Backstory

Protein structure prediction became crucial in various branches of Life Sciences, such as structural biology, bioinformatics, and drug discovery. By predicting protein structures, scientists can reveal their functions, identify drug targets, and design therapeutic molecules.    

Previously, methods for protein structure prediction faced several limitations, including time-consuming computations, high computational costs, and accuracy issues. To overcome these challenges, innovative approaches were in high demand. 

That's why advanced AI models like AlphaFold, developed by DeepMind for predicting protein structures from amino acid sequences, were introduced. AlphaFold has transformed protein analysis, enabling scientists to understand and calculate molecular structures with unprecedented precision and speed. 

In this blog, we will explore how innovative solutions based on cloud technologies can facilitate protein structure prediction in an intuitive, convenient, and cost-effective manner both for business and science.  

We share a case study of how the Quantori solution was implemented to achieve machine learning inferences (MLI) from AlphaFold with the desired characteristics and initial settings. 

Challenge 1: Interface Complexity 

Traditionally, AlphaFold users often had to configure a virtual machine clustering setup, relying on command-line interfaces (CLIs) and inputting numerous commands to select the right files and initiate the model inference. The entire process is complex and lacks transparency, particularly for researchers without extensive technical expertise. Thus, those seeking a simpler method to obtain MLI outcomes were left without a viable solution. Keeping in mind that many prefer immediate results without the hassle of intricate setups and programming, this challenge remained unresolved.  

Quantori’s Interface Solution

Our objective was to streamline the UX so researchers could simply access a webpage, input a protein sequence of interest and relevant equipment details, initiating their analysis. 

_{Folding Process}

Quantori team has developed a Graphical User Interface (GUI) enabling researchers to obtain MLI from AlphaFold with pre-defined settings and desired characteristics, such as protein types, algorithm variants, etc. 

This tool was created to interact with AlphaFold models quickly and conveniently without the need for coding. Moreover, it allows users to perform protein structure calculations seamlessly without relying on cloud services or a distributed computing infrastructure.  

Plus, a storage solution was developed for storing proteins of interest, and AlphaFold performed all related calculations. This enabled users to view statistics, repeat calculations, and access the history of their computations, providing a comprehensive overview of their work.  

Challenge 2: Computations

Typically, all protein structure prediction calculations are performed on a single GPU machine, which is both bulky and expensive. However, the initial phase of the computing process doesn't require GPU power; it's better suited for CPU processing. Therefore, Quantori proposes dividing the task into two parts, seamlessly connecting them at the backend and utilizing optimized AWS resources. This approach will not only nearly double the speed of calculations but also optimize overall spending. 

Quantori’s Computing Solution

To achieve these objectives, the Quantori team experimented with ML models and their infrastructure, exploring their operations, computation stages, and how to provide clear process updates to users

The team optimized the computations by breaking down the algorithm into parts and configuring each one with the best DevOps setup.

All operations were conducted on AWS infrastructure. For the optimization of time spent and costs for calculations, AWS Batch and AWS EC2 instances (both CPU and GPU) were utilized. The interface allows users to choose a particular type of instance for business optimization.

_{General Solution Architecture}

_{Application Architecture: Frontend & Backend}

Summarizing the Key Benefits of our Solution for Researchers:

1. No coding required: The platform is ready-to-use for those without extensive coding skills. 
2. Interactive, web-based, user-friendly interface. 
3. Powerful computational capabilities with the utilization of AWS Batch. 
4. Customization and efficiency through the ability to select an algorithm based on research needs.   

Perspectives of the Solution

• Protein calculations can be both time-consuming and expensive. We envision a traffic light system (red, amber, green) to provide users with visual cues about the tradeoffs between time and cost. This system would enable scientists to easily choose between saving money with slower processing or spending more for faster results, helping them better predict and manage research costs.     
• The first stage of the two-staged AlphaFold computations is time-consuming and includes numerous operations. While some operations are interdependent, many can be performed separately. These atomic processes can be improved, especially since they were developed a while ago. Quantori has the potential to propose new algorithms and approaches to further optimize the computations for predicting protein structures.