Revolutionizing Drug Discovery with Hyperdimensional Computing

Drug discovery is a multifaceted process that leverages knowledge from biology, chemistry, and pharmacology to identify effective and safe medications. Traditionally, this process includes an expensive and often inefficient virtual screening phase, where candidates are selected from extensive chemical databases like ChEMBL and OpenChem to build smaller, more focused in-house databases for further synthesis.

In recent years, various machine learning algorithms such as random forest, support vector machines, k-nearest neighbors, and gradient boosting have been explored to enhance drug discovery efforts. These models use molecular representations to predict properties, but they often fall short due to their limited ability to capture the complex structural nuances of molecules. Consequently, deep learning models, particularly Graph Neural Networks (GNNs), have gained popularity due to their superior performance in learning detailed molecular features. However, GNNs require significant pre-processing and computational resources, limiting their efficiency and accessibility.

Introducing MoleHD: A Paradigm Shift

Our research introduces MoleHD, an innovative, ultra-low-cost model based on hyperdimensional computing (HDC) that significantly reduces pre-processing efforts and computational demands. HDC is inspired by brain-like attributes such as high-dimensionality and distributed holographic representation, which allows it to generate, manipulate, and compare symbols represented by high-dimensional vectors. Compared to deep neural networks (DNNs), HDC offers several advantages, including smaller model sizes, reduced computational costs, and the capability for one-shot or few-shot learning.

How MoleHD Works

MoleHD begins by tokenizing SMILES strings into numerical tokens. These tokens are then encoded into high-dimensional vectors, or hypervectors, which represent the realistic features of molecules. The hypervectors are used to train an HDC model for molecule classification tasks. This approach bypasses the need for backpropagation and complex arithmetic operations, making MoleHD highly efficient.

Key advantages of MoleHD include:

1. Backpropagation-Free Training: MoleHD does not rely on backpropagation to train its parameters. Instead, it uses one-shot or few-shot learning to establish abstract patterns representing specific symbols.
2. Efficient Computing: Unlike neural networks that require complex operations like convolutions, MoleHD performs simple arithmetic operations such as vector addition. This efficiency allows MoleHD to run on commodity CPUs and complete both training and testing in minutes, compared to GNNs that require extensive GPU time.
3. Smaller Model Size: MoleHD needs to store only a set of vectors for comparison during inference, unlike state-of-the-art neural networks that require large memory for storing numerous parameters.

Significant Contributions and Results

1. Novel Learning Model: MoleHD presents a cost-effective alternative to existing learning methods in drug discovery, demonstrating promising results.
2. Complete Pipeline for HDC-Based Drug Discovery: MoleHD tokenizes SMILES strings into substructure-representing tokens, encodes them into hypervectors, and uses these hypervectors for training and evaluation.
3. Extensive Evaluation: MoleHD was tested on 29 classification tasks from three widely-used molecule datasets under various split methods. Compared to eight baseline models, including state-of-the-art neural networks, MoleHD achieved the highest ROC-AUC scores on average across random and scaffold splits, with significantly reduced computing costs.
4. Design Space Exploration: We developed and evaluated two tokenization schemes (MoleHD-PE and MoleHD-char) and two gram sizes (uni-gram and bi-gram) to explore their impact on performance.

Conclusion

MoleHD represents a significant advancement in drug discovery by providing an efficient, low-cost, and highly effective model for predicting molecular properties. This innovative approach not only outperforms traditional methods but also reduces the computational burden, making it accessible for broader applications. As we continue to refine and expand the capabilities of MoleHD, it holds the potential to transform the landscape of drug discovery.

For those interested in exploring MoleHD further, I have developed a Streamlit app that showcases how the model works and provides an interactive way to experience its capabilities.

Feel free to check it out and see how MoleHD can revolutionize your drug discovery process.

Streamlit App Research Paper