Dock2D: Synthetic Data for the Molecular Recognition Problem.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in structural biology

Background:

• Predicting protein-protein interactions is crucial in computational biology.
• Current methods often use large, complex datasets from the Protein Data Bank, hindering rapid prototyping and non-expert interpretation.
• There is a need for more accessible datasets for algorithm development.

Purpose of the Study:

• To present two novel, simplified datasets, Dock2D-IP and Dock2D-IF, for developing and evaluating algorithms predicting molecular interactions.
• To facilitate the selection and prototyping of new learning-based algorithms for protein-protein interaction prediction.
• To demonstrate a unified approach to learning interaction energy functions.

Main Methods:

• Development of two 2D "toy" datasets: Dock2D-IP (interaction pose) and Dock2D-IF (interaction fact).
• Utilizing 2D shapes as input for molecular interaction prediction tasks.
• Formulating interaction prediction as either an energy-minimization docking problem or a binding free energy estimation problem.

Main Results:

• The proposed datasets allow for efficient algorithm selection and prototyping for molecular interaction prediction.
• A single underlying energy function can be learned by addressing both the interaction pose and interaction fact tasks.
• Baseline solutions demonstrate the feasibility of the proposed approach.

Conclusions:

• Dock2D-IP and Dock2D-IF offer accessible alternatives to large, complex datasets for computational biology research.
• These datasets can accelerate the development and benchmarking of algorithms for predicting protein-protein and other molecular interactions.
• The study highlights the potential of simplified models to uncover fundamental principles in molecular recognition.