Annotating Gene Sets with Biological Processes Using an LLM

This notebook shows how to use AnnDictionary’s functions to have LLMs label a set of genes based on the biological process that they may represent. To demonstrate this functionality, we’ll re-annotate some known gene sets from GO Biological Process (GOBP)

Skip the tutorial

In case you want to skip the tutorial, here’s all the code you need to run this type of annotation.

To annotate a list of genes:

import anndict as adt
gene_list = ['gene1', 'gene2', 'gene3']
annotation = adt.ai_biological_process(gene_list=gene_list)

To annotate the results of rank_genes_groups in an AnnData object:

import anndict as adt

#The results will be returned and also stored in `adata.obs['ai_biological_process']`.
annotation_df = adt.ai_annotate_biological_process(adata, groupby='disease vs. control', n_top_genes=10, new_label_column='ai_biological_process')

Begin the tutorial

import anndict as adt
import gseapy as gp

# Configure LLM backend
adt.configure_llm_backend(provider='anthropic',
                          model='claude-3-5-sonnet-20240620',
                          api_key='my-anthropic-api-key',
                          requests_per_minute=100
                          )

We’ll use gseapy to access the GOBP database, and re-annotate some gene lists of known process.

# Download the latest Human GOBP gene set collection (2023 release on Enrichr)
gobp = gp.get_library(name="GO_Biological_Process_2023", organism="Human")  # dict: {term: [genes]}

# Inspect how many gene sets were retrieved
print(f"Total GOBP terms: {len(gobp):,}")

Total GOBP terms: 5,406

# Grab any three terms
terms_of_interest = list(gobp.keys())[:3]

# Build a dictionary of the selected gene sets
selected_gene_sets = {term: gobp[term] for term in terms_of_interest}

# Display the gene symbols for each selected term (truncate for readability)
for term, genes in selected_gene_sets.items():
    print(f"\n{term} ({len(genes)} genes)")
    print(", ".join(genes[:15]), "...")       # first 15 genes as a preview

'De Novo' AMP Biosynthetic Process (GO:0044208) (6 genes)
ATIC, PAICS, PFAS, ADSS1, ADSS2, GART ...

'De Novo' Post-Translational Protein Folding (GO:0051084) (32 genes)
SDF2L1, HSPA9, CCT2, HSPA6, ST13, ENTPD5, HSPA1L, HSPA5, PTGES3, HSPA8, HSPA7, DNAJB13, HSPA2, DNAJB14, HSPE1 ...

2-Oxoglutarate Metabolic Process (GO:0006103) (14 genes)
IDH1, PHYH, GOT2, MRPS36, GOT1, IDH2, ADHFE1, GPT2, TAT, DLST, OGDHL, L2HGDH, D2HGDH, OGDH ...

# Now, Let's annotate these known gene sets with an LLM
llm_annotation = {}
for term, genes in selected_gene_sets.items():
    llm_annotation[term] = adt.ai_biological_process(gene_list=genes)

/Users/geocr/repos-local/anndict/anndict/llm/base_llm_initializer.py:49: LangChainBetaWarning: Introduced in 0.2.24. API subject to change.
  return InMemoryRateLimiter(
/Users/geocr/repos-local/anndict/anndict/llm/llm_manager.py:309: LangChainDeprecationWarning: The method `BaseChatModel.__call__` was deprecated in langchain-core 0.1.7 and will be removed in 1.0. Use invoke instead.
  response = llm(langchain_messages, **kwargs)

Now, we can view the results. Here, the keys are the label of the pathway in GOBP, and the values are the LLM-derived label of the same pathway. These example results have remarkable agreement.

# And View the results
llm_annotation

{"'De Novo' AMP Biosynthetic Process (GO:0044208)": 'Purine biosynthesis pathway',
 "'De Novo' Post-Translational Protein Folding (GO:0051084)": 'Protein folding and chaperone-mediated quality control.',
 '2-Oxoglutarate Metabolic Process (GO:0006103)': 'Mitochondrial tricarboxylic acid (TCA) cycle and related metabolic pathways.'}