Skip to content

The Complement-ARIE Landscape Analysis as a curation source

Written 2026-10-30 against dismech main at commit ae8c3458fe3. Sources: the NIH Common Fund Complement-ARIE Landscape Analysis (report dated 29 February 2024) and its Tableau Public dashboard, whose "Explore Questions" view offers the per-paper answers as a data download.

The Landscape Analysis surveys the published literature on non-animal approaches, putting the same 19 questions to each of about 3,000 papers: what method was described, which tissues, cell types and diseases it examined, how it was validated, and what limits it. That makes it a ready-made map of the field for the Monarch NAMs work to build from. Models of disease are a large part of what DisMech records, through each entry's experimental_models, animal_models and computational_models and the links that tie a model to the mechanism it reproduces, so a structured survey of how disease is modelled without animals bears directly on what DisMech can hold.

This report describes the data download, how it was turned into tables, and which of the 19 questions could feed DisMech curation. It sits beside the Complement-ARIE Challenge winners report, which covers a different Complement-ARIE source (20 prize-winning proposals), and the NAMs search vocabulary. The Landscape Analysis PDF itself has not been read for this report; everything below comes from the dashboard download.

Summary

  • The download gives answers to 19 questions for 2,737 papers. It lists 3,051 citation keys; 294 have no answers and 20 group several papers under one key, which leaves 2,737 papers with answers that trace to one paper.
  • Nearly every one of those papers has a citable identifier. 2,427 have a PMID and 2,725 a DOI; 2 have neither. Most are 2023 publications (2,447 of 2,737).
  • The answers are summaries that point to papers. They read as model-generated ("Yes, the paper examines melanoma. The abstract clearly states …"); the download does not say how they were produced. They tell a curator which paper to read. A DisMech snippet is still quoted from the paper itself.
  • Abstracts should be fetched by PMID or DOI. The download's Abstract column does not always line up with its row, while the answers follow the title and PMID, so the identifiers are the reliable join.
  • The questions closest to DisMech's data model are diseases, tissues, cell types, cellular endpoints, and the validation and limitations of each method. Each is answered for between about half and four-fifths of the papers.

The 19 questions

Counts are papers with an answer, out of the 2,737 papers. Every question also has 294 NA answers, all from the same 294 papers without answers, so those are not shown.

Question Full wording Papers with an answer
Method Details Described Provide details on the method described 2,682
Method Technical Challenges Described What were the main technical challenges described in the paper that this method was trying to overcome? 2,477
Method: New Technology or Refinement of Existing Approach Is the method novel (new technology) or is it an improvement on an existing approach (refinement)? 2,416
Tissue Types Examined What tissue types does this paper examine? 2,152
Remaining Challenges after Method Development What challenges remain even after the development of this method? 2,130
Cell Types Examined What cell types does this paper examine? 1,961
Human Physiology Elements Considered What elements of human physiology does this paper consider? 1,847
Disease(s) Examined Is a disease examined in this paper? If so what disease? 1,572
Method Limitations What are the limitations of the method as described in this paper? 1,498
Cellular Endpoints Measured Are cellular endpoints being measured in this paper? If so, what are the endpoints? 1,472
Method Validation Methods Described Did the authors describe the validation methods, such as the use of controls, taken for this method? If so, what are they? 1,391
Research Funding Details What are the details on the funding of this research, including any grants or sponsors? 1,303
Population Modeled Did this paper model outcomes in a specific population? If so, what population? 553
Model Metabolic Capacity Addressed Did this model seek to address metabolic capacity of the tissue? If so, in what way? 373
Economic Considerations Discussed Did the authors discuss economic considerations? If so, what did they say? 174
Ethical Considerations Discussed Did the authors discuss ethical considerations? If so, what did they say? 119
Method Regulatory Use Discussed Did the authors discuss the method in terms of a regulatory use? If so, what did the authors say? 96
Population Diversity and Variability Incorporated Did the authors incorporate considerations of population diversity or interindividual variability into the technology? If so, how? 48
Workforce Considerations Discussed Did the authors discuss workforce considerations? If so, what did they say? 11

Every row in the download has Answer-simple set to yes, which suggests the dashboard view lists answered questions only. A paper missing from a question's count may therefore not have been answered "no".

