nestauk / ai_genomics Goto Github PK

Per our kick off call on 13-06, Ada folks to send over literature reviews for our review.

We currently have access to Conor Griffin's Biotechnology & AI review paper

based off of ai genomics patent ids, develop methodology to sample abstracts, subject to google bigquery limits.

Assess options to enrich publication data with information about influence / impact / controversy

• Citation data already addressed in #4
• Assess options to enrich publication data with information about social media debates from Crossref Event data:
  • What information is available directly from CRED API?
• Assess how controversial a paper is using scite.ai
  • This will involve some costs. Estimate costs under different scenarios

JMG will be working on this w/c 28 March

• OpenAlex EDA PR

Assess options to enrich OpenAlex data with institutional / geographical information from GRID

• This will involve fuzzy matching OpenAlex institutions with e.g. GRID institutions

Create template

The format of the /inputs/gtr/gtr_projects-projects.json file on S3 has changed, meaning gtr_definintions.py does not run

as @Jack-Vines and I will be on OJO.

Structure:

1. Context

• Project goals
• Research questions

3. Methodology

• High level methodological narrative
• Data sources: collection and enrichment
• Analysis

4. Outputs
5. Timelines and budget

• Assess CrunchBase dataset:
  • What are the right crunchbase tags
  • How long are the descriptions of CB companies
  • ...

can we do one for like 5 patents? filter patent ids in genomics - precondition for issue #25

• close issues that have been completed
• delete remote branches that have been merged

Using ai genomics patent ids, query BigQuery for additional fields that will be relevant for analysis.

• interrogate patent data schema to identify fields most relevant for analysis
• interrogate how complete those fields are on a sample of ai genomics patent ids

Based on feedback from the first epic, we need to clarify a collaborative report writing process. We will also need to schedule an onboarding call to make sure everyone knows how we will be collaborating.

• Establish report writing process
• Set up call with team to onboard us
• implement process

• Get lit reviews
• Get timelines for other project streams
• Get lists of terms they are using to define AI / Genomics
• Get research questions / assumptions
• Schedule dates for epic meetings
• Create project slack and invite them (do we want to do this? :-))
• Agree if there is an advisory board / its composition / how we engage with it
• Decide comm strategy: where does the report live and how do we share it with others

Sanity check patent ids that should be ai and genomics.

• pick N .csvs and check patent ids are relevant to ai AND genomics (perhaps pulling a sample of titles to review or even searching patent ids in Google's front end: https://patents.google.com/ for more complete info)
• Investigate distribution of cpc/ipc codes in "golden-shine-355915.genomics.*" - are all codes represented to at least some extent?

• Collect data relevant across research data, patent data, business data and funding data
• Conduct EDA and DQA across the relevant datasets
• Improve collective domain knowledge and subject expertise of AI and genomics
• Solidify agile ways of working
• Define whether this data collection work should sit within this project repo or beyond it

in patent, openalex, crunchbase taxonomies

As a baseline, we could use the same metrics karlis used in innovation sweet spots - we could also use his code from the repo to calculate i.e.:

• #of orgs founded per topic (pending topic definition: i.e. crunchbase category, cluster etc.)
• #of funding rounds per topic per topic (pending topic definition: i.e. crunchbase category, cluster etc.)
• Overall funding per topic

We could also join on patent assignees and look at the number of patents per topic.

After that, we could also calculate more complex indicators around emergence metrics, look into novelpy

• Update getters to fetch data from S3 without storing locally
• Generate summary by year including
  • Number of papers
  • Top institutions
  • Top countries
  • Top concepts
• Update getters to fetch data and query for specific concepts, store locally (necessary for genomics)
• Explore impact of different definitions on corpus
  • E.g. number of papers in intersection of AI and genomics
  • Coverage of papers in literature review

Assess OpenAlex. Some questions to consider:

• What is its coverage?
• What is its timeliness?
• How will we collect it?
• How will we store it?
• How will we find AI / genomics papers in it?

start simple:

• simplest approach: just use crunchbase topics or openalex topics

Think about approaches:

• what about using pygram? n-grams in abstracts and look at trends -- what's emerging? (second simplest)
• spectar (sp?) embeddings of documents (just pull vectors based on DOI from semantic scholar) and just cluster those guys (stretch-ish goal)
• topic modelling
• some "minimum" threshold of topics - can we at least find these types of categories from the literature review?

crunchbase --> do we do this in a qualitative way?

Explore options to enrich OpenAlex with semantic scholar data including:

• Citations
• High impact citations
• Fields of study
  As part of this, request access to free + fast API (it takes some time for them to confirm access)

JMG will be working on this w/c 4 April

from Luca: yesterday I forgot that Sam is on leave until Friday so I guess we could discuss wikipedia topics for this project next week (15-08)

Options include:

• Gateway to Research
• NIH
• EU (through Cordis)
• Wellcome Trust

As part of the project deliverable, we are also interested in joining datasets to analyse. We would then need to:

• Identify common entities across all the datasets we've collected so far i.e. a company name
• We will need to identify a strategy to join datasets based on an entities i.e. use jacchammer? Are there other entity matching algorithms we want to explore? What preprocessing will we need to do to join?
• Justify why we need joins - is this to help with analysis? Which analysis? Is this to expand data collection?
• Develop an evaluation strategy for joining i.e. develop a ground truth dataset of the same entity across different datasets - report on how many entities do we in/correctly join etc.? How do these metrics change subject to different entity matching algorithms and preprocessing methods?

[from @Juan-Mateos] Potential joins:

• orgs from patent data w/ crunchbase
• expanding list of crunchbase companies based on patent join

Other ideas:

• join named patent inventors w/ OpenAlex's publisher or display_name

• Check pipeline to enrich abstract data with MeSH terms: what is required to run this, how long does it take?

Some options:

• Patent lens
• Global Patent Index
• USPTO using the open AI patent dataset (https://www.uspto.gov/ip-policy/economic-research/research-datasets/artificial-intelligence-patent-dataset)
• Google Patent dataset @BIG query
• Check with Karlis / Discovery Hub about their plans to analyse patent data

• Find AI + genomics papers in Papers with Code
• Get their GitHub repos
• Are there any benchmark data related to AI + genomics?

second convo - spotlight tool on thursday morning