DNA and Protein Tools for AI-Driven Biologics Design

DNA and Protein Tools for AI-Driven Biologics Design 

Introduction

In 2020, the prediction of protein structure from amino acid sequence was largely solved by Google DeepMind’s machine learning (ML) model, AlphaFold.¹ Now, scientists are putting artificial intelligence (AI) to the test on the hunt for the next holy grail: the de novo design of functional proteins.² Achieving this goal promises to transform research by accelerating target discovery, optimizing candidate de­sign, and reducing development time and cost.³ AI/ML models will do so by helping unpick and predict sequence-function relationships, but this approach will require high-quality, purpose-built datasets to train AI to design new candidates, predict sequence performance, and uncover novel therapeutic pos­sibilities. Service providers that can provide solutions from DNA to data, such as Twist Bioscience, are needed to meet the AI community’s experimental needs wherever they are in the process.

The infographic illustrates an iterative, end-to-end AI-driven biologics discovery workflow, beginning with DNA synthesis and library generation and progressing through primary screening, model development, and candidate selection. Experimental characterization and preclinical data continuously feed back into model training, enabling refinement and improved predictive accuracy. These interconnected design–build–test–learn loops accelerate the identification of lead candidates and support their advancement into clinical trials.

Key Requirements for AI-Ready Datasets

• High sequence accuracy and customization to establish reliable links between sequence and function 
• Sufficient library depth to support effective model learning while matching assay constraints 
• Speed and production readiness to enable rapid design–test–learn iteration 
• Focused experimental validation that prioritizes high-value candidates, reducing time and resource use 

Enabling Experimental Scale for AI Workflows

• Wet-lab data is required for ground truth, translating AI-proposed sequences into measured performance 
• Traditional labs are not optimized for the scale, diversity, and structure of datasets needed for AI/ML workflows 
• Specialized providers enable scale, synthesizing large sequence panels and generating standardized assay data 
• Consistent experimental feedback supports model refinement, candidate triage, and reliable hit selection 

Accelerating Biologics Discovery with High-Quality Data

• Deeper developability insight through stability, aggregation, specificity, and expression data 
• Reduced wet-lab bottlenecks by combining AI-driven candidate prioritization with rapid DNA, protein production, and characterization 
• Earlier elimination of weak candidates using large, standardized, protein-level datasets 
• More efficient discovery progression enabled by precise DNA construction and scalable, high-quality experimental data that supports model refinement 

References

1.           Senior AW, et al. Improved protein structure prediction using potentials from deep learning. Nature. 2020;577(7792):706-710.

2.           Kortemme T. De novo protein design—From new structures to programmable functions. Cell. 2024;187(3):526-544.

3.           Mirakhori F, Niazi SK. Harnessing the AI/ML in drug and biological products discovery and development: The regulatory perspective. Pharmaceuticals. 2025;18(1):47.