Artificial intelligence (AI) has advanced considerably, AlphaFold2 protein models are as good as X-ray results, language models like ChatGPT can pass MBA and medical exams, and deep learning models Midjourney, and Stable Diffusion can emulate artistic styles. Given current progress, could text-based inputs be used for the generative design of artificial proteins, pathways, or even organisms, with traits designed purely by AI? Existing strategies for biotechnology design are founded in knowledge-based approaches, such as rational enzyme engineering or whole pathway design using synthetic biology, often borrowing “parts” from other organisms. Alternatively, desired traits are achieved via random mutagenesis with iterative selection procedures. Both are costly in terms of acquiring knowledge and undertaking experimentation. Recently, advances in protein language models have allowed AI to implicitly “learn” properties that allow sequences to be folded alongside other embedded learning techniques for function prediction from primary sequences. Thus, AI offers varied routes to predicting biological outcomes from DNA sequences. However, AI has not yet been extensively used to design novel functions, despite the wealth of functionally annotated protein products at our fingertips. Thus, generative protein language models for biodesign represent a promising future. We seek to explore current technological limits and challenges, investigate new avenues and methodologies to make this possible and broach discussion around wider issues arising from AI-designed life.
|Number of pages||2|
|Journal||Research Directions: Biotechnology Design|
|Early online date||2 May 2023|
|Publication status||Published - 2 May 2023|