How can machine learning improve protein structure prediction?

1 Sequence alignment

One of the first steps in protein structure prediction is to find similar sequences of amino acids that have known structures, called templates, and align them with the target sequence. This can provide clues about the structure and function of the target protein. However, finding and aligning sequences can be difficult, especially when the similarity is low or the sequences are long. Machine learning can help with this task by using neural networks, which are models that can learn complex patterns and relationships from data, to generate more accurate and efficient sequence alignments. For example, AlphaFold, a machine learning system developed by DeepMind, uses neural networks to align sequences and predict distances and angles between amino acids, which are then used to construct the protein structure.

2 Ab initio modeling

Another approach to protein structure prediction is to use physical principles and energy calculations to model the protein structure from scratch, without relying on templates. This is called ab initio modeling, and it is useful when there are no suitable templates available or when the target protein has a novel structure. However, ab initio modeling is very computationally intensive and often produces many possible structures that need to be evaluated and refined. Machine learning can help with this approach by using generative models, which are models that can create new data based on existing data, to generate realistic and diverse protein structures. For example, trRosetta, a machine learning system developed by the Baker lab, uses generative models to create protein structures based on predicted contacts between amino acids, which are then refined by energy minimization.

3 Quality assessment

A final step in protein structure prediction is to assess the quality and accuracy of the predicted structures, and compare them with experimental data or other methods. This can help to select the best structure or improve the prediction process. However, quality assessment can be subjective and difficult, as there are many factors and criteria that can affect the structure and function of a protein. Machine learning can help with this step by using supervised learning, which is a type of learning that uses labeled data to train a model to make predictions, to score and rank the predicted structures. For example, ProQ4, a machine learning system developed by the Wallner lab, uses supervised learning to predict the quality of protein structures based on features such as contacts, torsion angles, solvation, and secondary structure.

Machine learning is a powerful tool that can improve protein structure prediction in biotechnology, by helping with sequence alignment, ab initio modeling, and quality assessment. By using machine learning, biotechnologists can gain more insights into the structure and function of proteins, and design new proteins for various applications, such as therapeutics, diagnostics, and biocatalysts. Machine learning can also enable faster and cheaper protein structure prediction, which can reduce the need for costly and time-consuming experiments.

4 Here’s what else to consider

This is a space to share examples, stories, or insights that don’t fit into any of the previous sections. What else would you like to add?