To help remedy the cumbersome, time-consuming, and expense of creating new drugs against biological pathogens, research into artificial intelligence (AI), machine learning (ML), and deep learning (DL) technologies have allowed scientists to focus on rapid drug design and development by combining DL with the principles of geometry to reduce failure rates, length of schedule, and cost-associated risks of drug development.

It is increasingly difficult to develop medical countermeasures (MCMs) against the biological pathogens available to our adversaries. To deter these biological pathogens against the Joint Force and maintain a scientific and technological edge, the Defense Threat Reduction Agency’s (DTRA) Chemical and Biological Technologies Department in its role as the Joint Science and Technology Office (JSTO) for Chemical and Biological Defense, an integral component of the Chemical and Biological Defense Program, invested in basic research with the University of Pennsylvania (UPenn) to rapidly create MCMs using aspects of machine learning.

The traditional drug discovery, design, and development process uses taxing laboratory tests and experiments to evaluate the safety and effectiveness of new drugs. Those with the desired traits are further evaluated in multiple clinical studies before achieving regulatory approval.

To help speed up the process, researchers at UPenn developed a model system that uses DL to identify new antimicrobial peptides (AMPs), which are a class of natural peptides that exhibit antimicrobial activity characterized by broad-spectrum actions including antibacterial, antifungal, antiviral, and antiparasitic effects. In addition, AMPs have low molecular masses and often possess high antimicrobial, antibiofilm, and anti-inflammatory properties.

The UPenn researchers used several DL methods to predict, identify, design, and develop AMPs that offer protection against bacterial pathogens which cause human diseases. Even though they were promising while needing further development, these first-generation computational tools did not include the precise 3D properties of AMPs to improve the accuracy of peptide binding properties, target affinity, and drug-like efficacy.

To overcome this hurdle, the UPenn researchers harnessed Geometric Deep Learning (GDL) that combines DL with the principles of 3D geometry to predict AMPs that possess high degrees of target specificity and affinity, which translates into greater antimicrobial potency. Also, GDL can capture the complex relationships between geometric properties and function of these molecules, as it more adequately represents the 3D non-Euclidean nature of peptide structures than traditional machine learning methods that operate in the 2D Euclidean spaces.

In contrast to traditional DL methods, GDL considers the fundamental structural features in three ways:

1. Manifolds, used to represent the conformational space of AMPs, which is the space of all possible 3D structures that an AMP can adopt

2. Sequences, used to represent the amino acid sequences of AMPs that dictate their structural and functional properties

3. Graphs representing the interactions between AMPs and their target molecules

This comprehensive approach enables GDL to identify the key features essential for antimicrobial activity and is an efficient DL approach for drug design, development, and biological molecule prediction. Candidates identified through GDL that exhibit the desired drug-like characteristics against a wide array of biological pathogens hold promise as effective MCMs, and GDL represents a rapid method to defeat the ever-evolving biological pathogens facing the Joint Force.

Sidebar1:

The general meaning of AI, in the context of scientific research, is the ability of a digital computer to think, learn, and execute simple and complex tasks, and the ability to learn from data and improve performance over time without specific programming, usually associated with human directions. ML, a subfield of AI, is a powerful tool that is trained on large data sets to identify specific patterns resulting in an improved performance of specific tasks. DL is a subfield of ML that uses so called “artificial neural networks” to process complex data. In this regard it mimics the interconnection of the neurons in a brain. It can learn from a huge amount of data and can capture fundamental relationships between different data types.

Sidebar2:

In mathematics and computer science, non-Euclidean geometry focuses on the characteristics of spherical and hyperbolic objects and their associated curvature properties, such as angles, coordinates, points, and surfaces. In contrast, Euclidean geometry focuses on planar surfaces with no curvature. Computational tools and ML algorithms for predicting the 3D properties of AMPs must include all these non-Euclidean factors to account for the molecular structural curvature due to the peptide bonds formation, spatial angles of individual amino acids, and other molecular characteristics.

POC: Annette von dem Bussche-Huennefeld, PhD, [email protected]