HIGH-THROUGHPUT STRUCTURAL DATA

High-throughput protein expression × Crosslinking MS

Protein structural data.High throughput. At scale.

Nirvana 1.0 combines gene synthesis, rapid protein expression, crosslinking MS and AI-powered analysis to generate structural data for proteins and complexes at scale.

Explore Nirvana 1.0 Compare methods

EXPRESS · CROSSLINK · MEASURE · LEARNExplore

OUR MISSION

More data.More capable models.

Nirvana 1.0 brings together high-throughput gene synthesis, 3-hour cell-free protein expression and crosslinking MS to produce structural data at ultra-high throughput. SuperNovA, our high-speed distributed analysis system, and RNovA, our AI model for mass spectrometry interpretation, turn experimental signals into structural data.

We aim to generate experimental evidence for underexplored proteins and interactions, giving AI more to learn from.

THE STRUCTURAL DATA GAP

Bigger models cannotreplace missing data.

01 / LIMITED SUPPLY

Limited structural data

Language models learn from vast text corpora. Experimental structure archives contain hundreds of thousands of entries. More parameters cannot create new experimental observations.

PDB structures250k+structure entries

Llama 3 training data15T+text tokens

02 / UNEVEN COVERAGE

Uneven data coverage

6,892Antibody–antigen complex structures
Across 700+ antigen targets

16%SARS-CoV-2 share of structures

Structural data are concentrated around a subset of research targets. More observations of familiar systems do not provide equivalent coverage of new targets, interfaces or conformations.

03 / DIFFERENT OUTCOMES

Uneven task performance

AlphaFold 3 · FoldBenchSuccess rate · DockQ ≥ 0.23

Protein–protein interfaces72.93%

Antibody–antigen interfaces47.90%

An overall benchmark score can hide differences between tasks. Challenging problems such as antibody–antigen binding still show substantial performance gaps.

1 / 3 · Limited structural dataSwipe sideways

Sources

RCSB · 250k milestone Meta · Llama 3 model card

THE STRUCTURAL DATA BOTTLENECK

Bigger models cannotreplace missing data.

Adding parameters alone cannot make up for missing experimental data. Limited samples and uneven coverage constrain what models can learn about protein structures and interactions.

01 / LIMITED EXPERIMENTAL DATA

Limited structural data

Language models learn from vast text corpora. Experimental structure archives contain hundreds of thousands of entries. More parameters cannot create new experimental observations.

Swipe sideways to explore the evidence SCROLL THROUGH THE EVIDENCE

Limited structural data

Language models learn from vast text corpora. Experimental structure archives contain hundreds of thousands of entries. More parameters cannot create new experimental observations.

A gap in the supply of experimental data

PDB · Experimental structures250k+experimental structure entries

Llama 3 · Pretraining corpus15T+text tokens

RCSB · 250k milestone Meta · Llama 3 model card

Sources

RCSB · 250k milestone Meta · Llama 3 model card

THROUGHPUT & EFFORT · FROM SAMPLE TO DATA

PDB-scale protein structure data in under a month.PDB-scale protein structure data in under a month.

Less stringent purity requirements · Automated preparation · Pooled analysis

Crosslinking MS

>400proteins / MS
/ hour

Crosslink separately. Digest, pool and measure.

Automated preparation: >96 samples
Sample requirements: No crystals · less stringent purity

What sets the pace in traditional methods?

X-ray crystallography

Fast collection. A long search for crystals.

Difficult crystallization projects can take 1–2 years, sometimes 3–5.

Cryo-EM

About 1 day / sample

Prepare and screen grids, collect each sample, then process the images.

