Versions Compared

Version Old Version 31 New Version 70
Changes made by

Vladimir Makarov

Vladimir Makarov

Saved on

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

...

  • Define the most common research questions in target discovery and validation. Establish an agreement between the project team that these are indeed the core target discovery business questions, and rank order them by vote by perceived relative importance. If such questions are many, pick the top ones. Establish an agreement on how many exactly. One can use this paper as a starting point for listing of relevant competency questions: https://www.sciencedirect.com/science/article/pii/S1359644613001542 (Failure to identify business questions , or picking too many or too few is was a project risk)

  • Open Targets can serve as a publicly available standardized data source for this use case. Validate that Open Targets either has a ready to use Knowledge Graph implementation, or can be converted into a KG with reasonable cost (this is known project riskwas a project risk; we established data availability for the Open Targets as a KG in BioCypher, and this risk is now closed)

  • Select a Large Language Model engine from publicly accessible sources. (Failure to identify a suitable open LLM is was a project risk. The available LLMs were analyzed and this risk is now closed)

Prompt-tuning procedure:

  • Define a plain text data source for mining; one of the choices can be the entire set of paper abstracts indexed in PubMed plus perhaps the entire collection of open-source papers. (Failure to download this large volume of data can be a risk)

  • Ask a plain English question from the collection of business questions identified above

  • An LLM uses the question and the data source to produce a human-readable answer

  • Either same or some other LLM converts this answer to a Knowledge Graph. (KG generation from text is a source of risk)

  • This answer Knowledge Graph is compared to the KG of the original data source (such as Open Targets). This comparison must be local; in other words, irrelevant sections of the larger knowledge graph should not be considered. (Ability to find a ready KG comparison algorithm or to code it fresh is a risk).

  • An accuracy metric is computed + inputs and outputs saved in some DB

  • Prompt-tune the opening plain text question Retrieval Augmented Generation (RAG):

    • Ask plain English question using prompt-tuned version of one of the questions from the business questions collection

    • This question is converted into a structured query by an LLM (Failure to generate a proper query for a KG database system is a risk)

    • Execute this query over a structured controlled data source (e.g. Open Targets DB)

    • Convert raw output of the query into human-readable input (by an LLM or by other means)

    • An expert compares this answer or answers with the expected one(s)

  • An accuracy metric is computed + inputs and outputs saved in some DB

  • Prompt-tune the opening plain text question to maximize the output quality

  • Experiment with different modes of using an LLM (such as LLM agents, or query templates, or different representations of the data source schema) to maximize the output quality

  • Repeat this tuning cycle for all business questions in the collection

  • Use these optimized prompts in RAG, below

Retrieval Augmented Generation (RAG):

  • Ask plain English question using prompt-tuned version of one of the questions from the business questions collection

  • This question is converted into a structured query by an LLM (Failure to generate a proper query for a KG database system is a risk)

  • Execute this query over a structured controlled data source (e.g. Open Targets DB)

  • Convert raw output of the query into human-readable input by an LLM

  • Quality assurance: it is desirable to produce a few instances of questions with known answers, ideally, by human experts, and then use these question-answer pairs as sanity checks of the RAG pipeline

Historic (obsolete) versions of the above steps:

  • Historic vision for the query (this is obsolete and is only preserved here for scope traceability):

    • Define a plain text data source for mining; one of the choices can be the entire set of paper abstracts indexed in PubMed plus perhaps the entire collection of open-source papers. (Failure to download this large volume of data can be a risk)

  • Historic vision for the QA (likely not necessary):

    • Either same or some other LLM converts this answer to a Knowledge Graph. (KG generation from text is a source of risk)

    • This answer Knowledge Graph is compared to the KG of the original data source (such as Open Targets). This comparison must be local; in other words, irrelevant sections of the larger knowledge graph should not be considered. (Ability to find a ready KG comparison algorithm or to code it fresh is a risk).

