Source: https://github.com/aipoch/medical-research-skills

When to Use

• You need curated, AI-ready datasets for drug discovery tasks (e.g., ADME, toxicity, bioactivity, DTI/DDI).
• You want standardized benchmarks with consistent evaluation protocols (including multi-seed benchmark groups).
• You require meaningful data splits such as scaffold splits (chemical diversity) or cold-start splits (unseen drugs/targets).
• You are building models for single-instance prediction (molecular/protein properties) or multi-instance prediction (interactions).
• You are doing goal-directed molecular generation and need property Oracles for scoring/optimization.

Key Features

• Dataset access by task type
  • Single-instance prediction: ADME, Toxicity (Tox), HTS, QM, and more.
  • Multi-instance prediction: DTI, DDI, PPI, and more.
  • Generation: MolGen, RetroSyn, PairMolGen.
• Standardized splitting utilities
  • random, scaffold, and cold-start variants such as cold_drug, cold_target, cold_drug_target, plus temporal where applicable.
• Unified evaluation
  • Built-in Evaluator with common metrics (ROC-AUC, PR-AUC, RMSE, MAE, Spearman, etc.).
• Benchmark groups
  • Curated collections (e.g., ADMET group) with recommended evaluation protocols (commonly 5 seeds).
• Chem/data utilities
  • Molecular format conversion (e.g., SMILES → PyG), filtering, balancing, negative sampling, entity retrieval (CID→SMILES, UniProt→sequence).
• Molecular Oracles
  • Property scoring functions usable for goal-directed generation workflows (see references/oracles.md).

Dependencies

Install (recommended):

uv pip install PyTDC

Upgrade:

uv pip install PyTDC --upgrade

Core runtime dependencies (installed automatically; versions depend on the PyTDC release you install):

• PyTDC (latest from PyPI)
• numpy
• pandas
• scikit-learn
• tqdm
• seaborn
• fuzzywuzzy

Optional dependencies may be pulled in automatically depending on which submodules you use (e.g., graph backends or chemistry toolchains).

Example Usage

A complete runnable example that:

1. loads an ADME dataset,
2. performs a scaffold split,
3. trains a simple baseline model,
4. evaluates with a standard metric,
5. queries an Oracle score for a SMILES.

# pip install PyTDC scikit-learn

from tdc.single_pred import ADME
from tdc import Evaluator, Oracle

from sklearn.pipeline import Pipeline
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.linear_model import Ridge

def main():
    # 1) Load a single-instance prediction dataset (ADME)
    data = ADME(name="Caco2_Wang")

    # 2) Create a scaffold split (train/valid/test)
    split = data.get_split(method="scaffold", seed=42, frac=[0.7, 0.1, 0.2])
    train, valid, test = split["train"], split["valid"], split["test"]

    # 3) Train a simple baseline model on SMILES strings
    #    (character n-gram + ridge regression; replace with your own model)
    model = Pipeline(
        steps=[
            ("featurizer", CountVectorizer(analyzer="char", ngram_range=(2, 5))),
            ("regressor", Ridge(alpha=1.0)),
        ]
    )
    model.fit(train["Drug"], train["Y"])

    # 4) Evaluate on the test set using a TDC Evaluator
    y_pred = model.predict(test["Drug"])
    evaluator = Evaluator(name="MAE")
    mae = evaluator(test["Y"], y_pred)
    print(f"Test MAE: {mae:.4f}")

    # 5) Oracle scoring example (property scoring for a SMILES)
    oracle = Oracle(name="DRD2")
    score = oracle("CC(C)Cc1ccc(cc1)C(C)C(O)=O")
    print(f"DRD2 Oracle score: {score}")

if __name__ == "__main__":
    main()

Related references and templates (if present in this skill package):

• Oracle catalog and usage: references/oracles.md
• Utility functions (splits, processing, retrieval): references/utilities.md
• Dataset catalog: references/datasets.md
• Workflow templates: scripts/load_and_split_data.py, scripts/benchmark_evaluation.py, scripts/molecular_generation.py

Implementation Details

1) Dataset Access Pattern

PyTDC datasets follow a consistent interface:

from tdc.<problem> import <Task>

data = <Task>(name="<DatasetName>")
df = data.get_data(format="df")
split = data.get_split(method="scaffold", seed=1, frac=[0.7, 0.1, 0.2])

• <problem> is typically one of:
  • single_pred (single-entity property prediction)
  • multi_pred (pairwise/multi-entity interaction prediction)
  • generation (molecule/reaction generation tasks)

2) Splitting Strategies (Key Parameters)

Use get_split(...) to obtain {"train": ..., "valid": ..., "test": ...}.

Common parameters:

• method: split strategy
• seed: random seed for reproducibility
• frac: [train, valid, test] fractions (when supported)

Typical methods:

• random: random shuffling split
• scaffold: Bemis–Murcko scaffold-based split to reduce scaffold leakage and improve chemical generalization
• Cold-start (commonly for interaction tasks like DTI/DDI):
  • cold_drug: test contains unseen drugs
  • cold_target: test contains unseen targets
  • cold_drug_target: test contains unseen drugs and targets
• temporal: time-based split for datasets with timestamps (when available)

Example:

split = data.get_split(method="cold_target", seed=1)

3) Standardized Evaluation

TDC provides a unified evaluator:

from tdc import Evaluator

evaluator = Evaluator(name="ROC-AUC")  # classification
score = evaluator(y_true, y_pred)

Choose metrics appropriate to the task type:

• Classification: ROC-AUC, PR-AUC, F1, Accuracy, etc.
• Regression: RMSE, MAE, R2, Spearman, Pearson, etc.

4) Data Schemas (What Columns to Expect)

While schemas vary by task, common conventions include:

• Single-instance prediction (e.g., ADME/Tox):

  • Drug (often SMILES) and label Y
  • sometimes Drug_ID / Compound_ID

• Multi-instance prediction (e.g., DTI):

  • Drug (SMILES), Target (protein sequence), label Y
  • plus identifiers such as Drug_ID, Target_ID

5) Oracles for Molecular Optimization

Oracles provide a callable scoring interface:

from tdc import Oracle

oracle = Oracle(name="GSK3B")
score = oracle("CCO...")
scores = oracle(["SMILES1", "SMILES2"])

Use Oracles to:

• score candidate molecules during generation,
• define optimization objectives,
• compare molecules under consistent property predictors.

For the full list of Oracles and their expected inputs/outputs, see references/oracles.md.