Prompt Engineering of Genomes framework

What is it about?

. Hereby, me and Dr Adhityo Wicaksono are proposing the frameworks to push bioinformatics away from totally blind brute forces, or incorrect heuristics. We are proposing ‘prompt-to-genome’ AI for all domains of life (bacteria, arachea, and eukarya), along with viruses . This frameworks serve also as vibecoding direction for bioinformatics, as extensive artificial genome generation usually generates extensive source codes as well. Some algorithmic considerations, for instance, the invocation of gene prediction method that usually invoke Hidden Markov Model (HMM), and gene annotation with GO database that invoke Directed Acyclic Graph (DAG). A two part system is proposed: First, the constraint compiler; Second, a modular validation stack. We also provided the phased roadmap and community benchmark to move forward with this “vibecoding in bioinformatics” initiative. Important points in our paper: a) every vibecoding effort (in current AI engineering terms, from prompting to graph engineering) must obey and consider the biological law. We provided some example prompts that could lead to a design failure b) Then, we also mention WHO and NASEM frameworks for bioethics and biosafety guardrails. c) a roadmap for moving forward is provided.

Featured Image

Photo by Daniil Komov on Unsplash

Photo by Daniil Komov on Unsplash

Why is it important?

When Dr Andrej Karpathy , artificial intelligence researcher and engineer that’s currently working in Anthropic, and formerly in OpenAI and Tesla, introduced the term “Vibecoding”, the whole social-media universe went into sudden exhilaration. Suddenly, every AI engineer, and even beginners that I know, practiced vibecoding almost immediately. Why is that so? Simply because coding and programming works that took weeks and months, could be resolved in days and hours. Vibecoding already been incorporated as an entry in several prominent dictionaries such as Collins, Meriam-Webster, and Oxford. Its application and frameworks are successfully implemented in computer science (the very traditional domain for coding/programming, of course), finance, HR, geography, physics, mathematics, and others. However, as a multidisciplinary studies that merge computer science and biology, not many frameworks that I know of deal with vibecoding in bioinformatics. In many instances, bioinformaticians are conducting trial and errors for developing their pipeline. It almost perfectly resembles a wasteful brute force algorithm or inaccurate heuristics. There are some initiatives to fill in the gap, and definitely those are published in good papers as well. But we believe that we can propose things that are still unaddressed.

Perspectives