Not in scope:

  • Proprietary data are not used to train any model and project participants are not asked to share proprietary data in any way. But proprietary data may be analyzed in the context of this set-up by the individual participants using private instances of the pipeline.

  • Training of a brand new LLM is out of scope. The plan is to only prompt-tune an existing LLM.

...

Phase

Milestones

Deliverables

Est Date

Initiation

Project charter

  1. A list of candidate pre-competitive projects

  2. One or more projects selected by vote

  3. Project charter is drafted for the winning idea

  4. Raise minimal funds for the Elaboration phase

12/11/23

(Complete)

Elaboration

  1. Development plan

  2. Cost estimates

  1. Risks analysis – see Risk Registry below

  2. Technology analysis to address the identified risks

  3. Work Breakdown Structure (WBS)

  4. Cost estimates

  5. Time estimates

  6. Gantt Chart for Construction with additional iterations as needed and a work schedule

  7. Make feasibility decisions before committing to build

Q1-Q3 2024

(Complete)

Construction

Target discovery pipeline

(POC)

  1. Learn best ways to use LLMs for NL queries

  2. Lessons learned published

  3. Target discovery pipeline – detailed deliverables are not yet known

  4. Lessons

  1. Results of experiments

  2. Lessons learned recorded and published

Q4 2024

(Almost complete in December 2024)

Construction (PROD)

  1. Target discovery pipeline

  1. Target discovery pipeline – detailed deliverables are to be specified by research project sponsors

TBD

Transition

Sustainability achieved

  1. Place the prototype into maintenance mode or outsource for continuous development by another organization (e.g. non-profit)

  2. Plan extension work, if any

TBD

Risk Registry

...

Description

...

Mitigation

Risks in green are resolved

Risks in yellow are in active research

Risks in white are general in nature

This may be unnecessary, TBDFailure to build a prototypical target discovery pipeline on in this phase (“project elaboration”)

Description

Mitigation

Responsible Party

Failure to identify business questions, or picking too many or too few

Draft appropriate business questions - DONE; but not all business questions can be answered with specific technologies, so must take this factor into account

Lee Harland, John Wise, Bruce Press, Peter Revill

Validate that Open Targets either has a ready to use Knowledge Graph implementation, or can be converted into a KG with reasonable cost

The Hyve
Jordan Ramsdell
Robert Gill
Brian Evarts

Open Targets/EBI:

  • Sebastian Lobentanzer

  • Ellen McDonagh

Failure to identify a suitable LLM

  • See this comparison

  • Recommend to focus on the Cypher query generation ability as the key risk (below)

  • Start with one open-source and one closed-source LLMs (say Mistral and GPT 4) and agree to explore others later, and meanwhile close this risk

Jon Stevens, Etzard Stolte, Helena Deus; Brian Evarts; Wouter Franke, Matthijs van der Zee

Failure to generate a proper query for a KG database system by an LLM

Technology research.

  • See refs 7, 8, 13, 14 below

  • BioCypher by EBI may have this capability already - needs evaluation

The Hyve
Jordan Ramsdell
Robert Gill
Brian Evarts

Open Targets/EBI:

Sebastian Lobentanzer

Ellen McDonagh

Does Open Targets use an ontology?

Yes in general

The Hyve

Failure to download a large volume of data (all of the PubMed as a maximum) for the prompt-tuning of the LLM

CLOSED RISK

This is not necessary

Failure to perform local KG comparison with calculation of a score

  1. Technology research

  2. If no ready-to-use technology exists, estimate bespoke development

  3. If estimates indicate infeasibility, this may become a gap

CLOSED RISK

This is not necessary - we can compare the output manually

Failure to build a prototypical target discovery pipeline on the limited budget in case of mounting technical difficulties

CLOSED: Schedule the project in phases. Aim to answer known unknowns and to establish risk mitigation strategies early

.

