Core Differentiation

pMHC 3D Structure Prediction

已验证

DiVo in-house pipeline module · External technical capability report (sanitized)

Step 5 self-developed unique module of DiVo Gen²AI's neoantigen 8-step pipeline. Traditional pipelines stop at affinity values for binding judgment; DiVo builds atomic-level pMHC 3D structures for each candidate neoantigen, with in-house cross-validation--highest confidence 95.4, complex overall score 0.977, approaching X-ray crystallography precision. 400+ complexes validated at scale. This page serves three audiences: patients and public wanting to understand pMHC, CAR-T/neoantigen teams seeking structure prediction partners, and investors and peers evaluating technical capabilities.

Patients & Public

Read "What is pMHC Structure" below--understand why affinity values aren't enough, 3D structures are also needed.

Partners / Hospitals

Focus on "DiVo's Role", "3-Step Pipeline", "Evidence"--we have 400+ complex real-world data.

Investors / Peers

Focus on "Differentiation", "Benchmarks", "Glossary"--evaluate the technical barrier of pMHC structure prediction.

pMHC structure prediction is a core differentiation step for the following flagship services

What is pMHC Structure Prediction

In the tumor neoantigen pipeline, MHC binding prediction gives affinity values--quantitative metrics of peptide-MHC molecule binding. Stronger affinity theoretically means more likely to be recognized by T cells. But affinity values only tell you "they bind", not "how they bind".

The spatial conformation of a peptide embedded in the MHC groove determines whether it can be stably presented on the cell surface and correctly recognized by TCR. Peptides that meet affinity thresholds but have unstable spatial conformations--"crooked keys" in the groove--are false positives that the immune system would never see in vivo.

pMHC 3D structure prediction builds full-atom 3D models of "peptide-MHC complexes" for each candidate peptide, validates conformational credibility with confidence metrics, then cross-validates to eliminate false positives. This is not just visualization--it's a qualitative change from "may bind" to "confirmed presentation".

Why Structure

Affinity values only answer "they bind", not "how they bind". Peptides with unstable spatial conformations are false positives--affinity passes but conformation is skewed, won't be stably presented.

Why Cross-Validation

Single prediction may overfit. Structure prediction + conformational stability cross-validation, two checkpoints filtering layer by layer, ensuring remaining candidates are truly credible at the atomic level.

DiVo Gen²AI's Role

pMHC structure prediction is Step 5 of DiVo's neoantigen 8-step pipeline, and CT5 core capability of the CAR-T 5-step pipeline. We complete the full loop from structure prediction to cross-validation--validating conformational credibility, eliminating false positives, with structure scores directly participating in immunogenicity ranking at 25% weight. This is not third-party tool stitching, but a complete capability loop orchestrated by in-house pipeline.

Underlying implementation details--including model architecture, toolchain, in-house weights and scheduling logic--are not disclosed in this report, available under cooperation frameworks as needed.

Core Capability · 3-Step Pipeline

From candidate peptides to conformational validation structure prediction loop · Click steps to view underlying validation

PS1

Candidate Peptide Input

已验证->

Receives MHC binding prediction output

Affinity-qualified candidate peptide list

PS2

pMHC 3D Structure Prediction

已验证

DiVo in-house structure prediction engine (Protenix)

Full-atom pMHC 3D structure + pLDDT/ipTM

PS3

Conformational Stability Cross-Validation

已验证

DiVo in-house validation engine

Eliminates conformationally unstable false positives

Differentiation

Core differences from traditional "affinity-value-only" pipelines

3D

Atomic-Level 3D Structure Replaces Pure Value Judgment

Traditional pipelines stop at affinity values for binding judgment, not knowing the spatial conformation of peptides in the MHC groove. DiVo builds atomic-level pMHC 3D structures for each candidate, validating conformational credibility with confidence metrics--a qualitative change from "may bind" to "confirmed presentation".

