HDBind: encoding of molecular structure with hyperdimensional binary representations.

Area of Science:

• Computational chemistry
• Machine learning
• Computer hardware engineering

Background:

• Traditional drug discovery relies on biophysical methods for binding energy calculations, demanding significant computational resources.
• Deep learning models offer promise but face hardware acceleration challenges due to their complexity.
• Hyperdimensional Computing (HDC) presents an algorithmically simple alternative using high-dimension binary vectors and low-precision arithmetic.

Purpose of the Study:

• To introduce Hyperdimensional Computing (HDC) for fast and efficient screening of drug-like compound libraries.
• To propose novel graph-based encoding methods for molecular data within the HDC framework.
• To evaluate the performance and energy efficiency of HDC against established methods on benchmark datasets.

Main Methods:

• Developed novel graph-based encoding techniques for molecular data representation in HDC.
• Implemented HDC models for molecular screening tasks.
• Compared HDC performance against traditional and deep learning approaches on MoleculeNet and LIT-PCBA datasets.
• Demonstrated HDC on Graphics Processing Units (GPUs) and Field Programmable Gate Arrays (FPGAs).

Main Results:

• Achieved substantial improvements over previous methods using HDC graph-based encoding.
• Demonstrated at least an order of magnitude greater energy efficiency compared to baseline neural network models.
• Validated HDC's effectiveness on multiple hardware platforms, including GPUs and FPGAs.

Conclusions:

• HDC is a viable and highly energy-efficient method for pre-screening molecular libraries in drug discovery.
• Novel graph-based HDC encodings significantly enhance molecular representation learning.
• This work advocates for further research into HDC for developing ultra-efficient drug discovery tools.