research rather than immunology peptide modeling. It is structurally and functionally distinct from the epitope and immunogenicity-focused peptides discussed in this guide, which target antigen presentation and T-cell recognition rather than metabolic receptor signaling.