KutumLab-T035: DREAM Challenge TARGET 2035 Submission

This repository contains the code and documentation for the KutumLab-T035 team's submission to the DREAM Challenge: TARGET 2035. The goal of this challenge is to predict small molecule binders for the protein WDR91. Our method secured the top 3 spot in Step 1 of the challenge.

Methodology

Our approach is divided into two main steps, each employing different strategies for hit prediction.

Step 1: Fingerprint-based Ensemble Model

In the first step, we developed a machine learning pipeline based on molecular fingerprints.

• Feature Engineering: We selected the top 500 most variable features from 9 different molecular fingerprint types. These were combined and scaled to create the final feature set.
• Model Training: A weighted XGBoost classifier was trained using a 5-fold cross-validation strategy to handle class imbalance and build robust models.
• Prediction: The final score for each compound was the average predicted probability from the five models in the ensemble.

For a detailed explanation, please see the Step 1 Write-up.

Step 2: Max-Probability Ensemble and Embedding Closeness

For the second step, we submitted two distinct predictions:

1. Max-Probability Ensemble: This approach used the same 5-fold XGBoost ensemble from Step 1, but the final score was determined by the maximum predicted probability across the five models, representing the highest confidence prediction.
2. Molformer Embedding Closeness: This method used a pre-trained Molformer model to generate embeddings for test compounds and 14 known reference ligands. The final score was the maximum cosine similarity of a test compound to any of the reference ligands.

For a detailed explanation, please see the Step 2 Write-up.

Repository Structure

.
├── LICENSE
├── README.md
├── STEP-1/               # Notebooks and write-up for Step 1 submission.
├── STEP-2/               # Notebooks and write-up for Step 2 submissions.
└── utils/                # Utility functions for data processing and analysis.

Authors

The KutumLab-T035 team consists of the following members:

• Gautam Ahuja [1,2,3,4,*]
• Rik Ganguly [3,4,*]
• Zonunmawia [2,4,5]
• Bableen Kaur [2,4]
• Sagarika Toor [4,6]
• Vinita Sharma [4,6]
• Aakansha Rai [2]
• Rintu Kutum [1,2,3,4,7,#]

Affiliations:

• [1] Department of Computer Science, Ashoka University, India
• [2] Koita Centre for Digital Health at Ashoka (KCDH-A), Ashoka University, India
• [3] Mphasis AI & Applied Tech Lab at Ashoka, Ashoka University, India
• [4] Augmented Health Systems Laboratory, Ashoka University, India
• [5] Department of Computer Science, International University of Applied Sciences, Bad Honnef, Germany
• [6] Department of Biology, Ashoka University, India
• [7] Trivedi School of Biosciences, Ashoka University, India
• [*] Equal contribution
• [#] Correspondance: rintu.kutum@ashoka.edu.in, rintu.kutum@augmented-health-systems.org

License

This project is licensed under the MIT License - see the LICENSE file for details.