Review Papers Claim by Claim

Visual-first: core quantitative performance (data from paper)

Visual-first: method schematic (compact)

Critical appraisal — evidence-based, citation-backed

Strengths

• Fast, scalable ranking across full sequences/proteomes because scoring reduces to matrix lookup + sum; valuable for alleles lacking experimental binding sets (PRE-DEP speed/utility).
• Transparent, interpretable scoring: position-wise contributions and choice of pairwise potentials are explicit; this aids mechanistic reading of why an amino acid is favored at a position Interpretability of score matrix.

Limitations and blindspots (explicit)

• Template dependence — requires at least one solved peptide–MHC complex (or reliable structural model) for an allele; coverage limited in 2004 and still a factor (modern curated template sets mitigate but do not remove this) Template dependence (Altuvia & Margalit).
• Energy model coarseness — statistical pairwise potentials provide relative interaction preferences, not physical free energies; observed correlations (r ≈ −0.62 or −0.5) show moderate explanatory power but leave large residuals (i.e., many false positives/negatives) Reported correlations.
• Position independence assumption — summing independent per-position scores ignores coupling (epistatic) effects between peptide positions and backbone-dependent context; later structure-aware methods (backbone templating, Rosetta relax, ML on 3D data) address couplings and often yield better accuracy HLA3DB / RepPred (contrast).
• Pocket chemistry bias — initial Miyazawa–Jernigan potentials emphasize hydrophobic interactions, explaining better performance for hydrophobic-pocket alleles; authors improved this with Betancourt–Thirumalai potentials but residual allele‑specific biases remained Potentials and pocket bias.
• Validation dataset & selection bias — the validation used an internal Sette dataset and literature-derived immunogenic peptides; authors note historical bias where experimental sets over-sampled anchor-preferred peptides, which can inflate sequence-method performance and complicate fair benchmarking Validation data bias.

Where this paper sits in the modern landscape

The 2004 threading approach is a clear forerunner of modern structure-aware epitope prediction workflows. Recent resources and methods—HLA3DB/RepPred for backbone templating, geometric/graph deep‑learning, and AF/AF-derived pipelines—explicitly model backbone/side‑chain coupling and learn energy-like scoring from large structural datasets, increasing accuracy and generalizability beyond simple position-wise potentials HLA3DB (2023).

Practical recommendations for users who might consider PRE-DEP today

1. Use PRE-DEP when: an experimental binding dataset for the allele is absent but a high-quality solved peptide–MHC structure (or close homology model) exists; it will efficiently produce a ranked candidate list for experimental testing PRE-DEP use-case.
2. Prefer modern structure‑aware methods when: backbone templates or large structural databases are available (RepPred, Rosetta-based exhaustive modeling, AF3/AF2/structure+ML pipelines) because they model backbone/side-chain coupling and generally yield higher accuracy RepPred advantage.
3. Validate predictions experimentally and use orthogonal filters (proteasomal cleavage/TAP/abundance/immunopeptidomics) before investing in immunogenicity assays — binding ranking alone is necessary but not sufficient for immunogenicity Binding ≠ immunogenicity (Altuvia & Margalit).

What would falsify the main claims

• Large-scale blinded tests across diverse alleles with solved structures showing poor correlation / poor ranking enrichment (e.g., immunogenic peptides not enriched among top ranks) would falsify practical usefulness; authors note this as the natural falsification test and provide a plan to improve scoring Falsification logic (Altuvia & Margalit).

Concise actionable improvements to the method (technical)

1. Replace position-independence by template-aware local coupling terms (e.g., pairwise/triwise position couplings) or use Rosetta/PyRosetta relax steps on top candidates to capture induced-fit.
2. Use backbone-class templates (HLA3DB-style) to include distinct backbone-dependent sequence biases rather than a single peptide trace.
3. Replace static statistical potentials by hybrid score: structure-derived energies + ML re-calibration on available binding datasets (transfer-learning) to reduce allele bias.

Concluding evaluation (compact)

Altuvia & Margalit 2004 introduced an interpretable, fast, structure-based threading approach (PRE-DEP) that meaningfully narrowed candidate peptide lists without prior binding data and was empirically useful for hydrophobic-pocket alleles; however, position independence, coarse statistical potentials, and template dependence limit accuracy compared with later backbone-aware and ML‑augmented structure methods. The paper is historically important and practically useful when modern templates or binding datasets are absent, but users should validate predictions experimentally and consider modern alternatives when available Overall judgement (Altuvia & Margalit).

Author review quick-links

All method descriptions and performance numbers above are taken directly from: Altuvia Y., Margalit H., "A structure-based approach for prediction of MHC-binding peptides" Methods 34 (2004) 454–459 (DOI:10.1016/j.ymeth.2004.06.008) Altuvia & Margalit 2004 (primary)