	X-RAY CRYSTALLOGRAPHY X-ray crystallographyX-ray Target-specific crystallization	CRYO-ELECTRON MICROSCOPY Cryo-EMCryo-EM Per-sample preparation and collection	CROSSLINKING MS Crosslinking MSXL-MS No crystals · parallel preparation · pooled analysis

	X-ray crystallography	Cryo-EM	Crosslinking MS
Sample preparationSample prepPurity · conditions · processing	1–2 years, up to 3–5 For difficult targets in our team’s experience, repeated screening and crystal optimization can take years. Screen and optimize crystal conditions High purity · high homogeneity	Iterative grid preparation Optimize ice thickness, particle distribution and orientation, then select grids for data collection. Optimize ice and particle orientation High purity · low heterogeneity	No crystals. >96 in parallel. Crosslink each sample separately, follow with standard digestion, then pool the peptides for analysis. Crosslink separately → digest → pool Less stringent purity requirements
Data acquisitionOnce the sample is ready	Fast acquisitionOnce suitable crystals are available Data collection can be fast once crystallization is established. Finding suitable conditions may take much longer. Fast once suitable crystals are ready	≈1 day / sampleTypical microscope acquisition time Collect particle images per sample, followed by image processing and reconstruction. Particle images → density reconstruction	>400 proteins / MS / hourPer MS · continuous batch processing One MS run measures crosslinks from multiple independently prepared samples. Read multiple systems in one run

Sample preparationSample prepPurity · conditions · processing

1–2 years, up to 3–5

For difficult targets in our team’s experience, repeated screening and crystal optimization can take years.

Screen and optimize crystal conditions

High purity · high homogeneity

Iterative grid preparation

Optimize ice thickness, particle distribution and orientation, then select grids for data collection.

Optimize ice and particle orientation

High purity · low heterogeneity

No crystals. >96 in parallel.

Crosslink each sample separately, follow with standard digestion, then pool the peptides for analysis.

Crosslink separately → digest → pool

Less stringent purity requirements

Data acquisitionOnce the sample is ready

Fast acquisitionOnce suitable crystals are available

Data collection can be fast once crystallization is established. Finding suitable conditions may take much longer.

Fast once suitable crystals are ready

≈1 day / sampleTypical microscope acquisition time

Collect particle images per sample, followed by image processing and reconstruction.

Particle images → density reconstruction

>400 proteins / MS / hourPer MS · continuous batch processing

One MS run measures crosslinks from multiple independently prepared samples.

Read multiple systems in one run

	X-ray crystallography	Cryo-EM	Crosslinking MS
Total effortFrom sample to data	Crystallization sets the pace Time Years of condition screening1–2 years, or 3–5 for some difficult targets our team has worked on, plus acquisition and analysis. Workload Repeated purification and screeningPurify → screen ⇄ optimize crystals → collect. New targets often require a new search.New targets often need a new search	Grid optimization + acquisition Time Preparation + ~1 day / samplePreparation and screening + about 1 day per sample + processing and reconstruction. Workload Per-sample optimizationPurify → prepare ⇄ screen grids → collect → process images, sample by sample.Prepare and collect each sample	More samples, one shared workflow Time >400 proteins / MS / hourEach MS processes more than 400 proteins per hour. Workload >96 samples in parallelAutomate more than 96 samples per batch. Crosslink and digest each sample separately, then pool for analysis.Pool peptides after crosslinking and digestion

Total effortFrom sample to data

Crystallization sets the pace

Time: Years of condition screening1–2 years, or 3–5 for some difficult targets our team has worked on, plus acquisition and analysis.
Workload: Repeated purification and screeningPurify → screen ⇄ optimize crystals → collect. New targets often require a new search.New targets often need a new search

Grid optimization + acquisition

Time: Preparation + ~1 day / samplePreparation and screening + about 1 day per sample + processing and reconstruction.
Workload: Per-sample optimizationPurify → prepare ⇄ screen grids → collect → process images, sample by sample.Prepare and collect each sample

More samples, one shared workflow

Time: >400 proteins / MS / hourEach MS processes more than 400 proteins per hour.
Workload: >96 samples in parallelAutomate more than 96 samples per batch. Crosslink and digest each sample separately, then pool for analysis.Pool peptides after crosslinking and digestion

	X-ray crystallography	Cryo-EM	Crosslinking MS
Data output	Diffraction → electron density and structural modelsDiffraction → structure	Particle images → 3D density and structural modelsDensity → structure	Crosslinks → distance restraints and AI structural modelsCrosslinks → restraints and structure

TWO ROUTES TO STRUCTURE

Different evidence. One structural goal.

