Therapeutics machine learning

PyTDC

An open-source platform for training, evaluation, and inference of multimodal biological foundation models, built on the Therapeutics Data Commons.

Get startedGitHub
Python package pytdc-nextml 1.2.0Import as tdc_mlPython 3.9 or newer
Schematic of PyTDC data, models, and evaluationIllustrative diagram, not measured results. Heterogeneous biological data sources feed API-first datasets. A model server retrieves weights and runs inference. Task-specific benchmarks evaluate models.Biological datacells · moleculesAPI-first datasetstasks · splitsModel serverweights · inferenceEvaluationbenchmarks
Schematic of relationships among data, models, and evaluation. Illustrative, not a measured result.

01 / Install

Installation and a first dataset

The documented release is pytdc-nextml 1.2.0 on PyPI. It requires Python 3.9 or newer. Import the library as tdc_ml. The older pip install PyTDC / from tdc import … interface belongs to the legacy TDC package.

Dataset retrieval talks to Harvard Dataverse. If a download fails, check Dataverse status rather than the Python import.

Full getting-started notes cover tasks, splits, and evaluation.

Shell
pip install pytdc-nextml
Python
from tdc_ml.single_pred import ADME

data = ADME(name="Caco2_Wang")
df = data.get_data()
splits = data.get_split()

print(df.columns.tolist())
print(sorted(splits.keys()))
print(len(splits["train"]), len(splits["valid"]), len(splits["test"]))

02 / Research

Featured papers

  1. Conference paper · 2025

    PyTDC: A multimodal machine learning training, evaluation, and inference platform for biomedical foundation models

    Alejandro Velez-Arce, Marinka Zitnik

    Forty-second International Conference on Machine Learning (ICML 2025)

    PaperPDFICML

  2. Workshop paper · 2024

    Signals in the Cells: Multimodal and Contextualized Machine Learning Foundations for Therapeutics

    Alejandro Velez-Arce, Xiang Lin, Kexin Huang, Michelle M. Li, Wenhao Gao, Bradley Pentelute, Tianfan Fu, Manolis Kellis, Marinka Zitnik

    NeurIPS 2024 Workshop on AI for New Drug Modalities

    Workshop paper, not a NeurIPS conference proceedings paper.

    PaperPDFSlidesPoster

All papers, talks, and foundational TDC citations

03 / Platform

Data, models, and evaluation

Data

API-first dataset loaders unify heterogeneous therapeutics sources into ML-ready views, with documented tasks and splits.

Dataset overview

Models

A model server retrieves research-ready biomedical foundation model weights and standardized inference endpoints.

Model server

Evaluation

Task-specific metrics, realistic splits, and documented dataset licenses distinct from the MIT software license.

How to cite

04 / Catalog

Tasks and how to cite

Single-instance prediction

ADME, toxicity, HTS, and other predictions on individual biomedical entities.

Multi-instance prediction

Interactions among drugs, targets, cells, and related entities, including single-cell tasks.

Generation

Molecule generation, retrosynthesis, and structure-based design tasks.

Foundation models

scGPT, Geneformer, and scVI weights on the Hugging Face hub.

Architecture

How PyTDC relates to Therapeutics Data Commons and the model server.

Cite PyTDC

Use the ICML 2025 paper for the platform; cite original dataset papers for individual resources.