AIPER: Agonist and Inhibitor Predictor

Mac Version: AIPER is a deep learning framework designed to predict functional regulatory effects (activation or inhibition) of drugs on proteins.

The framework integrates drug molecular fingerprints, protein ESM2 embeddings, 
and protein–protein interaction networks to infer functional effects.

Key features:
• Functional effect prediction (activation / inhibition)
• Drug repurposing analysis
• Integration of literature-derived interaction data
• Graph neural network propagation over PPI networks

1. Overview

(1) AIPER is a deep learning–based framework designed to predict the directional functional effects of drugs on proteins, specifically classifying compounds as agonists (activators) or inhibitors of a given target.

The model integrates:

• Protein sequence representations derived from ESM2 embeddings

• Small-molecule molecular fingerprints

• A supervised learning architecture trained on curated drug–protein functional interaction data

(2) Given a drug (SMILES string or chemical name) and a human protein target, AIPER predicts the functional regulatory direction of the drug.

(3) AIPER further includes a UDS (Upstream Drug Scoring) module, which prioritizes candidate drugs by integrating predicted functional effects with protein–protein interaction (PPI) networks and pathway information.

----------------------------------------------------------------

3. Installation and decompression

Please make sure all split files are downloaded to the same folder and then execute the following command according to your operating system(Contains 24 files in total, please confirm the number of downloads):

        copy /b AIPER_part_* AIPER_Win.zip

----------------------------------------------------------------

4. Input Preparation

AIPER supports two input formats:

(1) SMILES + protein name
The input file must contain a column named SMILES

(2) Drug name + protein name
The input file must contain a column named ChemicalName

Example 1:

Example 2:

(3) KD_prediction (Binding affinity category)

• kd < 10 → 3

• 10 ≤ kd < 100 → 2

• 100 ≤ kd < 1000 → 1

• kd ≥ 1000 → 0

• ic50 < 1 → 5

• 1 ≤ ic50 < 10 → 4

• 10 ≤ ic50 < 100 → 3

• 100 ≤ ic50 < 1000 → 2

• 1000 ≤ ic50 ≤ 10000 → 1

• ic50 > 10000 → 0

• 3: High — well-defined ligandable pockets and/or approved drugs available

• 2: Medium — identifiable and druggable binding pockets

• 1: Low — incomplete structural information or weak pockets

• 0: Not tractable for small-molecule targeting

• 3: High — extracellular or cell-surface proteins with strong evidence

• 2: Moderate — partial or intermediate localization evidence

• 1: Low — weak or uncertain accessibility

• 0: Not tractable for antibody-based targeting

7. Notes

• Warnings may appear when processing SMILES strings and do not affect prediction results, e.g.:[xxx] WARNING: not removing hydrogen atom without neighbors

----------------------------------------------------------------

8. Contact

For bug reports or feature requests, please contact the authors.