Directives de soumission d'échantillons
From Research Question to Multi-Omics Plan
Multi-omics is most useful when the research question spans biological layers that a single assay cannot resolve. The goal is not to collect the maximum number of datasets. We help you decide whether the project needs one focused modality, a two-layer validation design, or a broader integrated program.
Biomarker discovery
Identify molecular features or multi-layer signatures that distinguish prespecified research groups or outcomes, then evaluate stability and interpretability before prioritizing a candidate panel.
Target discovery and prioritization
Combine genomic, expression, regulatory, protein, pathway, and public evidence to rank candidate genes or network nodes for downstream validation.
Molecular subtyping
Use integrated molecular structure to discover reproducible sample groups, quantify cluster stability, and identify subtype-defining features and pathways.
Drug response and mechanism research
Separate baseline response-associated features from treatment-induced changes, then prioritize resistance programs, pathway effects, and follow-up experiments.
Cell-state and tissue-context studies
Use single-cell and spatial data to identify the cell populations, regulatory states, tissue domains, and neighborhoods that explain signals hidden in bulk measurements.
Our existing Multi-Omics Analysis capabilities provide the broader omics foundation, while this AI-assisted solution adds research-question routing, model selection, validation controls, and task-specific interpretation.
Choose the Evidence Layers
Each data layer should have a defined role in the study. We review how the layers complement one another before deciding which assays or existing datasets belong in the integrated analysis.
| Evidence Layer | Typical Input | What It Contributes |
|---|---|---|
| Genomics | FASTQ, BAM, VCF, mutation/CNV matrices | Genetic background, variants, copy-number changes, structural context |
| Transcriptomics | FASTQ, BAM, count or expression matrices | Differential programs, pathway activity, regulatory responses |
| Epigenomics | Methylation calls, ATAC-seq peaks, ChIP-seq signals | Regulatory state, chromatin accessibility, methylation-expression relationships |
| Proteomics / PTM | LC-MS/MS raw data or quantitative matrices | Protein abundance, signaling, pathway activation, functional-layer evidence |
| Metabolomics / Lipidomics | Raw MS data, peak tables, annotated matrices | Metabolic state, pathway remodeling, downstream biochemical phenotype |
| Single-Cell Data | FASTQ, count matrices, Seurat/AnnData objects | Cell types, states, trajectories, rare populations and heterogeneity |
| Spatial Omics | Platform outputs, matrices, images, coordinates | Tissue architecture, spatial domains, neighborhoods and cell-cell context |
| Phenotype / Metadata | Groups, outcomes, treatment, time, dose, covariates | Defines contrasts, supervised endpoints, confounders and interpretation context |
Proteomics and metabolomics as integrated capabilities: AI-enhanced DIA proteomics analysis can be included after DIA data generation or from existing quantitative matrices. Machine-learning metabolomics analysis can likewise be included for feature selection, classification, pathway interpretation, and cross-omics integration. These capabilities do not require separate standalone service pages to be used inside a broader project.
Project Input Requirements
| Input Category | Accepted Starting Point | Information Needed at Kickoff |
|---|---|---|
| Biological samples | Study-dependent material for selected omics assays | Sample type, species, preservation, group design, available material, pairing across assays |
| Raw sequencing data | FASTQ/BAM plus platform metadata | Reference build, library method, run/batch identifiers, sample manifest |
| Processed sequencing data | VCF, count matrix, peak file, methylation matrix | Processing method, genome build, filtering and normalization history |
| Proteomics / metabolomics data | Raw MS files or quantitative feature matrices | Instrument/acquisition mode, batch order, annotation state, prior missing-value handling |
| Single-cell / spatial objects | Seurat, AnnData, matrices, coordinates or images | Platform, preprocessing history, donor/sample IDs, batch structure, annotation status |
| Public or external cohorts | Downloaded data plus accession/provenance | Cohort definition, platform, usage constraints and compatibility with the primary study |
Find the Right AI-Assisted Solution
The hub routes projects by the scientific task. A single program can move through more than one route as the evidence matures.
