Timeline to discontinued use of old Phase 2 sources / ingests in RTX-KG2

[mbrush]

It has come to my attention that ARAX is still using an RTX-KG2 KG that is comprised mostly of old Phase 2 ingests to support ARAX queries and reasoner predictions, and generate explanation/support graphs for its predicted results. So the majority of their knowledge graph is out of date with ingest and modeling conventions, standards, and practices implemented during the refactor.

Potential Issues:

• Many of the modeling refactors and knowledge representation conventions that were implemented during the Phase 3 re-ingest work will not be implemented in RTX-KG2 - leaving much data that is inconsistently or incorrectly represented in the system.
• A lot of the source filtering and content filtering that was done in the Phase 3 ingests will not be implemented either - leaving in low quality / relevance data.
• Most of the one hop predictions that I just worked with ARAX to regenerate will not work - as they rely on features implemented in the DINGO re-ingests (CTD, TMKP, CTKP).
• All of these problematic, duplicative, and inconsistent edges, which we worked hard to improve during the phase 3 ingest refactor, will continue to show up in UI results and support paths.

We should discuss implications of this for various Translator efforts, and agree on a timeline and priority for ARAX moving over to use the Phase 3 Translator ingests to build an updated RTX-KG2 artifact, and re-train their prediction algorithms over the new graphs for ARA reasoning. It is ok to still decorate this custom graph with annotations / additional edges needed for scoring/reasoning. But IMO the main knowledge content should draw on updated Phase 3 ingests as much and as quickly as possible.

[mbrush]

Specifics from Adil about what old Phase 2 sources are in RTX-KG2, and what has been replaced by new Phase 3 Sources (see Slack messages here).

Phase 2 ingests in RTX-KG2 but NOT in Phase 3 translator-ingests:
umls (includes omim, rxnorm, nci, hgnc, nddf, pdq, psy, vandf as sub-sources), drugbank, ensembl, hmdb, jensenlab, kegg, mirbase, reactome, smpdb, unichem, unii, uniprotkb, disgenet, clinicaltrialskg (KG2 version of ctkp), drugapprovalskg (KG2 version of dakp), 30+ ontologies (fma, bfo, chebi, go, mondo, doid, hp, uberon, cl, foodon, ordo, pato, ro, nbo, pr, ino, genepio, mi, bspo, ehdaa2, ddanat, taxslim), semmeddb (old RTX-KG2 ingest)

Phase 3 translator-ingests but NOT in RTX-KG2:
alliance, bgee, bindingdb, cohd, ctd, ctkp, dakp, diseases, drug_rep_hub, gtopdb, gene2phenotype, geneticskp, go_cam, hpoa, icees, panther, pathbank, sider, signor, tmkp, ttd, ubergraph, semmeddb (new Phase 3 ingest)

Overlap (in both):
chembl, dgidb, drugcentral, goa (go annotations), intact, ncbi_gene, semmeddb (semmeddb coming from RTX-KG2 ingest)

Different ingests of same source in RTX-KG2 vs translator-ingests:
semmeddb, ctkp, dakp

---

In short, the bulk of the data that ARAX is using for reasoning and to generate support paths shown in the UI is coming from old Phase 2 ingests.

[dkoslicki]

Just FYI, in the Shepherd version of ARAX (i.e. the refactored stuff), the only place that RTX-KG2 data shows up in the refactored system is via the creative queries. So

In short, the bulk of the data that ARAX is using for reasoning and to generate support paths shown in the UI is coming from old Phase 2 ingests.

is not quite accurate. Particularly since Shepherd ARAX gets results from Shepherd to do any one-hop inference or lookups. I.e. once we retrain xDTD on the DogPark graph, you will see all the issues like this go away.

[mbrush]

Thanks for the clarifications @dkoslicki. A few follow ups.

• Can you confirm that the one hop inference stuff that we worked on with Chunyu a couple weeks back are based on the new Phase 3 translator-ingests?
• What is the rough timeline for retraining xDTD? And will MVP2 queries also be retrained / generated from the new Phase 3 translator-ingests as well?
• after all retrainings, will all of the old Phase 2 sources in the current KG2 no longer be used, such that edges form these sources won't be surfaced in support paths in any of our UIs?

[dkoslicki]

1. Yes, those one hop inferences are based on the new Phase 3 translator ingests (or whatever lives in Retriever at any given moment). To be technically accurate, the answer is yes/and: that logic was implemented in the ARAX code base, so will come through if you hit legacy ARAX, the same logic will apply on legacy translator graphs. If you hit Shepherd ARAX, that logic will apply to the phase 3 ingest graph.
2. Retraining xDTD will take approximately a month (or at least, merely the compute phase took a month when it was trained on RTX-KG2, additional dev time may be needed). MVP2 queries is much the same (though hopefully faster)
3. Correct-ish. You'll still be able to see KG2 sources in the legacy ARAX UI. If some ARS endpoint somewhere points to an old system, you might see it, etc. But after all the retraining and stuff is done, any query that goes to the UI which hits a Shepherd-backed ARS endpoint will only surface phase 3 ingests

[saramsey]

@dkoslicki thank you for responding about xDTD.

Regarding RTX-KG2, the team at today's ARAX meeting decided two things that may be relevant to your post:

1. We (the OSU part of the ARAX team) are indefinitely pausing work on RTX-KG2.10.3 and no longer plan to roll it out to the legacy ARAX ARA. That decision, which has been some time in coming, is based on our team's collective assessment that the KG2.10.3 graph, despite being KGX-compliant, would not be an acceptable candidate for integration into the Translator system via the legacy ARAX (this is based in part on feedback from you and folks seeing creative-mode results in the UI that include in their support graphs knowledge from RTX-KG2 from Phase 2 ingests).
2. We (the ARAX team) are exploring integrating the DogPark Tier 0 graph into legacy ARAX for the purpose of xDTD, xCRG, and PathFinder.

@mbrush please let me know if you have any questions about this.
@dkoslicki please correct me if I have misstated anything.