The computational framework established in Chapter 3 demonstrates that natural language can align with protein sequences and guide their generation \textit{in silico}. However, the critical question remains: do text-guided designs produce functional proteins in living systems? Within our knowledge, this chapter provides the first experimental validation of text-to-protein generation, demonstrating that natural language prompts can guide the design of synthetic proteins with wild-type-level function \textit{in vivo} and \textit{in vitro}. We validate BioM3's text-to-protein capability using two experimental systems: SH3 domain binders in yeast osmotic stress response and chorismate mutase enzymes in rescuing the shikimate pathway for synthesis of aromatic amino acids for \textit{E. Coli}. Fine-tuning BioM3 on domain-specific families, we conduct systematic design campaigns with text prompts of varying specificity—from simple functional keywords to detailed mechanistic descriptions. Each design is experimentally tested using high-throughput selection assays coupled with next-generation sequencing, enabling quantitative fitness measurements across hundreds to thousands of synthetic variants. Our results reveal a fundamental trade-off in prompt engineering: detailed prompts specifying protein names and functions yield higher rescue rates (functional designs) but lower sequence novelty, while simpler prompts generate more divergent sequences that explore broader regions of sequence space. Remarkably, mechanistic prompts describing molecular pathways rather than specific proteins successfully generate functional variants despite substantial sequence divergence from training data, suggesting that BioM3 captures latent mappings between functional descriptions and sequence features. Key experimental successes include a "supercharged" SH3 design with binding affinity superior to wild-type (𝐾𝑑 = 0.31 𝜇 M vs. 0.9 𝜇 M) without using supervised assay labels for training, a highly divergent functional variant with only 28% sequence identity to natural SH3 domains, and mechanistic prompt designs achieving wild-type-comparable function. These results establish natural language as a viable interface for protein design, demonstrating that text prompts can specify functional intent that translates into working proteins in biological systems. The experimental validation bridges computational predictions with laboratory reality, providing proof-of-concept for the text-to-protein paradigm. However, important limitations remain: designs are constrained to functions represented in training data, optimal prompt engineering strategies require further investigation, and the relationship between computational confidence metrics and experimental outcomes remains imperfect. These challenges motivate the advanced methods and future directions explored in subsequent chapters.
