The promise of multi-omics integration in target discovery is real: combining transcriptomics, proteomics, genomics, and phenomics data should in principle give a more complete picture of target biology than any single data type alone. The practical challenge is that these data types are measured at different biological levels, have different noise properties, and are often collected in different experimental contexts. Integrating them naively produces misleading results.
What Each Layer Contributes
Transcriptomics -- whether bulk RNA-seq, single-cell RNA-seq, or spatial transcriptomics -- measures gene expression at the RNA level. It is sensitive, relatively inexpensive, and well-supported by analytical infrastructure. The limitation is that mRNA levels correlate imperfectly with protein abundance. For some gene families, the mRNA-to-protein correlation is high and transcriptomics data is a reasonable proxy. For others, post-transcriptional regulation, protein stability, or translational efficiency creates systematic differences between RNA and protein levels that make transcriptomics a poor guide to protein activity.
Proteomics -- typically mass spectrometry-based -- measures the actual proteins present and their relative abundance. It is more directly relevant to drug targets, which are almost always proteins, but it is technically demanding, has lower throughput than transcriptomics, and covers a smaller fraction of the proteome than RNA-seq covers of the transcriptome. Label-free quantitative proteomics has improved substantially, but deep proteome coverage in a challenging tissue type remains a significant undertaking.
Genomics and GWAS data contribute the causal layer: which genetic variants are associated with disease risk, and which genes those variants likely affect. This is conceptually the strongest evidence type for target validity, but GWAS loci often require substantial additional work to assign to a specific gene and mechanism.
Phenomics -- systematic phenotypic characterization of genetic perturbations, typically from resources like the International Mouse Phenotyping Consortium or high-throughput CRISPR screen databases -- provides functional evidence for target biology. A gene whose knockout produces a phenotype relevant to the disease indication has functional support that is more direct than expression data alone.
The Integration Problem
These four data types do not live on a common scale, are not measured in the same experimental systems, and often conflict with each other in ways that are genuinely informative rather than attributable to noise. The naive approach to integration -- normalize each data type to a [0,1] scale and average them -- loses the information contained in conflicts and artificially smooths over uncertainty that should be visible in the composite score.
A more principled approach treats each data type as an independent evidence layer contributing a weighted component to a composite score, with weights informed by the reliability and relevance of each layer for the specific disease context. Genetic evidence carries higher weight because of its causal properties. Functional phenomics data carries higher weight when the phenotype is disease-relevant in a validated model. Transcriptomics and proteomics data are more context-dependent and should carry weights that reflect the data quality of the specific experiment rather than a fixed tier assignment.
Handling Sparse and Missing Modalities
A practical challenge in multi-omics integration is that real datasets are almost always incomplete. Most programs have data for some modalities and not others. A team might have excellent RNA-seq data from patient tissue and published GWAS summary statistics but no in-house proteomics data for the relevant cell type. An integration framework that requires all four modalities to produce a score will either return no result or impute missing values in ways that introduce systematic bias.
The right approach is to design the scoring model to handle missing modalities gracefully: using available evidence to compute a partial score while widening the confidence interval to reflect reduced evidence depth. A target with strong genetic and transcriptomic evidence but missing proteomics data should receive a score that reflects those two strong tiers while making the missing third tier visible as uncertainty, not treating it as zero evidence or as average evidence.
Batch Effects and Cross-Study Harmonization
When integrating omics data from multiple sources -- different studies, different cohorts, different experimental platforms -- batch effects are a serious concern. Two RNA-seq studies of the same tissue type can show systematic differences in gene expression that reflect differences in library preparation, sequencing depth, cell population composition, or sample handling rather than true biological differences. Integrating these studies without batch correction will produce confounded results.
The standard approaches -- ComBat, limma's removeBatchEffect, or platform-specific normalization methods -- can reduce technical variation when batch assignment is known. The harder case is when batch effects are confounded with biological variables of interest, or when studies were collected under different clinical definitions of the same disease. These scenarios require judgment about which normalization steps are appropriate and honest reporting of the uncertainty that remains after correction.
What Transcriptomics and Phenomics Add Together
The combination of transcriptomics and phenomics data is particularly informative when the two agree: a gene that is consistently upregulated in disease tissue and whose knockdown produces a relevant phenotype in a model system has converging evidence from two independent experimental approaches. When they disagree -- high expression but no phenotype on knockdown, or strong knockdown phenotype but low expression in disease tissue -- the disagreement points to specific mechanistic questions that should drive follow-up experiments rather than being averaged away.
Multi-omics integration done well does not produce a single confident number. It produces a structured evidence profile that makes explicit what is known and what is uncertain, and guides researchers toward the experiments that will most efficiently reduce that uncertainty.
Practical Guidance for Integration Projects
For teams starting a multi-omics integration project, the most important early decisions are about data quality and standardization rather than about integration methodology. Spending time on careful quality assessment and batch effect correction before beginning the integration will produce better results than applying a sophisticated integration method to poorly prepared data. The integration method can be chosen after the data quality issues are understood; the reverse order usually produces results that look convincing but are not reproducible.
Start with the simplest integration approach that appropriately handles your data structure -- typically a weighted combination of standardized scores with explicit missing-data handling -- before considering more complex methods like multi-omics factor analysis or network-based integration. Simpler methods are easier to audit, easier to explain to collaborators, and often perform comparably to more complex methods when the underlying data quality is high. Reserve complex methods for cases where the simpler approaches demonstrably fail to capture important patterns in the data.