| Research Task | Recommended Solution Route | Typical Output |
|---|---|---|
| Integrate two or more matched omics layers | AI-Assisted Multi-Omics Data Integration | Harmonized matrices, latent factors, cross-omics networks, pathway interpretation |
| Identify a stable candidate signature | Machine Learning Biomarker Discovery | Stability-ranked features, nested-CV performance, interpretable candidate panel |
| Rank candidate targets using convergent evidence | AI-Assisted Target Discovery and Prioritization | Target ranking, evidence matrix, pathways/networks, validation priorities |
| Discover molecularly distinct sample groups | AI-Assisted Molecular Subtyping | Cluster stability, subtype assignments, defining features, reproduction plan |
| Resolve cell-state-specific biology | AI-Assisted Single-Cell Multi-Omics Analysis | Integrated cell states, annotation, trajectories, regulatory programs |
| Preserve tissue architecture and molecular location | AI-Assisted Spatial Omics Analysis | Spatial domains, deconvolution, neighborhoods, cell-cell interaction context |
| Study treatment response, resistance, or mechanism | AI-Assisted Multi-Omics Drug Response and MOA Analysis | Response signatures, MOA/resistance networks, follow-up priorities |
If your project needs an experimental single-cell foundation, explore our Single-Cell Sequencing capabilities. Tissue-context studies can draw on Spatial Multi-Omics Sequencing Services. High-throughput perturbation programs can use Drug-seq to generate transcriptomic response data for later integration.
For example, a drug-response program may begin with data integration, move into biomarker discovery, and then use single-cell or spatial evidence to localize the response signal to a specific cell state or tissue neighborhood.
Project Entry Options
Data-to-Insight
You provide existing raw files or processed matrices. We review provenance, metadata, group structure, missingness, and compatibility before selecting the integration and modeling strategy.
Sample-to-Insight
You provide biological samples. Qualified partner platforms generate the selected omics data, after which the project enters the same QC, integration, validation, and interpretation framework.
Hybrid Project
Some layers already exist while another layer is generated specifically to close an evidence gap. This can avoid repeating data generation that is already fit for purpose.
Our Biomarker Research and Drug Development pages provide application context, while this hub focuses on coordinating the data and analytical decisions.
How AI and Machine Learning Fit the Workflow
We do not label routine preprocessing as AI. Machine learning is introduced only where it has a defined analytical function and can be evaluated against the study design.
Pattern discovery and integration
- Unsupervised integration: Factor models, network methods, embeddings, and related approaches can identify shared and layer-specific variation without a predefined outcome.
- Subtype and similarity analysis: Integrated molecular structure can support clustering, latent-factor discovery, and sample-similarity analysis.
- Multimodal representation: Deep generative approaches can be considered when data scale, nonlinear structure, or missing modalities justify the added complexity.
Prediction and prioritization
- Supervised modeling: Classification or regression is used only when a defined endpoint and adequate design support it.
- Interpretability: Feature attribution, pathway enrichment, network context, and evidence matrices convert model outputs into research priorities.
- Validation: Preprocessing, feature selection, and tuning remain inside training folds; external or hold-out validation is used where feasible.
A high model-importance score is treated as prioritization evidence, not proof of biological causality. Simpler models may be preferable when cohort size is limited or interpretability is the primary requirement.
Integrated Multi-Omics Workflow
One horizontal workflow connects project planning, data or sample intake, layer-specific preprocessing, AI-assisted integration, and biological interpretation.
Step 1: Research Question & Study Design
Define the biological decision, contrasts, outcome variables, covariates, sample relationships, and which molecular layers can add nonredundant evidence.
Step 2: Data / Sample Reception & QC
Review submitted files and metadata, or coordinate partner-platform data generation; document batches, missingness, and provenance.
Step 3: Layer-Specific Preprocessing
Process each omics layer with modality-appropriate normalization, filtering, batch assessment, and feature annotation before integration.
Step 4: AI-Assisted Integration & Validation
Apply factorization, network, supervised, or multimodal methods as justified; protect train/test separation and evaluate model or cluster stability.
Step 5: Biological Interpretation & Delivery
Map integrated signals to pathways, networks, targets, biomarkers, cell states, spatial context, or treatment mechanisms and deliver reproducible outputs.
Validation and Study Design Safeguards
Multi-omics can produce visually compelling patterns even when the underlying design is weak. We therefore treat study design and validation as part of the service rather than an afterthought.
- Batch and platform effects: assessed within each layer before integration; correction is not applied blindly when batch is confounded with condition.
- Missing modalities: documented by sample and layer; the integration strategy is selected around the observed overlap pattern.
- Data leakage: feature selection, imputation, scaling, and model tuning are confined to training data for supervised analyses.
- Small cohorts: exploratory, unsupervised, or hypothesis-generating analyses may be more defensible than predictive modeling.
- Class imbalance and covariates: outcome imbalance and relevant donor, tissue, treatment, or technical variables are incorporated where supported.
- External validation: independent cohorts or hold-out data are used when available; otherwise the evidence level is reported accordingly.
- Reproducibility: software versions, model settings, code, and methods-ready descriptions are documented according to project scope.
