Biology Atlas

Map / Concepts

concept · intermediate · partly written

How omics layers are joined: by identifier, sample, cell, position, latent space and graph

Six ways layers are joined, the key each uses, what it must assume, and where it is documented to break.

The six joins

Layers are joined by identifier (a gene stands for its protein), by sample, by cell barcode, by position, by a statistical latent space, and by a knowledge graph. Each assumes something the data may not give, and measured results exist for only a few. 1234

Reported, not yet replicated

By identifier more

Key: a gene or protein identifier. Assumes the gene stands for its protein. In yeast, in one condition and about 150 protein spots, mRNA level was insufficient to predict protein level, with the same mRNA level carrying more than 20-fold different protein (abstract only read). Shared peptides make protein presence an inference: in one small human-cell dataset about 20% of proteins were non-conclusive. No later correlation figures were opened for this record. 15

Reported, not yet replicated

By sample (matched multiomics) more

Key: a sample id. Assumes the same material was measured and tolerates a missing layer. MOFA handled about 40% of 200 leukaemia patients lacking a layer; with 10 factors it explained 41% of drug-response variance, 38% of mRNA, 24% of methylation and 24% of mutations, in the method authors' own evaluation. 2

Reported, not yet replicated

By cell (CITE-seq, multiome) more

Key: the cell barcode. Assumes both modalities came from one cell. The anchor-based method's authors note that chromatin changes may precede expression in developing systems, in which case they expect RNA-to-ATAC anchors would fail, and that the surface proteins CD25 (gene IL2RA, interleukin 2 receptor subunit alpha, HGNC:6008) and CCR7 (C-C motif chemokine receptor 7, HGNC:1608) were predicted poorly from RNA. In the NeurIPS 2021 challenge, run by the makers of the dataset, a mutual-nearest-neighbour baseline was beaten by only 14 of 25 teams, and the winning online CITE-seq entry used simple linear dimensionality reduction. 367

Reported, not yet replicated

By a statistical latent space more

Key: shared cells or samples. Assumes linear factors (MOFA) or a shared cell state (anchors). In scIB, run by authors who also build methods, 68 method and preprocessing combinations were scored on 85 batches and 1.2 million cells; selecting highly variable genes beat using all genes in 74% of comparisons; and 27% of ATAC outputs beat the best unintegrated result on peaks, against 85% for RNA tasks. Details of the three main joint-embedding methods and their benchmarks are in a separate record. 48

Reported, not yet replicated

By position more

Key: coordinates or a shared coordinate system. Assumes sections and modalities can be aligned and that a transcript falls in the right cell. The measured weak points are in the record on segmentation and registration error. 910

Reported, not yet replicated

By knowledge graph more

Key: a grounded identifier. Assumes the edge is not an artefact of connectivity. Aiyappa and colleagues show, for link prediction in general and not for knowledge graphs specifically, that a method using only node degree can reach near-optimal scores under common edge sampling (abstract only read). No Hetionet-specific figure was opened here. 11

Reported, not yet replicated

What was not found more

No benchmark that joins RNA to protein or to metabolites was found. A 2024 to 2025 benchmark of 40 methods on 86 datasets was seen only as a bioRxiv record, so none of its figures is used. The NeurIPS dataset has 12 donors in the results paper and 10 in two other sources, and that conflict is unresolved. 7

Reported, not yet replicated

Plain answer

Layers are joined by an identifier, a sample, a cell barcode, a position, a shared latent space or a graph; each join assumes the layers describe the same thing, and the documented breaks are protein-versus-mRNA, missing layers, chromatin leading expression and segmentation error. 123

Reported, not yet replicated

Why it matters

Theoretical biology needs data placed together; the join key decides what can be placed together and what is silently lost. 4

Our inference

Real numbers

scIB design and results 2026-11-18

68 combinations, 85 batches, 1.2 million cells; 27% of ATAC outputs beat the best unintegrated peaks result against 85% for RNA tasks. 4

Reported, not yet replicated

NeurIPS 2021 challenge baseline 2026-11-18

A mutual-nearest-neighbour baseline was beaten by 14 of 25 teams on the joint-embedding task. 7

Reported, not yet replicated

Still open here

Not yet written: Equation.

Linked from

Sources

  1. Gygi SP, Rochon Y, Franza BR, et al.. Correlation between protein and mRNA abundance in yeast. Mol Cell Biol (1999). paper; opened 2026-11-17 all records citing this
  2. Argelaguet R, Velten B, Arnol D, Dietrich S, et al.. Multi-Omics Factor Analysis-a framework for unsupervised integration of multi-omics data sets. Mol Syst Biol (2018). paper; opened 2026-11-18 all records citing this
  3. Stuart T, Butler A, Hoffman P, Hafemeister C, et al.. Comprehensive Integration of Single-Cell Data. Cell (2019). paper; opened 2026-11-18 all records citing this
  4. Luecken MD, et al.. Benchmarking atlas-level data integration in single-cell genomics. Nature Methods (2022). paper; opened 2026-11-18 all records citing this
  5. Prieto G, Aloria K, Osinalde N, Fullaondo A, et al.. PAnalyzer: a software tool for protein inference in shotgun proteomics. BMC Bioinformatics (2012). paper; opened 2026-11-18 all records citing this
  6. HGNC REST records for IL2RA, CCR7, MS4A1, CD28, CRYM, KRAS and NOTCH3. HGNC. database, not peer reviewed; opened 2026-11-18 all records citing this
  7. Multimodal single cell data integration challenge: Results and lessons learned. Proceedings of Machine Learning Research 176 (2022). paper; opened 2026-11-18 all records citing this
  8. Argelaguet R, Arnol D, Bredikhin D, Deloro Y, et al.. MOFA+: a statistical framework for comprehensive integration of multi-modal single-cell data. Genome Biol (2020). paper; opened 2026-11-18 all records citing this
  9. Design document for SpatialData.. web, not peer reviewed; opened 2026-11-18 all records citing this
  10. Jones DC, Elz AE, Hadadianpour A, Ryu H, et al.. Cell simulation as cell segmentation. Nat Methods (2025). paper; opened 2026-11-18 all records citing this
  11. Rachith Aiyappa, Xin Wang, Munjung Kim, Ozgur Can Seckin et al.. Implicit degree bias in the link prediction task. arXiv (2024). preprint, not peer reviewed; opened 2026-11-18 all records citing this
Where it sits on the map
History of this record (1)

This record has been changed 1 time, on the dates below. The note kept against each change is written for the people who maintain the atlas, in their own shorthand, so it is not shown here; it is in the record itself.

Status
partly written
Last updated
2026-11-18
Record
content/concepts/how-omics-layers-are-joined.json