DeepPurpose

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. 1. Classify the request: environment/install, task pipeline, dataset format,

pretrained model, notebook/demo adaptation, or troubleshooting.

1. 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

1. 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.

1. 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

1. 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. INLINECODE32

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:

• - INLINECODE33
• INLINECODE34
• INLINECODE35
• INLINECODE36
• INLINECODE37
• INLINECODE38
• INLINECODE39
• INLINECODE40
• INLINECODE41
• INLINECODE42
• INLINECODE43
• INLINECODE44
• INLINECODE45
• INLINECODE46