Scaling laws do not automatically exist everywhere. Much of the machine learning engineering challenge is finding the specific domain, architecture, and data configuration where adding more compute and data reliably yields better performance.
Rather than relying strictly on pristine, curated structures, training protein models on low-quality metagenomic sequences improves downstream protein design. The scale and diversity of noisy data boost generalization even if fragments do not represent real, complete proteins.
A biological model predicting high raw accuracy is dangerous if its confidence estimates are miscalibrated. Calibrated uncertainty, such as AlphaFold's pLDDT score, provides critical behavioral characterization so practitioners can trust where predictions are reliable and avoid wasted laboratory effort.
AlphaFold reproduced static structures deposited in the Protein Data Bank rather than modeling complete protein dynamics or actual conformational ground states. Proteins are complex, disordered, and context-dependent, meaning biological understanding of dynamics remains an open challenge.
Standard databases like the PDB represent processed structural interpretations that discard nuances. Building machine learning models capable of scaling directly over raw cryo-EM micrographs could unlock protein dynamics and distributional information lost in processed structures.