concept · intermediate · partly written
Context-dependent edges: storing the conditions under which a statement holds
A fact stored with the species or cell type it was seen in; every large schema has the slot, and what is open is how often it is filled and used.
His idea, in his words
"Can we make data more dynamic so a certain piece of data is connected to all these other pieces of data, but only in this context? You have conditional graph connections or entanglements, similar to quantum." 1
Amit's hypothesis
The idea stated so it can be wrong more
A statement A-relation-B is stored with the conditions it was observed under; a query returns only statements whose conditions match; and a model conditioned on context beats a model that pools across contexts when whole contexts are held out of training. This is our reading of his words, not a quotation. 23
Our inference
Where it already exists, under other names more
INDRA keeps context on each piece of evidence, not on the merged statement: an Evidence object carries disease, cell line and similar fields "when available from the source". Nanopublications split a statement into assertion, provenance and publication-information graphs, and more than 10 million exist; the specification defines no dedicated field for conditions such as species or tissue, so context must be written as extra statements. Tissue-specific network layers exist as OhmNet (107 tissue layers) and as PINNACLE (394,760 protein representations across 156 cell-type contexts). Hyper-relational models such as StarE report gains over triple-only models on Wikidata and Freebase-derived data, which are not biology. See also the records on Biolink, Wikidata and the statement shapes. 2456789
Reported, not yet replicated
How often context is actually recorded more
Measured fill rates are few. SIGNOR, a curated causal resource, states that approximately one third of its interactions record the experimental system. For the protein-interaction resource HIPPIE, cell-line information was recovered for 247,065 of 391,410 interactions (about 63%) by text mining after the fact, not recorded by the source databases. IntAct states that tissue and cell type are partly held as free text and publishes no rate. No fill rate was found for INDRA or Biolink. 101112
Reported, not yet replicated
What a query does with it: the pooled default more
The authors of QIAGEN's published causal-network method wrote that filtering literature findings by specific contexts "generally does not work well because it leads to networks that are too sparse", and pooled findings instead. INDRA's authors list incorporating biological context "in a principled manner" as work still to do and say polarity conflicts "can potentially be addressed by more thorough incorporation of biological context". 32
Reported, not yet replicated
The nearest controlled test more
Five single-cell foundation models and two deep models did not beat an additive or mean baseline on 62 held-out double perturbations in K562 cells. This tests learned representations of perturbation effects, not stored context as such, and is the closest trivial-baseline test found. 13
Reported, not yet replicated
Where it fails against real data more
Capture comes first: the fill rates above are a third to two thirds at best, so a context filter removes most statements (the sparsity the QIAGEN authors saw). Second, a degree-only baseline scores very high on a pooled graph (AUROC 0.95 or more on 17 of 20 Hetionet edge types), and embedding models rank densely connected entities highly "no matter the context"; splitting a graph by context gives each context its own sparser degree sequence, so the null must be rebuilt per context. No paper opened does that. 14153
Reported, not yet replicated
What remains open (our inference) more
On a graph that stores context, one edge-prediction task scored three ways, pooled, context-conditioned and degree-only, with whole contexts withheld, was not found. Nor was a published audit of context fill in INDRA or Biolink. 214
Our inference
Cheapest test (our inference) more
Count host organism, tissue and cell-type fields, free text and controlled terms apart, in a public IntAct export or a public INDRA statement dump for one gene family. Where context is filled, rank pooled against context-conditioned with whole cell types held out, against the degree-preserving prior and then against belief score alone. If context is filled for a small minority, the idea fails at capture before it reaches modelling. 12142
Our inference
Plain answer
Every large biomedical schema can attach context to a statement. What is thin is the filling in of that context and the use of it when a question is asked: the published numbers show one third to about two thirds filled, and the main builders of mechanism graphs describe using context as unfinished work. 10112
Reported, not yet replicated
Why it matters
Still open here
- What share of INDRA evidences carry a filled context object, and of Biolink edges a context qualifier?
- Does a context-conditioned model beat a pooled one on held-out contexts against a per-context degree null?
Not yet written: Equation, Real numbers.
Linked from
Sources
- Amit's dictation of 2026-11-18: how biology data is represented, and where ontologies can be innovated. Kept in the atlas repository; not published.
- INDRA paper on automated assembly of mechanistic knowledge from text and databases (PMC10167483). PubMed Central (2023). paper; opened 2026-11-17 all records citing this
- Krämer A, et al.. Causal analysis approaches in Ingenuity Pathway Analysis. Bioinformatics (2013). paper; opened 2026-11-18 all records citing this
- Kuhn T, Meroño-Peñuela A, Malic A, et al.. Nanopublications: A Growing Resource of Provenance-Centric Scientific Linked Data. arXiv (2018). preprint, not peer reviewed; opened 2026-11-18 all records citing this
- nanopub.net. Nanopublication Guidelines (working draft). nanopub.net. web, not peer reviewed; opened 2026-11-18 all records citing this
- Zitnik M, Leskovec J. Predicting multicellular function through multi-layer tissue networks. Bioinformatics (2017). paper; opened 2026-11-18 all records citing this
- Li MM, Huang Y, et al.. Contextual AI models for single-cell protein biology. Nature Methods (2024). paper; opened 2026-11-18 all records citing this
- Galkin M, Trivedi P, Maheshwari G, et al.. Message Passing for Hyper-Relational Knowledge Graphs. arXiv (EMNLP 2020) (2020). paper, not peer reviewed; opened 2026-11-18 all records citing this
- INDRA developers. INDRA documentation: Statements. indra.readthedocs.io (2026). web, not peer reviewed; opened 2026-11-18 all records citing this
- Lo Surdo P, et al.. SIGNOR 3.0, the SIGnaling network open resource 3.0: 2022 update. Nucleic Acids Research (2023). paper; opened 2026-11-18 all records citing this
- Federico A, Monti S. Contextualized Protein-Protein Interactions. Patterns (2021). paper; opened 2026-11-18 all records citing this
- Del Toro N, et al.. The IntAct database: efficient access to fine-grained molecular interaction data. Nucleic Acids Research (2022). paper; opened 2026-11-18 all records citing this
- Ahlmann-Eltze C, Huber W, Anders S. Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. Nat Methods (2025). paper; opened 2026-11-17 all records citing this
- Zietz M, Himmelstein DS, Greene CS, et al.. The probability of edge existence due to node degree: a baseline for network-based predictions. GigaScience (2024). paper; opened 2026-11-18 all records citing this
- Bonner S, et al.. Implications of topological imbalance for representation learning on biomedical knowledge graphs. arXiv 2112.06567 (2022). preprint, not peer reviewed; opened 2026-11-18 all records citing this
Where it sits on the map
History of this record (5)
This record has been changed 5 times, 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
- 2026-11-19
- Record
- content/concepts/context-dependent-edges.json