Target Druggability Screening
Compile a preliminary evidence dossier for a protein target. STRING associations
and BindingDB measurements alone do not establish druggability or therapeutic
relevance. This recipe retrieves a limited sample, not an exhaustive search.
Workflow
- Confirm the target's UniProt accession and species. Resolve a gene name
such as CDK4 through
host.science.search("uniprot", ...) first; review ambiguous
matches. BindingDB accepts UniProt accessions or PDB IDs, not arbitrary gene names.
Use the same confirmed accession for both searches below.
- Retrieve STRING interaction partners with an explicit species. The returned
functional associations do not measure physical binding, whole-network hub
centrality, or pathway enrichment.
- Query BindingDB for one endpoint (
Ki, Kd, IC50, or EC50) at a time.
Even within an endpoint, review assay conditions before comparing values.
- Use optional RDKit for molecular weight, calculated logP, HBD, HBA, and
rotatable bonds. The Rule of Five uses the first four descriptors, allowing
at most one violation. Install the existing
chemistry extra in the active
Python environment (uv sync --locked --extra chemistry) to enable this step.
No RDKit, missing SMILES, and invalid structures produce Unknown, never
invented descriptors or a passing verdict. Core imports remain stdlib-only.
- Preserve qualified affinities (
<, >, ~), missing values, citations, and
repeated assay records. These are evidence, not distinct confirmed leads.
Scores require an exact positive nM measurement and a complete descriptor
screen; otherwise they remain Unscored. The score is an unvalidated
within-endpoint heuristic, not a probability or a biological conclusion.
- Save each complete query envelope with its provenance, then save the dossier
with lineage to those exact artifact versions. Propagate connector failures;
do not replace an error with an empty successful search.
Interactive / Cell Recipe
Run inside a persistent Python kernel cell. Host connectors mediate all network
requests; the sidecar performs no direct networking.
import json
from pathlib import Path
from target_druggability_screening.kernel import format_dossier, rank_candidates
# Confirm both values before running: this example is human CDK4.
target = "P11802"
species = "9606"
affinity_type = "Ki"
ppi_res = host.science.search(
"string", target, limit=10, filters={"species": species}
)
binding_res = host.science.search(
"bindingdb", target, limit=20,
filters={"cutoff": 100, "affinity_type": affinity_type},
)
ranked_leads = rank_candidates(
binding_res["results"], top_k=10, affinity_type=affinity_type
)
source_versions = []
for result in (ppi_res, binding_res):
source_path = Path(f"{target.lower()}-{result['database']}-source.json")
source_path.write_text(json.dumps(result, indent=2, allow_nan=False), encoding="utf-8")
saved = host.save_artifact(
str(source_path), content_type="application/json", source=result["provenance"]
)
source_versions.append(saved["version_id"])
dossier_content = format_dossier(
target=target,
ppi_records=ppi_res["results"],
ranked_leads=ranked_leads,
provenance=[ppi_res["provenance"], binding_res["provenance"]],
)
dossier_path = Path(f"target-druggability-{target.lower()}.md")
dossier_path.write_text(dossier_content, encoding="utf-8")
artifact = host.save_artifact(
str(dossier_path), content_type="text/markdown", input_version_ids=source_versions
)
print(f"Dossier saved as artifact version: {artifact['version_id']}")
Sidecar Functions
calculate_smiles_descriptors(smiles): Parsed RDKit descriptors and engine
version, or a reason for missing/invalid/unavailable chemistry.
evaluate_lipinski(descriptors_or_smiles): Four-criterion screen; an incomplete
input has pass_rule_of_5=None.
score_lead(affinity_val, affinity_type, lipinski_pass, violations_count):
Exact nM potency contributes 60/50/35/20/5 for <10/<100/<1000/<10000/>=10000.
A complete Rule of Five screen contributes 40 (pass) or 15 (fail), minus
10 per violation, floored at zero. Unknown inputs return None.
rank_candidates(ligands, top_k, affinity_type=...): Select one endpoint,
rank scored assay records, and retain unscored evidence afterwards. Mixed
endpoints without explicit selection raise an error. Repeated assays are
retained, so row count is not a distinct-compound count.
format_dossier(target, ppi_records, ranked_leads, provenance=...): Report
evidence, unknowns, source links, retrieval hashes, and interpretation limits.
Descriptor definitions: RDKit Lipinski API
and RDKit molecular descriptors.