It is not yet known whether a product will be built. For now the scope is focused on the technology analysis for the POC

Some proprietary LLMs may be censored, thus introducing uncontrollable bias in the answers that they produce

DONE:

CLOSED. Censorship may already be included in the performance scores, so this is taken care of in the comparison of the LLMs. However, there is team preference for open-source and uncensored LLMs

Identified and resolved in the LLM sub-team

...

Stakeholder mailing list in Google Groups: https://groups.google.com/a/pistoiaalliance.org/g/llm-projectMS Teams: Large Language Models | General | Microsoft Teams

Meetings

Every other week at 8 am PST (= 11 am EST = 4 pm London = 5 pm Berlin) starting on January 10th, 2024

Lessons Learned

  • The highest risk item is generation of the structured query (Cyphrer or SPARQL) from a plain English request. Some publications estimate success rate of about 48% on the first attempt.

  • The structure of the database used for queries matters. LLMs can easier produce meaningful structured queries for databases with flat, simple structure.

  • Practically useful system requires filtering or secondary mining of output in addition to natural language narration.

References

...

https://www.sciencedirect.com/science/article/pii/S1359644613001542

...

https://www.nature.com/articles/s41573-020-0087-3

...

Github

https://github.com/PistoiaAlliance/LLM

Final Report

Lessons Learned

  • Modern LLMs have sufficient knowledge of biology embedded in them to be able to answer almost any question we (humans) can think about. This is a source of problems: hallucinations are indistinguishable from true answers; we cannot fully test the innate ability of the LLMs to translate the natural language questions into structured queries (unless we obscure the terms with synonyms unknown to the LLM).

  • The highest risk item is generation of the structured query (Cyphrer or SPARQL) from a plain English request. Some publications estimate success rate of about 48% on the first attempt.

  • The structure of the database used for queries matters. LLMs can easier produce meaningful structured queries for databases with flat, simple structure.

  • The form of the prompt matters. LLMs can easier produce meaningful answers from prompts that resemble a story, rather than a dry question, even if the details of the story are irrelevant to the main question asked.

  • Practically useful system requires filtering or secondary mining of output in addition to natural language narration.

  • It is extremely important to implement a reliable named entity recognition system. The same acronym can refer to completely different entities, which can be differentiated either from the context (hard) or by asking clarifying questions. Must also map synonyms. Without these measures naïve queries in a RAG environment will fail.

  • LLMs may produce different structured queries starting from the same natural language question. These queries may be semantically and structurally correct, but may include assumptions on the limit of the number of items to return, or order, or lack of these. These variations are not deterministic. As a result on different execution rounds the same natural language may result in different answers. It is necessary to explicitly formulate the limits, order restrictions, and other parameters when asking the question, or to determine the user’s intentions in a conversation with a chain of thought. A question related to this topic, is whether specifics in the implementation of usual RAG models with a vector database may introduce implicit restrictions on what data is explored by the LLM and what data is not, and thus artificially limit the answers. This may be happening without the user knowing the restrictions (and perhaps even without the system’s authors knowing that they introduced such restrictions embedded in the specifics of the system architecture).

  • Need for an API standard.

  • There is no good biological test-set for LLM evaluation
    Existing test sets are saturated

    Background knowledge contaminates the results
    Frontier models have ~100% biological background knowledge, which makes evaluation of cypher query generation difficult as these models can bypass cypher queries generation and hallucinate correct results

    Models need an independent way for entity resolution as KG may not have specific synonyms provided by user

    Small models (mistral, llama 13B etc) underperform; even their adapters / fine tuned versions trained on cypher generation fail on OT KG

    Providing automatically-generated graph schema doesn't really help for OT KG

    Graph Schema matters
    OT KG has non-trivial schema that makes LLMs very confused

  • Template-based strategies achieve 100% performance in current evaluation, but require hard-coded templates for information retrieval
    Agent-based strategies achieve 83-98% performance, and do not require prior knowledge to work with KGs
    DSPy framework could be used to optimize prompts and increase success rate

