I thought this was one of the most provocative images from the EVOLVEpro paper by @omarabudayyeh and @jgooten. Supported by experimental data, they assert that a protein's fitness and its activity are lowly, or even negatively, correlated. How could this be? Shouldn't nature have selected for proteins with maximum activity? Shouldn't these two landscapes be one-in-the same? As it turns out, the answer is no. - - - - - First, protein language models (PLMs) are trained on large databases of naturally occurring proteins. The 'fitness' they estimate is how likely a sequence occurs in nature or how well it fits the pattern of existing proteins. This fitness score is a proxy for evolutionary fitness. Second, protein activity, which is often experimentally determined, refers to a how potent a specific function or property is that a designer is hoping to optimize. This could be catalytic efficiency, binding affinity, stability or something else. Intuitively, one might expect that proteins that are more 'fit' according to a PLM would also be more active. There are several reasons why this isn't necessarily the case." - - - - - 1. Natural Selection v. Engineered Function - natural selection optimizes proteins for overall organismal fitness; not for maximum activity for a specific function - Eg. an enzyme in nature might be moderately active but stable because of evolution whereas an engineered version may be highly active and not have to worry as much about stability 2. Real Trade-Offs - proteins face evolutionary trade-offs between different properties because they often have multiple roles - high activity might come at the cost of reduced stability or energy consumption; a PLM trained on natural sequences might favor balanced proteins - Eg. a natural protease might have moderate activity to avoid damaging other proteins in a cell whereas an engineered version may not have to worry about this 3. Contextual Differences - proteins evolve their functions in specific environments; these can be different than in an experimental set up - pH, temperature, pressure, and other factors can cause a divergence in data - Eg. a thermophilic enzyme might appear highly fit to a PLM, but be lowly active in room temperature 4. Novel Functions - some mutations might appear unfavorable based on natural sequences (and thus unfit to PLMs); but actually they unlock novel functionality - Eg. a rare mutation could create a new binding pocket in an enzyme, drasstically increasing its activity despite appearing unfit to a PLM All of these contribute to the observed discrepancy between PLM-predicted fitness and experimental activity. To achieve breakout performance, EVOLVEpro combines the strengths of evolutionary embeddings from a PLM with a few-shot domain expert layer that harness small tranches of experimental data. This means it combines the best of both worlds to help us engineer proteins.