When Multi-Omics Is Not the Best Starting Point
We may recommend a narrower design when one well-powered assay can already answer the question, when omics layers come from incomparable samples, when a key phenotype is undefined, or when cohort size is too small for the proposed supervised model. Adding modalities does not automatically add evidence if the design cannot connect them.
Deliverables
- Study-design and technology-selection summary
- Per-layer QC and preprocessing outputs
- Harmonized matrices and sample/feature metadata
- Cross-omics factors, correlations, networks, or integrated embeddings
- Candidate biomarkers, targets, subtypes, response features, or cell/spatial states according to the selected route
- Model validation, stability, calibration, and interpretability outputs when supervised ML is used
- Pathway and network interpretation with evidence traceability
- Reproducible analysis code and environment information according to scope
- Methods-ready analysis description and final research report
References
- Baião AR, Cai Z, Poulos RC, et al. A technical review of multi-omics data integration methods: from classical statistical to deep generative approaches. Briefings in Bioinformatics. 2025;26(4):bbaf355. Baião et al., 2025
- Liu Y, Zhu K, Peng W, Liu Z, Mao X. Multi-omics and artificial intelligence for precision drug discovery and potential clinical applications. Signal Transduction and Targeted Therapy. 2026;11:210. Liu et al., 2026
- Argelaguet R, Arnol D, Bredikhin D, et al. MOFA+: a statistical framework for comprehensive integration of multi-modal single-cell data. Genome Biology. 2020;21:111. Argelaguet et al., 2020
- Walsh I, Fishman D, Garcia-Gasulla D, et al. DOME: recommendations for supervised machine learning validation in biology. Nature Methods. 2021;18:1122–1127. Walsh et al., 2021
Demo Results
A typical report can connect three views in one evidence chain: integrated sample structure, cross-omics relationships, and a decision-oriented summary of the next priorities. The example below represents output types rather than guaranteed performance benchmarks.
Integrated sample map
Latent factors or low-dimensional embeddings show whether major study patterns are shared across molecular layers and whether technical batches dominate the structure.
Cross-omics evidence network
Genes, regulatory features, proteins, metabolites, or pathways are connected according to the selected integration strategy and evidence thresholds.
Priority summary
Candidate biomarkers, targets, subtypes, mechanisms, or follow-up experiments are ranked with their supporting evidence and validation status.
AI-Assisted Multi-Omics Solutions FAQs
1. Do I need matched samples across every omics layer?
Matched samples provide the cleanest vertical integration, but complete overlap is not always required. We first map which samples are shared by each layer, then choose an integration strategy that can handle the actual overlap pattern without treating unobserved modalities as measured data.
2. How do I know whether I need multi-omics at all?
Start with the biological decision. If one assay can directly answer the question with adequate power, adding extra layers may increase analytical burden without improving the evidence. Multi-omics is most useful when the hypothesis crosses regulatory, expression, protein, metabolic, cellular, or spatial levels.
3. Can you analyze data generated by different vendors or public databases?
Yes. Data-to-Insight projects can include externally generated files when formats, metadata, reference builds, processing history, and sample provenance are sufficiently documented. Cross-platform differences are assessed before direct integration.
4. Can DIA proteomics and metabolomics be included even without separate AI service pages?
Yes. DIA proteomics and metabolomics can function as data-generation or analysis modules inside a broader project. AI-assisted feature selection, classification, pathway analysis, and cross-omics integration can be scoped without requiring a separate standalone service page.
5. Which AI method will you use?
There is no single default model. We select among correlation-based, matrix-factorization, network, classical machine-learning, or deep multimodal approaches according to sample size, number of layers, missingness, outcome type, and the required level of interpretability.
6. What if my sample size is small?
Small cohorts can still support descriptive, pathway-level, correlation, factor, or hypothesis-generating analyses, but they may not justify a high-dimensional predictive model. We adjust the analytical goal rather than presenting unstable prediction as a validated result.
7. Can single-cell and spatial data be integrated with bulk omics?
Yes, when the study design provides a meaningful bridge between modalities. Single-cell data can resolve the cell states behind a bulk signal, while spatial data can place those states back into tissue context. The strategy depends on whether datasets are matched by sample, donor, tissue, or comparable biological condition.
Related Publications
A technical review of multi-omics data integration methods: from classical statistical to deep generative approaches
Journal: Briefings in Bioinformatics
Year: 2025
Multi-omics and artificial intelligence for precision drug discovery and potential clinical applications
Journal: Signal Transduction and Targeted Therapy
Year: 2026
MOFA+: a statistical framework for comprehensive integration of multi-modal single-cell data
Journal: Genome Biology
Year: 2020
For Research Use Only. Not for use in diagnostic or clinical procedures.
