Self-Tuned

mRNA Sequence Optimization

已验证

RNALens · Spearman 0.92 · CAI 0.7->0.95 · Self-tuned model

mRNA vaccine sequence design is not just "translating amino acids to bases". Codon usage frequency affects translation efficiency, UTR structure affects stability and loading efficiency. DiVo Gen²AI uses the self-tuned RNALens model to predict ribosome loading efficiency--not calling third-party APIs, but part of 560K lines of self-developed code. This page serves three audiences: patients and public wanting to understand mRNA optimization, vaccine/CAR-T teams seeking sequence optimization partners, and investors and peers evaluating technical capabilities.

Patients & Public

Read "What is mRNA Sequence Optimization" below--understand why vaccine design is not just translating bases.

Partners / Hospitals

Focus on "DiVo's Role", "4-Step Pipeline", "Benchmarks"--self-tuned, not API wrapper.

Investors / Peers

Focus on "Differentiation", "Benchmarks", "Glossary"--evaluate the technical barrier of self-tuned models.

mRNA sequence optimization is a key step in the following flagship services

What is mRNA Sequence Optimization

The core of an mRNA vaccine is a messenger RNA sequence--after being delivered into human cells, it directs cells to synthesize target proteins (e.g., tumor antigens), triggering immune response. But the same amino acid can be encoded by multiple different codons, and codon usage frequencies vary greatly.

mRNA sequence optimization is not simply translating amino acid sequences to bases--it is selecting the sequence with the highest translation efficiency and best stability from all candidate sequences that can encode the same protein. Codon choice affects translation speed (CAI), UTR structure affects mRNA half-life and ribosome loading (MRL), and different target tissues have different codon preferences.

DiVo uses the self-tuned RNALens model to predict ribosome loading efficiency (Spearman=0.92), not calling third-party APIs--fine-tuning data, training pipeline, and inference service are all self-deployed.

Why Codon Choice Matters

The same amino acid has 2-6 codons. Human cells translate some codons much faster than others--choosing the right codons can improve translation efficiency from 0.7 to 0.95 (CAI index).

Why Self-Tuning Matters

General pre-trained models have limited prediction accuracy for mRNA sequences. Three-stage fine-tuning makes RNALens specifically adapted to mRNA MRL tasks--Spearman=0.92 is not a zero-shot performance of a general model, but a specialized result after domain adaptation.

DiVo Gen²AI's Role

mRNA sequence optimization is Step 7 of DiVo's neoantigen 8-step pipeline, and the core capability for CAR construct mRNA optimization in the CAR-T 5-step pipeline. We provide end-to-end from codon optimization to ribosome loading prediction to UTR optimization--not just doing CAI optimization, but simultaneously optimizing coding and regulatory regions, with self-tuned model prediction for functional validation.

We do not do mRNA delivery system design (LNP formulation, etc.), do not do in vivo expression efficiency experimental validation. We deliver computationally optimized mRNA sequence design proposals.

Core Capability · 4-Step Pipeline

From amino acid sequence to tissue-adapted optimal mRNA sequence

MO1

Codon Optimization (CAI)

已验证

In-house codon optimization algorithm

CAI 0.7 -> 0.95 optimized sequence

MO2

Ribosome Loading Efficiency Prediction

已验证

RNALens (DNABERT-Z 117M fine-tuned)

MRL prediction + Spearman=0.92

MO3

UTR Optimization + Stability Assessment

已验证

GEMORNA + ViennaRNA

5'/3' UTR optimized sequence + secondary structure score

MO4

Tissue-Adapted Model Selection

已验证

HEK293T / Muscle / PC3

Target tissue-adapted optimal sequence

Differentiation

Core differences from "CAI-only" and "third-party API" approaches

AI

Self-Tuned, Not Third-Party API

RNALens is based on DNABERT-Z 117M parameter three-stage fine-tuning. Fine-tuning data, training pipeline, and inference service are all self-deployed. Not a wrapper around third-party model APIs, but part of 560K lines of self-developed code.

UTR

Optimizing Regulatory Regions, Not Just Coding

Most mRNA optimization services only do CDS codon optimization. DiVo uses GEMORNA to generate UTR + ViennaRNA to assess stability, simultaneously optimizing coding and regulatory regions--both translation efficiency and stability.

Core Capabilities

CAI Codon Adaptation Index Optimization

已验证

Improved from 0.7 to 0.95--codon usage frequency aligned with host cell preferences, greatly improved translation efficiency.

MRL Ribosome Loading Efficiency Prediction

已验证

RNALens model based on DNABERT-Z 117M parameter three-stage fine-tuning. Spearman=0.92, R²=0.87--predicts ribosome loading efficiency of mRNA sequences.

5'/3' UTR Optimization

已验证

GEMORNA generation + ViennaRNA stability assessment. Not just optimizing coding regions, but also regulatory regions.

3 Tissue-Adapted Models

已验证

HEK293T / Muscle / PC3 three target tissue/cell line adapted models--different tissues prefer different codons.

Why We Emphasize "Self-Tuned"

->

Not Calling Third-Party APIs

RNALens is a self-developed model based on DNABERT-Z 117M parameter three-stage fine-tuning. Fine-tuning data, training pipeline, and inference service are all self-deployed.

->

Not General Pre-Trained Weights

Three-stage fine-tuning: (1) DNABERT-Z 117M pre-training -> (2) mRNA MRL downstream task fine-tuning -> (3) DiVo self-developed data domain adaptation fine-tuning.

->

Has Real-World Validation

Asparaginase project: 6194 mutations × 11 iterations. CAI 0.7->0.95. Not a toy project.

Benchmarks

MetricValueNote
MRL Spearman0.92Ribosome loading efficiency prediction correlation
0.87Coefficient of determination
CAI optimized0.95Improved from 0.7
DNABERT-Z parameters117MThree-stage fine-tuning
Tissue-adapted models3HEK293T / Muscle / PC3
Real-world validationAsparaginase 6194 mutations11 iterations
0.92
MRL Spearman
0.87
0.95
CAI optimized
117M
DNABERT-Z params

Honest Boundaries

What we can and cannot do, clearly stated

What We Can Do

CAI codon adaptation index optimization (0.7 -> 0.95)
MRL ribosome loading efficiency prediction (Spearman=0.92)
5'/3' UTR sequence optimization + stability assessment
3 tissue-adapted models (HEK293T / Muscle / PC3)
Self-developed RNALens fine-tuned model end-to-end

What We Don't Do

No mRNA delivery system design (LNP formulation, etc.)
No in vivo expression efficiency experimental validation
No in vivo mRNA stability measurement
No clinical-grade GMP manufacturing process optimization
No direct-to-patient mRNA therapy proposals

Glossary

5 core terms in mRNA sequence optimization

Abbr.Full NameTranslationExplanation
CAICodon Adaptation IndexCodon Adaptation IndexMeasures how well codon usage frequency matches host cell preferences; higher means better translation efficiency
MRLMean Ribosome LoadingMean Ribosome LoadingAverage number of ribosomes on an mRNA, reflecting translation initiation efficiency
UTRUntranslated RegionUntranslated RegionRegulatory sequences at both ends of mRNA that don't encode proteins; 5' UTR affects translation initiation, 3' UTR affects stability
RNALensDiVo RNALensDiVo RNA Prediction ModelSelf-developed mRNA property prediction model based on DNABERT-Z 117M three-stage fine-tuning
GEMORNAGenerative mRNAmRNA Generative ModelGenerative model for generating and optimizing mRNA UTR sequences

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