dmccloskey
commented
1 month ago Task categories and types for AI agents for life sciences
- General software requirements: [Streamlit](), [Ollama]()
- General database requirements: vector database (our internal stack is preferential towards FAISS, but ChromeDB, Milvus, can be used just as well).
AI agents for computational modelling and simulation
- Special software requirements: https://github.com/copasi/basico
- Data for computational modelling: ODE mathematical models in SBML format from BioModels e.g., insulin dynamics
- Data for model annotation: cell ontology, ChEBI, and UniProtKB
- Other data: PubMed for original articles
Task type 1
- Description: Simulation of a mathematical model and reporting of the biomarker trajectories and predicted clinical efficacy
- Input: simulation parameters such as initial concentrations
- Output: time-course of simulation species
Task type 2
- Description: Creating a mathematical model from scratch
- Input: Original article describing the mathematical model and list of equations
- Output: SBML model with annotated species
AI agents for omics and foundation models
- Special software requirements: scGPT
- Data for analysis: cell by gene
- Other tools/analyses: differential gene set enrichment analysis using GO, UMAP
Task type 1
- Description: Integration of multiple scRNA seq datasets, correction for batch effects, annotation of cells, and reporting of the results as a UMAP
- Input: multiple cellxgene datasets for a particular disease (e.g., Rheumatoid Arthritis)
- Output: UMAP visualization with cell annotation
Task type 2
- Description: Simulation of gene perturbation and reporting of the predicted differentially expressed genes using pathway enrichment analysis
- Input: cell x gene dataset for a particular disease, knockout gene list
- Output: list of differentially expressed genes and pathway enrichment analysis visualization
AI agent for Biomedical knowledge graph reasoning and construction
- Special software requirements: LLMGraphTransformer and ULTRA
- Biomedical knowledge graph dataset: PrimeKG specifically the subset used in STARK
- Other data: PubMed for original articles
- Graph database: NetworkX
Task type 1
- Description: Knowledge graph Q&A and retrieval of K-hop subgraph explanations
- Input: Natural language question (see subset used in https://arxiv.org/abs/2404.13207 for PrimeKG)
- Output: Ranked nodes answers and visualization of k-hop subgraphs
Task type 2
- Description: Disease knowledge graph construction from text using a LLM to graph model and link prediction model to fill in gaps
- Input: List of disease MeSH terms and associated articles from PubMed and list of nodes and edges (same as in PrimeKG)
- Output: NetworkX representation of the knowledge graph and visualization
Task type 3
- Description: Same as type 1 but including protein embeddings from https://www.uniprot.org/help/embeddings and additional vector similarity search of drug targets embeddings
- Input: Natural language question (see subset used in https://arxiv.org/abs/2404.13207 for PrimeKG)
- Output: Ranked nodes answers and visualization of k-hop subgraphs
Other (backup)
- Retrieval of relevant literation involved in the target drug from PubMed (e.g., using Semantic Scholar recommendation API) and reporting of the results as a table with paper citations
- Retrieval of relevant clinical trials for a disease from ClinicalTrials.gov, parsing of the results into molecule, trial criteria, and trial outcomes, and reporting of the results as a table with clinical trial citations similar to https://www.cell.com/patterns/fulltext/S2666-3899(22)00018-6
Description
We would like to define the tasks for the Hackathon. The tasks should specify the inputs, expected outputs, foundation models, and external data resources
Checklist
Acceptance criteria
RACI