Expand entity support: complexes & families

[jvwong]

Summary

Goal: Extend grounding support for complexes & families.
Why: Second largest class of Biofactoid entity grounding errors. Ubiquitous in the literature (protein gene - 56%; Family/complex - 17.7%). Been tabled for years.
How: Re-use FamPlex, a curated resource for disambiguation of (human) complexes and families.

Background

It is common for researchers to refer to complexes and members of a family. These authors may not be concerned or possibly even aware of the precise individual component(s) of a complex or member(s) of a family to which they refer, but rather wish to convey information about a general class of function or structure. The result is that authors name entities using these broader terms, with the implicit assumption that there are individual components/members.

Example: NF-κB

Nuclear factor kappa-light-chain-enhancer of activated B cells (NF-κB) is a protein complex that controls transcription of DNA, cytokine production and cell survival.

There are five proteins in the mammalian NF-κB family

Class	Protein	Aliases	Gene
I	NF-κB1	p105 → p50	NFKB1
I	NF-κB2	p100 → p52	NFKB2
II	RelA	p65	RELA
II	RelB		RELB
II	c-Rel		REL

Various NF-κB complexes

Fig. 1. A general model by which different NF-κB dimers contribute to selectivity of the transcriptional response to an NF-κB-inducing stimulus. The model shown is based on published studies of the selective functions of different NF-κB dimers, as discussed in the text. [Smale, S. T. Dimer-specific regulatory mechanisms within the NF-κB family of transcription factors. Immunol Rev 246, 193–204 (2012).]

In Biofactoid

Complexes and families represent the second largest class of errors in entity grounding. See PathwayCommons/factoid#1003 (comment).

Example

NF-κB-p62-NRF2 survival signaling is associated with high ROR1 expression in chronic lymphocytic leukemia. Sanchez-Lopez et al. Cell Death Differ. 2020 Jul;27(7):2206-2216

Phosphorylated RB Promotes Cancer Immunity by Inhibiting NF-κB Activation and PD-L1 Expression. Mol Cell
. 2019 Jan 3;73(1):22-35.e6.

Implementation

FamPlex is a resource that helps improve named entity recognition, grounding, and relationship resolution. The repository provides several comma-separated files that can be used to populate our grounding resource (Table I).

Table I. Relationship between FamPlex data and ground-search fields

FamPlex file.	Description	Count	ground-search field(s)
entities.csv	FamPlex namespaced entities	754	name/id.
descriptions.csv	Description text	431	summary
grounding_map.csv	Synonyms	2163 (FamPlex)	synonyms
equivalences.csv	Mappings	2489(FamPlex)	xrefs
relations.csv	Components, members	4711	?type?

Top entities referenced

Rank	Name	Count
1	ERK	6301
2	AKT	5839
3	NFkappaB	5768
4	TGFB	2877
5	PI3K	2486
6	JNK	2401
7	p38	2345
8	VEGF	2326
9	Cyclin	2087
10	Wnt	1622
11	Integrins	1498
12	RAS	1402
13	Actin	1299
14	PKC	1234
15	PKA	1058

Caveats

• Human bias, mostly if not exclusively
• categorization: "type" is contiguous rather than discrete ((a) family [ERK] (b) family of complexes [NFKappaB]) (c) Unique complex [IL23:IL12B+IL23A] - more the case that something is a family
• Includes other types to exclude (processes, etc)

References

• Bachman, J. A., Gyori, B. M. & Sorger, P. K. FamPlex: a resource for entity recognition and relationship resolution of human protein families and complexes in biomedical text mining. BMC Bioinformatics 19, 248 (2018).

[jvwong]

Entity type

When it comes to integrating with factoid model, it's important to assign some sort of 'type' to a Famplex entity. To attempt this, will we use the Famplex provided relations.csv containing:

• entities (describing complexes and families)
• (directed) relationships ('isa' and 'partof') between entities

A first attempt is to follow a simple heuristic: A complex namedComplex is an entity that has some other entity that is partof it. Otherwise, it is a family.

---

Example below for "AMPK":

[jvwong]

Turning this into an itemized issue, to stage the changes and reduce risk (in particular, changes in factoid):

• Add grounding-search as datasource (including tests), but do not expose in aggregate search
• Add Famplex entities of type family
• Add Famplex entities of type complex namedComplex