1---2name: target-druggability-screening3description: >4 Evidence-attributed target and ligand screening. Retrieves STRING interaction5 records and BindingDB assay measurements, optionally calculates molecular6 descriptors with RDKit, and saves a dossier with source snapshots and limitations.7origin: openai4s8category: biology9requirements: []10capabilities:11 network:12 mode: none13 domains: []14metadata:15 source: https://skillmd.com/skills/pku-yuangroup/target-druggability-screening16---1718# Target Druggability Screening1920Compile a preliminary evidence dossier for a protein target. STRING associations21and BindingDB measurements alone do not establish druggability or therapeutic22relevance. This recipe retrieves a limited sample, not an exhaustive search.2324## Workflow25261. Confirm the target's **UniProt accession and species**. Resolve a gene name27 such as CDK4 through `host.science.search("uniprot", ...)` first; review ambiguous28 matches. BindingDB accepts UniProt accessions or PDB IDs, not arbitrary gene names.29 Use the same confirmed accession for both searches below.302. Retrieve STRING interaction partners with an explicit species. The returned31 functional associations do not measure physical binding, whole-network hub32 centrality, or pathway enrichment.333. Query BindingDB for one endpoint (`Ki`, `Kd`, `IC50`, or `EC50`) at a time.34 Even within an endpoint, review assay conditions before comparing values.354. Use optional RDKit for molecular weight, calculated logP, HBD, HBA, and36 rotatable bonds. The Rule of Five uses the first four descriptors, allowing37 at most one violation. Install the existing `chemistry` extra in the active38 Python environment (`uv sync --locked --extra chemistry`) to enable this step.39 No RDKit, missing SMILES, and invalid structures produce **Unknown**, never40 invented descriptors or a passing verdict. Core imports remain stdlib-only.415. Preserve qualified affinities (`<`, `>`, `~`), missing values, citations, and42 repeated assay records. These are evidence, not distinct confirmed leads.43 Scores require an exact positive nM measurement and a complete descriptor44 screen; otherwise they remain **Unscored**. The score is an unvalidated45 within-endpoint heuristic, not a probability or a biological conclusion.466. Save each complete query envelope with its provenance, then save the dossier47 with lineage to those exact artifact versions. Propagate connector failures;48 do not replace an error with an empty successful search.4950## Interactive / Cell Recipe5152Run inside a persistent Python kernel cell. Host connectors mediate all network53requests; the sidecar performs no direct networking.5455```python56import json57from pathlib import Path5859from target_druggability_screening.kernel import format_dossier, rank_candidates6061# Confirm both values before running: this example is human CDK4.62target = "P11802"63species = "9606"64affinity_type = "Ki"6566ppi_res = host.science.search(67 "string", target, limit=10, filters={"species": species}68)69binding_res = host.science.search(70 "bindingdb", target, limit=20,71 filters={"cutoff": 100, "affinity_type": affinity_type},72)73ranked_leads = rank_candidates(74 binding_res["results"], top_k=10, affinity_type=affinity_type75)7677source_versions = []78for result in (ppi_res, binding_res):79 source_path = Path(f"{target.lower()}-{result['database']}-source.json")80 source_path.write_text(json.dumps(result, indent=2, allow_nan=False), encoding="utf-8")81 saved = host.save_artifact(82 str(source_path), content_type="application/json", source=result["provenance"]83 )84 source_versions.append(saved["version_id"])8586dossier_content = format_dossier(87 target=target,88 ppi_records=ppi_res["results"],89 ranked_leads=ranked_leads,90 provenance=[ppi_res["provenance"], binding_res["provenance"]],91)92dossier_path = Path(f"target-druggability-{target.lower()}.md")93dossier_path.write_text(dossier_content, encoding="utf-8")94artifact = host.save_artifact(95 str(dossier_path), content_type="text/markdown", input_version_ids=source_versions96)97print(f"Dossier saved as artifact version: {artifact['version_id']}")98```99100## Sidecar Functions101102- `calculate_smiles_descriptors(smiles)`: Parsed RDKit descriptors and engine103 version, or a reason for missing/invalid/unavailable chemistry.104- `evaluate_lipinski(descriptors_or_smiles)`: Four-criterion screen; an incomplete105 input has `pass_rule_of_5=None`.106- `score_lead(affinity_val, affinity_type, lipinski_pass, violations_count)`:107 Exact nM potency contributes 60/50/35/20/5 for <10/<100/<1000/<10000/>=10000.108 A complete Rule of Five screen contributes 40 (pass) or 15 (fail), minus109 10 per violation, floored at zero. Unknown inputs return `None`.110- `rank_candidates(ligands, top_k, affinity_type=...)`: Select one endpoint,111 rank scored assay records, and retain unscored evidence afterwards. Mixed112 endpoints without explicit selection raise an error. Repeated assays are113 retained, so row count is not a distinct-compound count.114- `format_dossier(target, ppi_records, ranked_leads, provenance=...)`: Report115 evidence, unknowns, source links, retrieval hashes, and interpretation limits.116117Descriptor definitions: [RDKit Lipinski API](https://www.rdkit.org/docs/source/rdkit.Chem.Lipinski.html)118and [RDKit molecular descriptors](https://www.rdkit.org/docs/GettingStartedInPython.html#list-of-available-descriptors).