deeppurpose

Help install, inspect, run, troubleshoot, and adapt the DeepPurpose molecular modeling library for drug-target interaction prediction, compound property pred...

Security Scan

This is an instruction-only helper for the DeepPurpose library that is internally consistent with its stated purpose, requests no credentials or installs, and mainly advises reading local repo files and using conda/pip to satisfy heavy ML/chemistry dependencies — but follow its advice in a sandboxed environment because runtime steps can download remote pretrained assets and require large native packages.

Purpose & Capability

The name/description match the SKILL.md: guidance focuses on installing, inspecting, running, and adapting DeepPurpose. The files referenced and workflows described (data_loaders, encodings, model entrypoints, pretrained downloads) are appropriate for a DeepPurpose support skill.

Instruction Scope

Runtime instructions direct the agent to read local repository files (README, DeepPurpose/*.py, DEMO/, toy_data/) and to prefer static validation before runtime imports. These file reads are directly relevant to the stated purpose. The SKILL.md explicitly warns about network downloads triggered by dataset and pretrained model helpers.

Install Mechanism

There is no install spec and no code files to execute as part of skill installation (instruction-only). That minimizes disk-write and remote-install risk. The skill documents upstream install commands (conda/pip) for the user, which is expected and reasonable.

Credentials

The skill declares no required environment variables, credentials, or config paths. The heavy dependencies (RDKit, PyTorch, Descriptastorus, DGL, etc.) are documented and expected for this kind of library; they are proportional to the described tasks. The SKILL.md also cautions about downloads performed by helper functions.

Persistence & Privilege

always is false and the skill is user-invocable. There is no request to modify other skills or system-wide settings. The guidance to operate on local repo files is normal for a repo-support skill and does not imply elevated privileges.

Assessment

This skill is an instruction-only advisor for the DeepPurpose repo and is internally consistent, but exercise normal caution: 1) Inspect the repository yourself before running code; the SKILL.md expects you to read local files and prefer a local checkout. 2) Do static checks first (python3 setup.py --name, compileall) — importing modules will trigger heavy native dependencies and may fail or perform network downloads. 3) Run any suggested installs or model downloads in an isolated environment (conda env / container) to avoid contaminating your system. 4) When a workflow would download pretrained assets or datasets, prefer providing local pretrained_dir or explicit permission before allowing automatic downloads. 5) Because the skill source is 'unknown' and there are remote-download behaviors documented (Harvard Dataverse for pretrained assets and BindingDB fallbacks), consider auditing any downloaded artifacts and running in a network-restricted sandbox if you have sensitive data or credentials on the same machine.

This skill is adapted from DeepPurpose, copyright (c) 2020 Kexin Huang, Tianfan Fu, licensed under BSD 3-Clause.

Prefer a local DeepPurpose checkout over web summaries. Treat a directory as the repo root when it contains setup.py, requirements.txt, DeepPurpose/, DEMO/, and toy_data/.

Workflow

1. Classify the request: environment/install, task pipeline, dataset format, pretrained model, notebook/demo adaptation, or troubleshooting.
2. Read only the relevant reference file:
   • installation, dependency sanity, or smoke tests: references/install-and-dependencies.md
   • task/module selection, encodings, splits, and core APIs: references/tasks-and-entrypoints.md
   • dataset loaders, custom text formats, pretrained downloads, and result outputs: references/data-and-pretrained.md
3. Verify advice against local files before answering. Prefer README.md, DeepPurpose/utils.py, DeepPurpose/dataset.py, and the task module the user actually needs.
4. Reuse the upstream API shape instead of inventing wrappers. The maintained paths are:
   • DTI: DeepPurpose/DTI.py
   • compound property prediction: DeepPurpose/CompoundPred.py
   • DDI: DeepPurpose/DDI.py
   • PPI: DeepPurpose/PPI.py
   • protein function prediction: DeepPurpose/ProteinPred.py
   • one-line repurposing and virtual screening: DeepPurpose/oneliner.py
5. Prefer the closest notebook in DEMO/ when the user wants an example or a starting point.

Execution Rules

• Build datasets with DeepPurpose.dataset helpers or local text files in the expected format.
• Encode and split with data_process(...), then build a config with generate_config(...), then call model_initialize(**config) or model_pretrained(...).
• Keep the task/module aligned:
  • DTI uses both drug and target inputs
  • compound property uses drug-only inputs
  • DDI uses X_drug plus X_drug_
  • PPI uses X_target plus X_target_
  • protein function uses target-only inputs
• For repurposing or screening, prefer the existing helpers: DTI.repurpose, DTI.virtual_screening, CompoundPred.repurpose, and oneliner.repurpose or oneliner.virtual_screening.
• Warn when a step triggers network downloads. Dataset helpers and pretrained model helpers fetch remote files.
• Distinguish static validation from runtime validation. DeepPurpose/utils.py imports heavy dependencies immediately, so a real import needs RDKit, PyTorch, Descriptastorus, and related packages installed first.

Source Files

Use these local files as the primary source of truth when present:

• README.md
• requirements.txt
• environment.yml
• setup.py
• DeepPurpose/utils.py
• DeepPurpose/dataset.py
• DeepPurpose/oneliner.py
• DeepPurpose/DTI.py
• DeepPurpose/CompoundPred.py
• DeepPurpose/DDI.py
• DeepPurpose/PPI.py
• DeepPurpose/ProteinPred.py
• toy_data/
• DEMO/