Both routes combine sequence, experimental evidence and structural rules to build a 3D model.

Loading the demonstration

CRYO-EM

From density to a 3D structure

Density reveals spatial features

Inspect the full map, trace the backbone through recognizable density and assign residues using the known sequence.

CROSSLINKING MS

From spatial clues to a complete structure

Crosslinks constrain residue distances

Combine quantified, calibrated events from crosslinkers of different lengths into a map of distance restraints.

Swipe vertically to explore · Use a large, fast swipe to leave

EVIDENCE IN ACTION

Experimental evidence.More accurate predictions.

Individual protein folds do not fully determine how a complex assembles. Crosslinking mass spectrometry adds spatial evidence to help AI predict how proteins bind.

TNF-α × infliximab Fab

Crosslink restraints improve the predicted antigen position relative to the crystal structure, 4G3Y. The experimental position is shown in silver. Drag to compare.

Structural agreement · DockQ ↑: 0.0260.668
Interface deviation · iRMSD ↓: 15.0 Å1.4 Å

Public AlphaLink2 prediction: violet TNF-alpha closer to the silver experimental position; antigen–Fab DockQ 0.668 — Drag to explore the difference。Or use the arrow keys. Home and End reveal one side in full.Blue: FabViolet: antigenSilver: experimental position

Public AlphaFold-Multimer prediction: blue antibody and violet TNF-alpha, displaced from the silver experimental antigen position; antigen–Fab DockQ 0.026 — Drag to explore the difference。Or use the arrow keys. Home and End reveal one side in full.Blue: FabViolet: antigenSilver: experimental position

SEQUENCE IN. STRUCTURAL DATA OUT.

Nirvana 1.0Protein structural data at ultra-high throughput

Turn human-designed or AI-generated sequences into protein structural data at scale.

START WITH SEQUENCES

Human-designed sequences

AI-generated sequences

LARGE-SCALE PARALLEL PRODUCTION An automated gene synthesis and protein expression system connected to a liquid chromatography–mass spectrometry instrument, illustrated in blue and violet with a stream of golden signals.

An automated gene synthesis and protein expression system connected to a liquid chromatography–mass spectrometry instrument, illustrated in blue and violet with a stream of golden signals.

NIRVANA 1.0 · 02 / 03

From sequence to structural data

01
High-throughput gene synthesis
Synthesize target genes in parallel
02
3-hour cell-free expression
Produce protein samples rapidly
03
High-throughput crosslinking MS
Crosslink separately · pool after digestion
04
SuperNovA MS analysis
Experimental signals → structural data

NIRVANA 1.0 · 03 / 03

Structural data, produced at scale

3hours

Rapid cell-free protein expression

>96samples in parallel

Automated batch preparation

>400proteins / MS / hour

Structural data acquisition

CORE ALGORITHMSRNovAAI for mass spectrometryExplore RNovA

STRUCTURAL DATA

Single proteins
Protein–protein
Protein–cyclic peptide
Protein–small molecule

THREE CORE CAPABILITIES

Three core strengths.Stronger together.

High-throughput production and protein characterization build our experimental dataset. As it grows, we plan to develop AI protein design and bring new sequences back into the lab for validation.

SuperNovA and RNovA turn mass spectrometry signals into data for model learning.

DATAEXPERIMENTAL DATAACCUMULATE · LEARN · REFINE

Produce

High-throughput gene synthesis3-hour cell-free protein expression

Characterize

Crosslinking MS analysisProtein characterization & sequencing

Design

IN DEVELOPMENT

AI-guidedprotein design

Planned following 3–4 more months of data collection

ABOUT NIRVANA

More structural data.Through AI and high-throughput science.

Based in Zhangjiang, Shanghai, NirvanaBio is developing Nirvana 1.0, an ultra-high-throughput platform for protein structural data, and RNovA, an AI model for mass spectrometry.

