Code and data-processing utilities for No Pose Left Behind: Integrating Activity and Structural Data with Uncertainty-Aware Multiobjective Learning for Kinase Inhibitor Prediction.
This repository contains the implementation of mOKDDD, a multi-objective E(3)-invariant graph neural network for kinase–ligand binding affinity prediction. The model is designed to learn from experimentally measured activity data paired with in silico-generated kinase–ligand complex structures of varying pose quality.
Unlike structure-based models that rely only on poses below a fixed RMSD cutoff, mOKDDD explicitly models structural reliability. For each kinase–ligand complex, the model jointly predicts:
- binding affinity, expressed as pIC50;
- activity uncertainty;
- pose quality, expressed as a structural reliability score.
The predicted pose quality is used to modulate the contribution of each structure to the activity loss, allowing the model to learn from heterogeneous structural data while giving greater importance to reliable complexes.
Structure-based machine learning for kinase inhibitor prediction is limited by the scarcity of experimentally resolved protein–ligand complexes. Computationally generated structures, such as docked poses, can reduce this limitation, but their usefulness depends strongly on pose quality.
mOKDDD addresses this by combining activity prediction with pose-quality estimation in a shared E(3)-invariant graph neural network.
The workflow consists of four main steps:
(a) Dataset construction.
Two complementary datasets are used during training: Kinodata, an activity dataset containing kinase–ligand complexes with experimental pIC50 labels and a pose-quality dataset containing generated cross-docked kinase–ligand poses with RMSD-derived pose-quality labels.
(b) Graph construction and featurization.
Each kinase–ligand complex is converted into a molecular graph. Atoms are represented as nodes, while covalent bonds and spatial contacts are represented as edges. The graph is featurized using atom-level descriptors, bond-order information, and interatomic distances.
(c) E(3)-invariant message passing.
Both activity and pose-quality mini-batches are processed by the same E(3)-invariant message-passing GNN.
(d) Multi-output readout and joint training.
A multi-output readout predicts binding affinity, activity uncertainty, and pose quality. These outputs are optimized jointly using a multi-objective loss that combines the activity and pose-quality objectives.
The activity objective learns binding affinity and uncertainty from kinase–ligand complexes with experimental activity labels, while the pose-quality objective learns to estimate the structural reliability of generated ligand poses. The predicted pose quality is then used to modulate the activity loss, allowing the model to learn from heterogeneous structural data while giving greater importance to reliable complexes.
For each kinase–ligand complex x, the model predicts:
mu(x) predicted activity
sigma^2(x) predicted activity variance
q_pose(x) predicted pose quality
We currently support installation from source.
git clone https://github.com/raquellrios/multi-objective-kinodata-3D.git
cd multi-objective-kinodata-3D
git checkout paper_release
You can use mamba or conda to set up the environment
mamba env create -f kinodata_env.yml
mamba activate kinodata_env
Then, install the package with
pip install -e .
The raw and processed data can be obtained from Zenodo. After downloading the archives, extract them in the root directory of this repository. See the Zenodo description for more details on the folder structure of the datasets.