Pinal turns protein-function descriptions into laboratory-tested candidates

Knowing what a protein should do is different from knowing how to build it. Pinal starts with a description of the desired function and generates candidate molecules. In a Nature paper published on 21 October, the researchers report laboratory activity across four tasks: fluorescence, PET-plastic breakdown, alcohol conversion and a step in carbon fixation.[1]

The model has 16 billion parameters. Its two-stage design first translates a language instruction into a structural representation, then proposes a sequence that fits that structure and the requested function. This gives researchers a way to specify a molecular goal without manually encoding every geometric constraint.[2]

What was tested in the laboratory

Laboratory screening remained part of the process. For the plastic-degrading enzyme, the team screened 94 candidates for esterase activity, then tested the selected pinalPETase36 on pretreated PET over seven days. They also measured substrate conversion using five purified alcohol-dehydrogenase designs. Those experiments distinguish an active molecule from a sequence that merely looks plausible on a computer.[1]

The H-protein comparison used the corresponding native Escherichia coli protein in a multi-enzyme CO2-fixation pathway. A designed replacement raised product titre 1.7-fold without iterative experimental optimisation. It is a result for that pathway, rather than a general improvement in carbon-fixation efficiency.[1]

How language becomes a design instruction

Training drew on approximately 135 million proteins and 1.7 billion synthetically annotated protein–text pairs. Multiple descriptions can therefore correspond to the same protein; the larger number is not a count of independent laboratory demonstrations. Synthetic descriptions broaden the training material while carrying the possibility of annotation errors.[1][2]

The model card separates the roles of its components: language guides structural generation, and a reverse-folding model uses that representation to design an amino-acid sequence. The laboratory then tests the property that matters to the researcher. Predicted structure and measured function answer different questions.[2]

A new starting point for molecular design

Pinal changes how a researcher can begin a search. A functional description can replace some manually specified sequence or structural constraints, widening the candidates considered. Expression, stability and activity still have to be established experimentally. The four tasks show that this route can reach useful molecules; screening results across more functions will determine how reliably it does so.[1]

Report an error or suggest a correction