Ten Common Failures in Neoantigen Prediction - And What Expert Analysts Do to Avoid Them

Introduction

Neoantigen prediction is becoming a critical part in immuno-oncology research and cancer vaccine development. It’s a beautiful idea: take the tumor’s own mutations, predict which of them can generate immunogenic peptides, and use those to target the cancer with unprecedented specificity.

But as simple as it sounds, real-world neoantigen prediction is full of failure points. We’ve worked with research groups and biotech teams who used top prediction tools - NetMHCpan, MHCflurry, pVACseq - and still failed to get results that matched immune response, patient outcome, or even basic expectations. Some were using perfect exome and RNA-seq data. The problem was in the pipeline logic, parameter choice, or lack of downstream filtering.

In this article, we summarize ten common pitfalls we often find in neoantigen prediction workflows. For each one, we explain the failure mode, the reason behind it, and how a more experienced approach can catch or avoid it - before it ruins your immunogenicity ranking or vaccine design.

Table of Contents

1. Misaligned Somatic Variant Calls

The Problem

Predicted neoantigens are derived from incorrect or misaligned somatic mutations, leading to non-existent peptides.

Why It Happens

If tumor-normal calling is not carefully tuned (e.g., Mutect2, Strelka2), or reference genomes are mismatched between variant calling and neoantigen tools, the wrong peptides are generated. Even indels are often mishandled, and annotation errors propagate through the pipeline.

Real Example

One group used GATK Mutect2 without panel of normals. They submitted 3,500 SNVs to a neoantigen pipeline. Over 20% mapped to pseudogenes or low-confidence regions, and several had incorrect coding effect annotations.

What We Do Differently

We use a validated tumor-normal variant calling setup with post-filtering and strict matching of genome versions (GRCh38 or hg19). We reannotate all variants using VEP or ANNOVAR before any peptide generation, and perform rigorous quality control on all inputs.

2. Incorrect HLA Typing or HLA Format

The Problem

Neoantigen predictions rely on patient-specific HLA types - but often the HLA calls are wrong, incomplete, or in the wrong format. As a result, predicted binding affinities have no biological meaning.

Why It Happens

Many teams use low-depth WES or RNA-seq to infer HLA alleles, but these methods lack resolution. Tools like OptiType or HLAPRG need specific read coverage and proper sample preparation. Worse, some analysts use ambiguous or misformatted allele names (e.g., 02:01 instead of HLA-A02:01), causing the binding predictor to fail silently.

Real Example

In one personalized vaccine trial, the team used RNA-seq-derived HLA calls but included only two alleles per patient. NetMHCpan expected six. Half of the predictions were discarded without warning, and true HLA-A restricted binders were missed.

What We Do Differently

We always validate HLA typing with multiple tools - e.g., OptiType + arcasHLA - and cross-check with clinical HLA reports when available. We reformat all alleles to the expected NetMHCpan nomenclature and log-check every call. If HLA types are uncertain, we flag predictions as low-confidence and exclude them from vaccine prioritization.

3. Misuse of NetMHCpan Score Thresholds

The Problem

Peptides are classified as binders or non-binders using fixed affinity cutoffs - often too strict or too permissive - which misclassify real neoantigens or flood the list with junk.

Why It Happens

NetMHCpan returns multiple values: affinity, percentile rank, and BA/R score. Many analysts use a simple IC50 cutoff (e.g., <500 nM) without understanding its context. Others apply fixed rank thresholds that were tuned for viral peptides, not self-derived tumor antigens.

Real Example

A biotech team filtered for peptides with <50 nM affinity, assuming stronger is always better. But most high-affinity binders came from wild-type sequences with low tumor specificity. Meanwhile, real immunogenic mutations with 200–300 nM affinity were discarded.

What We Do Differently

We optimize thresholds based on the allele, tumor type, and reference distribution. For each project, we analyze binding score distributions for both mutant and wild-type peptides. We prioritize delta-binding difference and tumor specificity - not just absolute affinity.

4. Ignoring Peptide Processing and Presentation

The Problem

Predicted peptides look like strong binders in silico - but are never actually produced or presented by the cell.

Why It Happens

Class I MHC presentation depends not only on binding but also on proteasomal cleavage, TAP transport, and peptide trimming. Tools like NetCTLpan or MixMHCpred model these steps, but many workflows skip them. So, peptides predicted as “binders” are biologically irrelevant.

Real Example

One group submitted a list of top mutant peptides for validation via tetramer staining. None triggered T-cell response. On inspection, most peptides had poor processing scores and were unlikely to be generated inside the cell.

What We Do Differently

We integrate proteasomal cleavage predictors, TAP scores, and presentation likelihood models. For class II, we consider flanking sequences and use tools like MARIA or MixMHC2pred. We exclude peptides that fail processing or are inconsistent with known presentation motifs.

5. Expression Not Filtered by RNA Evidence

The Problem

Neoantigens are prioritized without checking whether the mutated gene is even expressed in the tumor - or whether the specific variant transcript is present.

Why It Happens

Analysts often rely on WES-only pipelines. Even with RNA-seq, they check only gene-level TPM without verifying allele-specific expression. Mis-spliced transcripts, intronic variants, or low-expression genes slip through.

