Open-source AI expands protein universe by 1 billion structures: systemic shift in biotech access or new enclosure risks?
Original framing: “Move over, AlphaFold: open source model predicts shape of 1 billion proteins” — Nature
ESMFold2’s breakthrough relies on ESM-2, a transformer model trained on 650M proteins, but its accuracy drops for underrepresented protein families. The model’s reliance on UniRef and other proprietary datasets introduces bias, as these databases overrepresent proteins from model organisms. Peer-reviewed critiques highlight that protein folding is not purely a computational problem—it requires experimental validation, which the model cannot provide. The scientific community’s focus on 'scaling' risks overshadowing reproducibility and generalizability issues.
The ESMFold2 protein atlas exemplifies how AI-driven 'open science' can both democratize knowledge and deepen structural inequities, repeating historical patterns of colonial extraction and corporate enclosure.
While the model expands the protein universe by a billion structures, it does so within a Western biomedical paradigm that sidelines Indigenous epistemologies, marginalizes Global South researchers, and risks privatizing public goods. The narrative’s focus on technical breakthroughs obscures the geopolitical and ethical dimensions of who controls protein data and who benefits from its applications. A systemic solution requires reimagining protein science as a collaborative, decolonial enterprise—one that integrates Indigenous knowledge, redistributes computational power, and enshrines ethical licensing to prevent the repetition of past enclosure cycles. The trickster’s insight reveals that the 'open' label is a Trojan horse: the real battle is over who gets to define the future of life itself.