The corpus

Papers
Citation keys in the download 3,051
Keys grouping several papers (DOI and abstract read *) 20
Keys with NA for all 19 questions 294
Papers with answers 2,737
of which keyed by PMID 2,427
of which keyed by Web of Science accession 199
of which keyed by Scopus accession 111
of which have a DOI 2,725

Each paper answers between 1 and 17 of the 19 questions; most answer 8 to 12. Publication years run from 2004 to 2024, with 2,447 papers from 2023.

The 294 keys without answers and the 20 keys grouping several papers are separate sets: no key is in both, which is why 3,051 − 294 − 20 = 2,737. The 294 carry no title or accession, but 270 of them carry a DOI, so they can still be identified.

Edge cases handled by the parser

A dashboard download is built for viewing rather than for reuse as data, so a few of its features need handling before it can be read as tables. The parser handles each one deterministically, and they are listed here so that a reader of the tables knows what was changed.

Edge case Extent Handling
Abstract does not always line up with its row 1,020 keys share their abstract text with another key; for example, the row for Metabolic reprogramming of clear cell renal cell carcinoma carries an abstract about perovskite nanocrystals Not carried forward. abstract_shared=1 marks the keys where it can be detected. Fetch abstracts by PMID or DOI
Several papers under one key 20 Author et al. YEAR (NA) keys, 380 answers Listed with merged_key=1; answers left out, since they cannot be traced to one paper
Keys without answers 294 keys no_content=1
Title holds the PDF's first-page text (journal citation, DOI, "REVIEW ARTICLE", "Check for updates") 28 papers Moved to title_page_header; title left empty, to be taken from the PMID or DOI
Title begins with a journal logo 1 paper (ESC Personalizing therapy for atrial fibrillation, Barichello 2018) Removed by a per-paper rule, since a general rule would also strip titles that begin with an acronym
UTF-8 text read as Windows-1252, up to three times over (naïve for naïve) about 9,000 fields Converted back
NA written for a missing title, DOI or answer throughout Emitted as empty, or answer_is_na

Regenerating the tables

The download is about 128 MB, over GitHub's per-file limit, so neither it nor the tables derived from it are committed. data/complement-arie/MANIFEST.yaml records the source URL and the download's sha256. To rebuild:

  1. Download "Q&A-Question Details" from the dashboard's "Explore Questions" view into data/complement-arie/.
  2. Run just complement-arie-parse.

The script checks the download's sha256 against the manifest first and stops if they differ, since a different snapshot would not give the counts in this report. --no-verify parses a different snapshot anyway. It then writes three files beside the download, all gitignored:

  • papers.tsv: one row per citation key, with authors, year, accession, PMID, DOI, title, and the flags above.
  • answers.jsonl: one record per paper and question, with the answer text.
  • questions.tsv: the table in The 19 questions.

What this could offer DisMech

This resource meets a need already recorded in two open issues:

  • The NAMs inventory (#13602) anticipates "an inventory of NAMs from many sources" that will need sorting into models that fit a DisMech entry, fringe cases, and models that do not. The Landscape Analysis is such a source, already assembled: 2,737 papers, each with its method, disease, tissues and cell types described in the same form.
  • The NAMs search-term analysis (#14428) was opened in support of a NAMs landscape analysis. The Complement-ARIE corpus is a landscape analysis that has already been carried out, so it can serve as the literature base that the search vocabulary in the NAMs search vocabulary report is checked against.

Within that, the following are candidate uses, not conclusions; none has been tried yet.

  • Disease-to-model leads. The 1,572 "Disease(s) Examined" answers name a disease alongside a method described under "Method Details Described". Read together, they point at papers that may describe an experimental_models record for an existing entry, which is the question the winners report asked of 20 proposals, here across 1,572 papers. The disease names are free text and need matching to MONDO and to DisMech entries first.
  • Cell and tissue annotation. "Cell Types Examined" and "Tissue Types Examined" are the free-text counterparts of the CL and UBERON bindings an experimental model carries.
  • Model limitations. "Method Limitations" and "Remaining Challenges after Method Development" describe the translational caveats that a modeled_mechanisms link records in limitations and divergences.
  • Validation. "Method Validation Methods Described" is one input to the fidelity tier a model link is given: how faithfully the model captures the human mechanism.

In every case the answer points to a paper. Each fact that reaches a DisMech entry needs the paper fetched with just fetch-reference and an exact quote from it.

Open questions

  • Whether to work through the full corpus, or start with the papers that have a disease answer.
  • How to match free-text disease names to MONDO and to DisMech entries at this scale, and how to decide which matches are close enough to curate.