WT

Structure Score Participates in Immunogenicity Ranking

Not just visualization; pMHC structure scores directly participate in immunogenicity ranking at 25% weight. False positives with passing affinity but unstable conformations are eliminated at the structural level, reducing downstream无效 experiments.

DiVo In-House Capability Modules

Functional capabilities · All verified

Capability ModuleUnderlying EngineFunction DescriptionStatus
pMHC 3D Structure PredictionDiVo in-house structure prediction engineOutputs full-atom 3D structure of peptide-MHC complex, with confidence metrics, validating spatial conformational credibility 已验证
Conformational Stability Cross-ValidationDiVo in-house validation engineCross-validates structure prediction results, eliminating false positives with passing affinity but unstable spatial conformation 已验证
Multi-Chain Complex Large-Scale PredictionDiVo in-house batch prediction schedulerSupports peptide-receptor, tetramer+substrate and other complex systems batch prediction, validated at scale, not just point demos 已验证

Three capability modules together form DiVo's complete pMHC structure prediction capability loop: structure prediction gives conformation -> cross-validation eliminates false positives -> large-scale scheduling supports real-world batch delivery. Modules are orchestrated by DiVo's in-house pipeline, not third-party tool stitching.

Evidence · pMHC Structures

Peptide sequences sanitized · Confidence metrics are measured values

ComplexAntigen SourceConfidenceComplex Overall Score
HLA-A*02:01 + candidate peptide ①HTLV-1 Tax95.40.977
HLA-A*02:01 + candidate peptide ②WT194+0.96+
HLA-A*02:01 + candidate peptide ③WT194+0.96+
HLA-A*02:06 + candidate peptide ④MAGE-A194+0.96+
HLA-A*02:06 + candidate peptide ⑤HTLV-1 Tax94+0.96+

The candidate peptide sequences of the above 5 pMHC structures have been sanitized (numbered ①-⑤), antigen sources retained for credibility cross-reference. Confidence metrics are all measured values, not simulated.

Evidence · Large-Scale Validation

Project-level scale generalized and sanitized

MetricValueNote
Large-scale real-world projects千级突变 × 多版迭代Multi-project cross-validation
Complex batch prediction400+ 复合物Highest confidence 95.4
Structure confidence95.4Approaching X-ray crystallography precision
Complex overall score0.977Atomic-level reliability
Cross-validation通过Conformational stability validation completed

Honest Boundaries

What we can and cannot do, clearly stated

What We Can Do

pMHC full-atom 3D structure prediction
Confidence metric validation (pLDDT + ipTM)
Conformational stability cross-validation
Structure score participates in immunogenicity ranking (25% weight)
400+ complex large-scale batch prediction

What We Don't Do

No experimental validation (X-ray crystallography/cryo-EM)
No TCR-pMHC ternary complex structure prediction
No BCR/antibody-antigen complex structures
No direct-to-patient structure interpretation
Underlying implementation details not disclosed in this report

Glossary

5 core terms in pMHC structure prediction

Abbr.Full NameTranslationExplanation
pMHCpeptide-MHC complexpeptide-MHC ComplexAntigen peptide complex presented by MHC molecules, the target of T cell recognition
pLDDTpredicted Local Distance Difference Testpredicted Local Distance Difference TestStructure prediction confidence metric, >90 is high confidence, 95.4 approaches X-ray crystallography precision
ipTMinterface predicted TM-scoreinterface predicted TM-scoreProtein complex interface interaction confidence, >0.75 is high confidence
MHCMajor Histocompatibility ComplexMajor Histocompatibility ComplexThe "display board" on cell surfaces, presenting intracellular protein fragments to T cells
假阳性False PositiveFalse PositiveCandidate peptides with passing affinity values but unstable spatial conformations, not actually stably presented

CAPACITY TRACE

能力回溯

这项服务由哪些能力支撑——从硅片到你的场景