torchdrug
by K-Dense-AItorchdrug is a PyTorch-native toolkit for molecular and protein machine learning. Use the torchdrug skill to choose tasks, datasets, and modular models for graph neural networks, protein modeling, knowledge graph reasoning, molecular generation, and retrosynthesis. It is best for custom model development and reproducible configs, not just canned demos.
This skill scores 78/100, which means it is a solid listing candidate for directory users: it is clearly triggerable, covers real TorchDrug workflows, and gives enough structure to justify installation, though users should still expect some adoption friction from the lack of a simple install command or runnable quick-start path in the skill file.
- Strong triggerability: the frontmatter explicitly says to use it for PyTorch-native GNN work in drug discovery, protein modeling, and knowledge-graph reasoning.
- Good operational coverage: the skill body and references map to concrete workflows such as molecular property prediction, protein modeling, retrosynthesis, molecular generation, and link prediction.
- High install-decision value: the repository includes multiple topic-specific references plus explicit dataset and model coverage, helping agents understand where TorchDrug fits and where alternatives like deepchem or pytdc may be better.
- No install command is present in SKILL.md, so users may need outside setup knowledge before they can use it reliably.
- The repository is reference-heavy but script-light, so some tasks may require more manual execution or model selection judgment than a fully operational skill bundle.
Overview of torchdrug skill
What torchdrug is for
The torchdrug skill helps you work with TorchDrug as a practical PyTorch-native toolkit for molecular and protein machine learning. It is best for users who need to build, train, or adapt graph neural network pipelines for drug discovery, protein modeling, knowledge graph reasoning, molecular generation, or retrosynthesis—not just run a canned demo. If you want a torchdrug guide that helps you decide fit before installation, this is the right page.
Who should use it
Use the torchdrug skill if you are turning SMILES, protein sequences, PDB structures, reactions, or biomedical triples into trainable models. It fits researchers and engineers who want custom model development, task selection, dataset choice, and reproducible configurations. It is less useful if you only need generic cheminformatics utilities or a ready-made benchmark wrapper.
What makes it different
TorchDrug’s main value is its modular design: models, tasks, datasets, and config loading are separated, so you can swap components without rewriting the whole pipeline. That matters when you are comparing architectures, changing targets, or moving from molecular property prediction to protein tasks. For torchdrug for Machine Learning, the key advantage is speed of experimentation with domain-specific abstractions, not broad one-click automation.
How to Use torchdrug skill
Install and read first
Install the torchdrug skill with npx skills add K-Dense-AI/claude-scientific-skills --skill torchdrug. After install, start with SKILL.md, then read references/core_concepts.md and the domain file that matches your job: references/molecular_property_prediction.md, references/protein_modeling.md, references/knowledge_graphs.md, references/molecular_generation.md, or references/models_architectures.md. Those files tell you which task class, dataset, and model family to choose before you start coding.
Give the skill a concrete problem
A weak prompt like “use torchdrug on my dataset” usually misses the important setup choices. A better torchdrug usage prompt names the input type, target, split style, and output goal, for example: “Train a TorchDrug model for BBBP binary classification from SMILES, use scaffold split, report AUROC and AUPRC, and show a config-based workflow.” If you are doing protein modeling, say whether the input is sequence, structure, or both, and whether you want function prediction, stability, localization, or interaction prediction.
Use a workflow, not a guess
The torchdrug install is only useful if you follow the repository’s modular path: select a dataset reference, map it to the task definition, then choose a baseline architecture that matches the data shape. For example, molecular property prediction often starts with GCN, GAT, or MPNN-style models; knowledge graph reasoning starts with link prediction tasks; molecular generation often needs a generation-specific objective rather than a standard classifier. If you are unsure, ask for a minimal baseline first, then iterate toward a custom model.
Improve output quality early
Tell the model your constraints up front: GPU budget, dataset size, whether you need config reproducibility, and whether you want a training script, evaluation plan, or architecture recommendation. TorchDrug’s configurable system is especially useful when you want the same experiment expressed as code and as saved config. When possible, ask for the exact files or classes to inspect first so the torchdrug guide stays anchored to the repo’s real task structure.
torchdrug skill FAQ
Is torchdrug only for drug discovery?
No. TorchDrug is strongest in drug discovery, but it also covers protein modeling, molecular generation, retrosynthesis, and biomedical knowledge graph completion. If your work is outside graphs, sequences, structures, or reactions, a different library may be a better fit.
How is torchdrug different from a generic prompt?
A generic prompt may suggest a model idea, but the torchdrug skill is meant to map your problem to TorchDrug’s actual task and dataset abstractions. That reduces the common failure mode of choosing the wrong split, the wrong metric, or a model that does not match the input representation.
Is torchdrug beginner-friendly?
It is beginner-friendly only if you already know the task you want. The repo is approachable for starting with baselines, but it expects you to distinguish classification vs regression, sequence vs structure, and molecular vs protein vs knowledge graph problems. Beginners get the best results by starting with one dataset and one baseline architecture.
When should I not use torchdrug?
Do not choose torchdrug if you mainly need pretrained molecular embeddings, broad tabular ADMET tooling, or benchmark dataset browsing without model development. For those cases, deepchem or pytdc may be a better first stop than a torchdrug install.
How to Improve torchdrug skill
Give stronger task constraints
The most useful way to improve torchdrug output is to specify the task precisely: dataset name, label type, prediction target, metric, and split strategy. “Predict molecule activity” is too vague; “Train on Tox21 multi-label classification with scaffold split and AUROC” gives the model the decision points it needs. For protein work, name the exact endpoint, such as stability or GO prediction, instead of only saying “protein ML.”
Ask for the right baseline first
Common failure mode: jumping straight to a custom architecture before proving the data pipeline works. A better torchdrug usage pattern is baseline first, then specialization: simple model, known dataset, reproducible config, then custom features or a larger architecture. That sequence helps you separate repo integration issues from real modeling issues.
Iteratively refine from repo structure
If the first answer is broad, tighten it by asking for a specific reference path from the skill: for example, references/core_concepts.md for configs, references/datasets.md for dataset choice, or the domain reference that matches your task. This is especially useful when you need a torchdrug guide that produces code you can actually adapt, not just a high-level summary.