Real Example

A colorectal cancer team nominated 18 peptides from non-expressed genes. Three of them came from retained introns not present in tumor RNA. The clinical trial arm based on these peptides failed to show immune activation.

What We Do Differently

We use matched RNA-seq to confirm both gene and allele-level expression. We quantify mutant vs wild-type transcript ratios and exclude candidates lacking evidence of transcription. We also remove peptides from non-coding or nonsense-mediated decay targets.

6. Clonality and VAF Not Considered

The Problem

Peptides are chosen from mutations that exist only in a subset of tumor cells. Even if they’re good binders, they won’t generate uniform immune response.

Why It Happens

Most prediction workflows don’t distinguish between clonal and subclonal mutations. VAF (variant allele frequency) is not used, or is misinterpreted due to copy number variation and tumor purity issues.

Real Example

A patient received a peptide vaccine targeting a mutation with 8% VAF. Later analysis showed that only a minority of tumor cells carried the mutation. The immune system responded, but the bulk of the tumor escaped.

What We Do Differently

We estimate clonality using VAF, tumor purity, and copy number adjustments (e.g., from FACETS or PureCN). Only mutations present in most tumor cells are considered for high-priority vaccines. For mixed clones, we rank neoantigens by both immunogenicity and clonal coverage.

7. Immunogenicity Assumptions That Mislead

The Problem

Some peptides are assumed to be immunogenic based purely on predicted binding - but this misses the complexity of T-cell biology.

Why It Happens

MHC binding is necessary, but not sufficient. Tools like DeepHLApan or PRIME attempt to model immunogenicity, but are still imperfect. Some workflows treat all binders as equally likely to trigger response, which is not true.

Real Example

One dataset had multiple peptides with <50 nM predicted affinity. None showed IFN-γ release. But a lower-affinity peptide with a known pathogen-like motif triggered strong CD8+ response. Structural mimicry - not just binding strength - mattered.

What We Do Differently

We combine binding predictions with known immunogenic motifs, TCR availability, and epitope databases like IEDB. We annotate for pathogen mimicry and use in-house scoring models trained on validated neoantigen datasets.

8. Unfiltered Self-Similar or Tolerized Peptides

The Problem

Many predicted neoantigens differ only subtly from wild-type - and are ignored by the immune system due to central tolerance.

Why It Happens

Some tools don’t compare mutant peptides against self-peptides. A single amino acid change isn’t enough if the rest of the sequence matches self - especially in anchor positions. Tolerized or cross-reactive peptides create false positives.

Real Example

A breast cancer study used high-throughput peptide screening. Mutant and wild-type peptides had identical binding scores, and both failed to stimulate T cells. Sequence similarity analysis showed 95% overlap.

What We Do Differently

We compute mutant–wild-type similarity scores and penalize minimal-difference peptides. We filter peptides with high self-homology or likely T-cell tolerance. If the change occurs outside TCR contact residues, we deprioritize them.

9. Vaccine Target Selection Misprioritized

The Problem

Even with a good list of predicted neoantigens, the wrong ones are chosen for the actual vaccine - due to oversimplified ranking.

Why It Happens

Many pipelines rank peptides by binding affinity alone. Others use simple scoring functions without integrating all features - expression, clonality, processing, immunogenicity. Important tradeoffs (e.g., moderate binder but highly clonal) are missed.

Real Example

A personalized vaccine trial selected top 5 peptides by affinity, all from lowly expressed genes. Stronger candidates were skipped due to poor sorting logic.

What We Do Differently

We use multi-parameter ranking algorithms that weigh expression, processing, binding, immunogenicity, and clonal coverage. We generate ranked shortlists tailored to vaccine platform constraints (e.g., peptide length, manufacturing limits).

10. Overpromising Without Mass-Spec or T-Cell Validation

The Problem

Researchers publish neoantigen candidates with strong prediction scores - but no orthogonal validation. Reviewers and collaborators lose trust, and the clinical impact is questionable.

Why It Happens

In silico predictions are fast and easy, but experimental confirmation is not. Mass-spec or T-cell assays require effort, and are skipped. So results stay theoretical - or worse, misleading.

Real Example

One high-profile paper listed dozens of neoantigens without validation. A follow-up group tried to detect them via HLA ligandome mass-spec - and failed to find any. Confidence in the pipeline dropped.

What We Do Differently

We work with collaborators to validate top-ranked peptides using HLA peptidomics or T-cell assays. For clients, we flag peptides as “validated”, “inferred”, or “untested”. We also design spike-in controls and feasibility plans to test real presentation.

Final Remarks

Neoantigen prediction sounds like a simple pipeline - mutation in, peptide out. But in real-world projects, many things go wrong. Binding affinity is just one part of the story. If expression, processing, clonality, or tolerance are ignored, predictions fail. If HLA calls are off, or validation is skipped, results collapse.

We’ve worked with teams across academia and biotech to recover failed pipelines, build robust prediction workflows, and guide real vaccine development. It’s not just about getting a list - it’s about knowing which peptides matter, and why. When neoantigen analysis is done right, it doesn’t just predict-it explains and drives real immunotherapy success.

This blog article was authored by Justin T. Li, Ph.D., Lead Bioinformatician. To learn more about AccuraScience's Lead Bioinformaticians, visit https://www.accurascience.com/our_team.html.