Pipeline B

Anti-aging Peptides & Protein Engineering

已验证

Finding anti-aging molecules in 3D space · 100% tools deployed & verified

Target 3D structure triple-engine prediction -> molecular docking + affinity calculation -> peptide inverse folding design -> stability + immunogenicity dual screening. 4-step pipeline with all tools deployed and verified, from Protenix V2 to MHCflurry, 38 SOTA models running locally.

Functional Skincare Brands

Read "Core Capability" and "Benchmarks" below--evaluate our anti-aging peptide dry-computing screening capability and precision.

Peptide / CRO

Focus on "Differentiation", "Pipeline Steps", "Glossary"--evaluate molecular-level computing screening differentiation.

Investors / Peers

Focus on "Differentiation", "Benchmarks", "Capacity Trace"--evaluate protein engineering model barriers.

What is Protein/Peptide Screening

Anti-aging peptide R&D is essentially finding molecules in 3D space that precisely bind target proteins. Traditional methods rely on high-throughput wet-lab screening--long cycles, high costs, low hit rates. Dry-computing screening uses AI to predict target protein 3D structure, then simulates molecular docking in silico, and finally uses inverse folding design to generate candidate peptide sequences--compressing wet-lab workload by orders of magnitude.

DiVo uses three independent structure prediction engines for cross-validation--Protenix V2, ESMFold, Boltz-2--only entering downstream pipeline when pLDDT>90. From docking, design to immunogenicity screening, 4-step pipeline with 100% tools deployed, 38 SOTA models running locally.

Why Triple-Engine Cross-Validation

A single structure prediction model may perform differently across protein types. Three independent engines jointly predict, only entering downstream when pLDDT>90--covering different modeling principles and structural spaces, ensuring reliable starting points for downstream docking and design.

Why Inverse Folding Design

Traditional peptide design relies on empirical libraries and random mutations, with limited efficiency and search space. ProteinMPNN/LigandMPNN reverse-engineer optimal sequences from 3D binding pockets--given a structure, find the sequence, directly generating high-affinity candidates.

DiVo Gen²AI's Role

Protein/peptide screening is the core output of DiVo's four-in-one service Pipeline B, and a key step in functional skincare brand anti-aging R&D for dry-computing screening from target to candidate molecules. We provide end-to-end computing from 3D structure prediction to safe peptide candidates--triple-engine cross-validation, molecular docking, inverse folding design, stability + immunogenicity dual screening, all 4 steps with tools deployed.

We do not do wet-lab validation (brand partners confirm after dry-computing screening), do not do GMP production docking (provide sequences + QC standards), do not guarantee in-vivo activity (in-silico docking ≠ in-vivo efficacy). We deliver candidate peptide sequences and QC standards ready for brand partner wet-lab validation.

Core Capability · 4-Step Pipeline

Pipeline B · Protein/peptide screening · All tools deployed & verified

Step B1Target structure predictionProtenix V2 + Chai-1 cross-validation3D structure + active sites 已验证
Step B2Binder/peptide generationPXDesign + BoltzGen + ProteinMPNN + LigandMPNNCandidate binders + peptide sequences 已验证
Step B3Stability + binding assessmentPro-Prime + ProSST + Protenix V2 co-folding + IPSAEΔΔG + ipSAE + pDockQ stability + binding 已验证
Step B4Immunogenicity + safety assessmentMHCflurry + NetMHCpan/IIpan + BepiPred-2.0 + pVACToolsImmunogenicity pre-screening + safe peptide candidates 已验证
Step B5mRNA delivery design (optional)GEMORNA + RNALens + UTRGANmRNA sequence + stability + UTR optimization 已验证
Step B6Multi-dimensional rankingDiVoDesignEngine -> four-dimensional scoringTier 1/2 candidate report 已验证

Differentiation

Core differences from single-engine/undeployed protein engineering pipelines

SOTA

Pro-Prime Replaces FoldX--Stability SOTA

No longer using 2005 FoldX (65-70% accuracy), but deploying Pro-Prime+ProSST for ΔΔG prediction. DDGun (physics-based) as baseline control--if Pro-Prime and DDGun disagree, it signals model hallucination.

5W

BoltzGen 5 Weights Complete--All-Atom Design

BoltzGen 5 weights (7.9G) fully downloaded, 529M parameters. boltz2_aff.ckpt is the architectural basis for DiVoPPIHead affinity prediction. All-atom diffusion design capability ready.

87

87 Tools Full-Stack Zero Additional Investment

From structure prediction to peptide design to immunogenicity assessment to mRNA delivery design, 87 tools all locally deployed. No additional procurement or deployment waiting needed.

Benchmarks

Model benchmarks and deployment status

MetricValueNote
Protenix V2 pLDDT95.4Structure prediction accuracy
BoltzGen5 weights 7.9G completeAll-atom diffusion design, 529M params (boltz2_aff)
Pro-Prime ΔΔGSOTAReplaces FoldX (2005) as primary stability prediction tool
ProteinMPNN>52%Sequence recovery rate (literature benchmark)
Immunogenicity pre-screeningAUROC=0.7547MHCflurry+NetMHCpan (TESLA benchmark)
Deployed tools87Full tool localization across three satellite projects

Honest Boundaries

What we can and cannot do, clearly stated

What We Can Do

Protein 3D structure prediction (Protenix V2 pLDDT>90)
Binder/peptide generation (PXDesign+BoltzGen+ProteinMPNN)
Stability assessment (Pro-Prime+ProSST SOTA-level, DDGun baseline control)
Binding assessment (Protenix V2 co-folding+IPSAE)
Immunogenicity pre-screening (MHCflurry+NetMHCpan/IIpan+BepiPred-2.0)
mRNA delivery design (GEMORNA+RNALens+UTRGAN)
100% tools deployed & verified (87 tools)

What We Don't Do

No wet-lab validation (brand partners confirm after dry-computing screening)
No GMP production docking (provide sequences + QC standards)
No in-vivo activity guarantee (in-silico docking ≠ in-vivo efficacy)
No animal experiments
DiVoPPIHead affinity prediction architecture complete, pending PDBbind training (currently using pDockQ+ipTM)

Glossary

7 core terms in protein/peptide screening

Abbr.Full NameTranslationExplanation
pLDDTpredicted Local Distance Difference TestLocal Distance Difference TestAlphaFold-family model confidence metric, 0-100, >90 indicates highly reliable structure
ΔΔGDelta Delta GBinding free energy changeMetric measuring protein-ligand binding strength, more negative ΔΔG = stronger binding. Pro-Prime is current SOTA predictor
Pro-PrimePro-PrimeProtein stability prediction SOTANext-generation stability prediction model replacing FoldX (2005, 65-70% accuracy), significantly improved precision
BoltzGenBoltzGenAll-atom diffusion protein designMIT open-source protein design model, 5 weights (7.9G) complete. boltz2_aff contains 529M-param AffinityModule
ProteinMPNNProtein Message Passing Neural NetworkProtein Message Passing NNProtein inverse folding design model, generates optimal amino acid sequences from 3D backbones, >52% recovery
GEMORNAGEMORNAmRNA sequence design modelDiVo's in-house fine-tuned mRNA design model, paired with RNALens stability prediction + UTRGAN UTR optimization
MMPMatrix MetalloproteinaseMatrix MetalloproteinaseZinc-dependent endopeptidase family, commonly used as target in anti-aging peptide screening, regulates ECM degradation