# Enhancing the Predictive Power of Macrocyclic Drug Permeability by Knowledge Distillation from Analogous Pretraining Data

## TL;DR

This paper introduces a deep learning model called Multi_DDPP that predicts how well macrocyclic drugs can pass through cell membranes using 2D structures, making drug development faster and more efficient.

## Contribution

The novelty lies in using knowledge distillation and diverse molecular representations to improve permeability prediction for macrocyclic drugs.

## Key findings

- Multi_DDPP outperforms existing machine learning and deep learning methods in predicting macrocycle permeability.
- Node masking identifies key substructures influencing permeability, aiding drug design.
- The model avoids costly 3D modeling and enables efficient prioritization of macrocycles with good pharmacokinetic properties.

## Abstract

Macrocyclic drugs
offer powerful opportunities for modulating
protein–protein
interactions, yet their development is limited by poor and unpredictable
membrane permeability. Experimental testing is slow, and 3D modeling
of macrocycles is computationally demanding due to their large conformational
space. To address this, we present Multi_DDPP, a deep learning (DL)
model that predicts macrocycle permeability directly from 2D structures.
Multi_DDPP employs knowledge distillation to leverage permeability
data from multiple cell lines, improving generalizability, and uses
a task-specific swing-range strategy to reduce label noise. By integrating
diverse molecular representations, including physicochemical descriptors,
fingerprints, molecular graphs, and hybrid features, the model outperforms
existing ML and DL approaches. Node masking highlights the substructures
that contribute most to permeability, and regression extensions incorporating
physiological parameters further refine these predictions. Early 2D-based
permeability prediction with Multi_DDPP avoids the costly generation
of 3D conformers and enables the efficient prioritization of macrocycles
with favorable pharmacokinetic potential.