Gene List Prioritization

Score thousands of candidate genes against the literature in minutes.

When to use

You have a ranked list of candidate genes — from a screen, GWAS, DEG analysis, or co-expression module — and need to know which ones have existing literature support for a specific biological role.

Example: scoring differentiation drivers

The following pipeline scored ~1,500 candidate genes per pancreatic cell type for their potential as overexpression targets in directed differentiation. Genes come from an in-silico CRISPR KO screen; KO correlation measures how much each gene's loss disrupts the cell type's transcriptional identity.

Each gene receives a Differentiation Perturbation Score (DPS) from 1–10:

Score	Meaning
9-10	Direct published evidence of overexpression driving differentiation
7-8	Required for identity (KO causes loss/transdifferentiation), known co-regulator
5-6	Cell-type-specific expression, functional pathway association
3-4	General tissue expression, indirect association
1-2	No relevant evidence found

Prompt template

python

PROMPT = """You are a biomedical research agent scoring candidate genes as overexpression
targets for driving iPSC differentiation toward human pancreatic **beta cell** identity.

**Genes to analyze (10):** {item}

For each gene, search GLKB and PubMed for:
1. Direct role in beta cell differentiation, specification, or identity
2. Membership in the core beta TF network: PDX1, NKX6.1, NKX2.2, NEUROD1, NGN3, MAFA
3. Gene-disease associations: T1D, T2D, MODY, congenital hyperinsulinism
4. Role in insulin secretion or beta cell maturation
5. Expression specificity in human islet beta cells

Benchmark: PDX1=10, NKX6.1=9

**Output format - one block per gene:**
### [GENE_SYMBOL] - DPS: [X]/10
- **Evidence:** [1-2 sentences of strongest finding]
- **PMIDs:** [up to 3 PMIDs, or "none found"]
- **Rationale:** [what drove the score up or down]"""

10 genes are packed per {item} to amortize the GLKB + PubMed overhead across a mini-batch.

Code

python

import csv, json, re, requests

BASE_URL = "https://api.glkb.dev"  # <!-- DEV: replace with production base URL -->

def pack_batches(genes, batch_size=10):
	return [", ".join(genes[i:i + batch_size])
			for i in range(0, len(genes), batch_size)]

def parse_dps_response(text, gene_to_corr):
	records = []
	for block in re.split(r"\n(?=###\s)", text.strip()):
		m = re.match(r"###\s+(\S+)\s+[—\-]+\s+DPS:\s*(\d+(?:\.\d+)?)/10", block)
		if not m:
			continue
		gene, score = m.group(1), float(m.group(2))
		ev = re.search(r"\*\*Evidence:\*\*\s*(.+?)(?=\n-|\Z)", block, re.S)
		pm = re.search(r"\*\*PMIDs:\*\*\s*(.+?)(?=\n-|\Z)", block, re.S)
		ra = re.search(r"\*\*Rationale:\*\*\s*(.+?)(?=\n-|\Z)", block, re.S)
		records.append({
			"gene":          gene,
			"dps_score":     score,
			"ko_correlation": gene_to_corr.get(gene, ""),
			"evidence":      ev.group(1).strip() if ev else "",
			"pmids":         pm.group(1).strip() if pm else "",
			"rationale":     ra.group(1).strip() if ra else "",
		})
	return records

genes = ["PDX1", "NKX6-1", "RYR2", ...]  # up to 1,500 genes
gene_to_corr = {}                          # optional: map gene -> KO correlation score
batches = pack_batches(genes)

with open("scores.csv", "w", newline="") as f:
	writer = csv.DictWriter(
		f, fieldnames=["gene", "dps_score", "ko_correlation",
					   "evidence", "pmids", "rationale"]
	)
	writer.writeheader()

	with requests.post(
		f"{BASE_URL}/apps/glkb/users/me/batch",  # <!-- DEV: confirm URL -->
		headers={
			"Authorization": f"Bearer {os.environ['GLKB_API_KEY']}",
			"Content-Type": "application/json",
		},
		json={"items": batches, "prompt_template": PROMPT, "concurrency": 5},
		stream=True,
		timeout=7200,
	) as r:
		r.raise_for_status()
		for line in r.iter_lines():
			if line and line.startswith(b"data:"):
				event = json.loads(line[5:])
				if event.get("status") == "done":
					writer.writerows(
						parse_dps_response(event["response"], gene_to_corr)
					)
					f.flush()

Sample output

gene,   dps_score, ko_correlation, evidence,                                       pmids,    rationale
PDX1,   10.0,      0.521,          Overexpression drives beta cell differentiation, 34567890, Benchmark anchor.
RYR2,   6.0,       0.614,          RyR2 loss impairs GSIS in an insulinoma model,   28765432, Ca2+ handling in beta cells.
RSPH1,  1.0,       0.831,          Cilia structural biology - no pancreatic evidence, none found, No islet linkage.

At concurrency=5, 1,500 genes complete in approximately 30 minutes. The output CSV feeds directly into downstream ranking, GSEA, or differential filtering.