Using AI to Design Proteins Against Deadly Snake Venom  

Researchers at the University of Washington (Seattle, U.S.) have utilized AI to design proteins capable of neutralizing snake venom, successfully protecting mice from lethal snakebites. This study has the potential to transform antivenom production, an expensive task with often disappointing results.   

The Deadly Reality of Snakebites 

Currently, there are around 600 venomous snake species worldwide, with 200 of them capable of killing humans. In fact, according to the World Health Organization (WHO), between 81,000 to 138,000 people die annually from envenoming, while approximately 400,000 suffer permanent disabilities or disfigurements following a bite.  

This alarming statistic has placed snake bites on the list of global health priorities. The WHO aims to halve snakebite-related deaths by 2030 through initiatives such as, ‘The Snakebite Information and Data Platform’. This large database categorizes venomous species with their corresponding antivenoms and allows countries to track and share epidemiological trends on snake envenoming.  

Understanding the Toxicity of Snake Venom 

Snake venom contains cytotoxins, a type of protein toxic to cells that can cause lasting tissue damage. One particularly dangerous class of cytotoxin, known as three-finger toxins (3FTx), binds to and inhibits nicotinic acetylcholine receptors, which play critical roles in the brain. Interfering with these receptors can be fatal. 

This warrants a need for effective antivenom treatments. However, developing and manufacturing them is notoriously difficult.  

Current methods rely on neutralizing antibodies derived from animal immune systems; these antibodies bind to these toxins, preventing their interaction with the nicotinic acetylcholine receptors. While effective, this process is costly and time-intensive, making it inaccessible to many low-resource regions.  

Additionally, antivenom treatments mostly consist of large proteins that are not readily absorbed by tissues, substantially limiting their efficacy. This is a particular problem in the brain, due to its highly selective membrane acting as a protective shield, the blood-brain barrier. 

Can AI-Engineered Antivenoms Change the Game? 

A recent study, published in Nature, addresses these challenges. Led by senior author David Baker, the research team used an array of deep learning (DL) AI tools to design effective proteins that neutralize 3FTx.    

They started by feeding the chemical structures of various venom toxins into the DL model, two of which were true-to-life, and one computer-generated average structure of 86 cytotoxins of the same family.  

The true-to-life toxins were fed into one AI, called RoseTTAFold Diffusion (RFdiffusion), whilst the computer-generated structure was fed into a different model, called Alpha Fold 2 (AF2). Using these different models allowed the researchers to leverage their differing strengths, which can be found in the table below:  

Tool	Core concepts	Rationale for use
RFdiffusion	RFdiffusion is useful for generating new proteins based on 3D structures of known protein sequences.   This generative diffusion model introduces Gaussian noise into the input data.  The model then learns how to denoise the data using a mean-squared error loss, capturing the difference between predicted structures and the original protein.  This denoising is combined with geometric deep learning, where proteins are represented as graphs with nodes and edges . The model then learns the 3D structure of the inputs and can use this to predict molecules that bind to the	RFdiffusion is excellent at refining noisy or incomplete data, meaning it has more flexibility to produce large proteins with a range of spatial arrangements.  RFdiffusion is therefore highly suitable for predicting binders that interact with complex or previously unexplored proteins.
AF2	AF2 predicts the 3D structures of proteins with high accuracy. It incorporates different deep-learning techniques, including attention mechanisms, convolutional layers, and graph-based representations.   At the core of AF2’s architecture is an Evoformer module, a specialized neural network that integrates the input amino acid sequence with multiple sequence alignment information, which essentially provides a background for each amino acid based on existing experimental data.	AF2 achieves high accuracy when using known sequences of proteins.   In this case, the input data was the consensus sequence, the average sequence of amino acids across highly similar proteins. This is likely to be highly stable without much noise.

Using the cytotoxin structures as input data, AF2 and RFDiffusion generated the backbones for 2,000 proteins that could potentially interact with them. 

These designs underwent further refinement using additional AI tools:  

1. ProteinMPP was used to predict specific sequences of amino acids that would likely fit into the backbone’s dimensions. 
2. FastRelax optimized the stability of the predicted protein by resolving any molecular clashes between amino acids in the sequence. 
3. AF2 Initial Guess tested the other two models’ predictions by generating a 3D structure from the final sequence of amino acids. This could be compared with the initial backbone structure to see if the two align.
    

To validate the predictions made by the AI models, the team inserted genetic information into bacterial DNA, providing these cells with ‘instructions’ to produce the desired proteins. 

Next, the group extracted these AI-designed proteins and tested their ability to neutralize snake venom toxins.  

Impressively, the newly generated proteins displayed efficacy comparable to the current ‘best’ treatments available. These proteins also displayed high thermic stability and were considerably smaller, which may facilitate easier absorption by different tissues.  

Concluding Remarks: What Does this Mean for the Future of Antivenoms? 

It is an exciting era for AI’s emerging role in drug design, particularly when considering the size and complexity of protein-based drugs such as antivenoms. Here, AI has successfully generated new proteins that bind to a given target; crucially, these proteins are stable, effective against cytotoxins, and are perhaps better absorbed by the body.  

More generally, using AI to design these tricky compounds is much quicker and cheaper than traditional methods. This has the potential to ‘trim the fat’ around antivenom development, making treatments more accessible to areas with fewer available resources. 

It is fair to say that this study is certainly a step in the right direction for fighting the deadly battle against snakebites.