    Future directions:
    Development of a better test-set that could be used to improve LLM ability to generate cypher queries for biological knowledge
    Further improvement of performance of agent-based approach by using long-term history (generation of templates)
    Introducing variation into KG schema as a parameter for testing. Using KGs with strict schema and KGs that are automatically extracted from literature to test how reliable strategies are
    Introducing entity resolution to improve information retrieval

References

  1. https://www.sciencedirect.com/science/article/pii/S1359644613001542

  2. https://www.nature.com/articles/s41573-020-0087-3

  3. https://www.epam.com/about/newsroom/press-releases/2023/epam-launches-dial-a-unified-generative-ai-orchestration-platform

  4. https://epam-rail.com/open-source

  5. Open LLM Leaderboard: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard

  6. Chatbot Arena: https://chat.lmsys.org/?arena

  7. Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning

    https://arxiv.org/abs/2310.01061

  8. Knowledge-Consistent Dialogue Generation with Language Models and Knowledge Graphs

    https://openreview.net/forum?id=WhWlYzUTJfP&source=post_page-----97a4cf96eb69--------------------------------

  9. BioChatter Benchmark Results: https://biochatter.org/benchmark-results/#biochatter-query-generation

  10. MBET Benchmark (embeddings) https://huggingface.co/spaces/mteb/leaderboard

  11. Lora---------BioChatter Benchmark Results: Land and Lorax: https://predibase.com/lora-land

  12. A Benchmark to Understand the Role of Knowledge Graphs on Large Language Model's Accuracy for Question Answering on Enterprise SQL Databases. Summary: queries over a KG with GPT 4 are much more accurate than queries over a SQL database with GPT 4. https://biochatterarxiv.org/benchmark-results/#biochatter-query-generationMBET Benchmark (embeddings) /abs/2311.07509

  13. https://towardsdatascience.com/evaluating-llms-in-cypher-statement-generation-c570884089b3

  14. https://huggingfacemedium.cocom/spaces/mteb/leaderboard

  15. Lora-Land and Lorax: https://predibase.com/lora-land

  16. A Benchmark to Understand the Role of Knowledge Graphs on Large Language Model's Accuracy for Question Answering on Enterprise SQL Databases. Summary: queries over a KG with GPT 4 are much more accurate than queries over a SQL database with GPT 4. https://arxiv.org/abs/2311.07509 neo4j/enhancing-the-accuracy-of-rag-applications-with-knowledge-graphs-ad5e2ffab663

  17. linkedlifedata.com

  18. Kazu - Biomedical NLP Framework: https://github.com/AstraZeneca/KAZU

  19. https://towardsdatasciencegithub.com/evaluatingf/awesome-llms-in-cypher-statement-generation-c570884089b3chatgpt-prompts/tree/main

  20. Zhou, L., Schellaert, W., Martínez-Plumed, F. et al. Larger and more instructable language models become less reliable. Nature 634, 61–68 (2024). https://mediumdoi.comorg/neo4j10.1038/enhancings41586-the024-accuracy-of-rag-applications-with-knowledge-graphs-ad5e2ffab663

  21. linkedlifedata.com

  22. Kazu - Biomedical NLP Framework: https://github.com/AstraZeneca/KAZU

  23. https://github.com/f/awesome-chatgpt-prompts/tree/main 07930-y

  24. Karthik Soman, Peter W Rose, John H Morris, Rabia E Akbas, Brett Smith, Braian Peetoom, Catalina Villouta-Reyes, Gabriel Cerono, Yongmei Shi, Angela Rizk-Jackson, Sharat Israni, Charlotte A Nelson, Sui Huang, Sergio E Baranzini, Biomedical knowledge graph-optimized prompt generation for large language models, Bioinformatics, Volume 40, Issue 9, September 2024, btae560, https://doi.org/10.1093/bioinformatics/btae560

  25. https://www.promptingguide.ai/

  26. References on Named Entity Recognition in biological sciences: Pubmed

  27. Incremental Knowledge Graphs Constructor Using Large Language Models