OUR TEAM

Founders

Zeping MaoPhD

CEO & CTO

Associate Professor, Yangtze Delta Region Institute of Tsinghua University, Zhejiang
BSc, School of Life Sciences, Tsinghua University
Visiting Scholar in Synthetic Biology, Caltech
PhD in Computer Science, University of Waterloo, supervised by Prof. Ming Li
Former Core Algorithm Engineer, Bioinformatics Solutions Inc.
Publications in Nature Biotechnology, Nature Machine Intelligence, Nature Communications and other leading journals
Unlocking Disease-Linked Protein Changes Using Artificial Intelligence
Computer Scientists Develop Zero-Shot Algorithm for De Novo Sequencing of Post-Translationally Modified Peptides
Unlocking the Secrets of Cells with Artificial Intelligence

Founders

Ming LiFRSC

Co-Founder

Fellow of the Royal Society of Canada
Internationally Recognized Computer Scientist
Pioneer in Information Theory and Computational Biology
2024 IEEE W. Wallace McDowell Award, a top international computing honor
ACM, IEEE and ISCB Fellow
Killam Prize, a major honor in engineering and natural sciences
Leading Contributor to Kolmogorov Complexity Theory
University Professor Ming Li Elected Fellow of the Canadian Academy of Engineering
Ming Li Receives 2024 IEEE Computer Society W. Wallace McDowell Award

OUR TEAM

Core Team

Core Team 1 / 3

Jie Liu

Principal Scientist, Biotechnology R&D

PhD & BSc, Life Sciences
Tsinghua University

Yixin Zhang

Chief Business Officer(CBO)

MBA, Peking University
BSc, Tsinghua University

Core Team 2 / 3

Zhuo Chen

Artificial Intelligence Scientist

PhD, Computer Science
University of Waterloo

Yandong Zhu

Digital Technology Director

MMath & BCS
University of Waterloo

Core Team 3 / 3

Yuxin He

Strategy Development Director

PhD, Biomedical Engineering
Master of Management
City University of Hong Kong

Shuai Zhang

Biomanufacturing Director

MSc, Biology
Anhui Agricultural University
Pharmaceutical Marketing

UNIVERSITIES & RESEARCH INSTITUTES

Academic Partnerships

Tsinghua University
University of Waterloo
Yangtze Delta Region Institute of Tsinghua University
Chinese Academy of Sciences
City University of Hong Kong
Shandong University

Swipe to explore our academic partners

Bayer Co.Lab Side by side in Shanghai.

Group photo at NirvanaBio’s Bayer Co.Lab location in Shanghai

NirvanaBio, Bayer Co.Lab, Building R02, 990 Guoshoujing Road, Pudong, Shanghai

View map

Protein structural data.High throughput. At scale.

More data.More capable models.

Limited structural data

Uneven data coverage

Uneven task performance

Bigger models cannotreplace missing data.

Limited structural data

A gap in the supply of experimental data

More data for some targets. Gaps for others.

One model. Different results across tasks.

PDB-scale protein structure data in under a month.PDB-scale protein structure data in under a month.

Crosslinking MS

X-ray crystallography

Cryo-EM

X-ray crystallographyX-ray

Cryo-EMCryo-EM

Crosslinking MSXL-MS

Different evidence. One structural goal.

From density to a 3D structure

From spatial clues to a complete structure

Experimental evidence.More accurate predictions.

TNF-α × infliximab Fab

Nirvana 1.0Protein structural data at ultra-high throughput

Human-designed sequences

AI-generated sequences

From sequence to structural data

High-throughput gene synthesis

3-hour cell-free expression

High-throughput crosslinking MS

SuperNovA MS analysis

Structural data, produced at scale

Three core strengths.Stronger together.

Produce

Characterize

Design

More structural data.Through AI and high-throughput science.

Zeping MaoPhD

Ming LiFRSC

Jie Liu

Yixin Zhang

Zhuo Chen

Yandong Zhu

Yuxin He

Shuai Zhang

Bayer Co.Lab Side by side in Shanghai.

Let’s build structural datasets.At scale. Together.