Investigation of biomolecular sequence space of extant proteins has been facilitated by the ever-increasing amount of experimentally determined sequencing data. Along with structural data, this enrichment has led to important advances in protein structure prediction, conformational sampling, and the inference of the effects of mutations in biomolecules. In addition to the increase of sequencing data, novel machine learning algorithms have pushed research questions from analysis and predictive capabilities toward generation of non-extant functional sequences. We describe how different methodologies based primarily on machine learning modeling of sequence alignments can help us expand our understanding of protein families as well as plausible evolutionary trajectories that explore novel sequence space. I will focus on two generative methodologies: (1) latent generative landscapes (LGL) and (2) sequence evolution with epistatic contributions (SEEC). LGLs provide interpretable maps to explore features in a latent space of protein families that are useful to characterize functional classification, phylogenetic clustering and serve as a guideline for the design of novel sequences with preferred characteristics. Diverse examples for different systems are presented, ranging from temperature sensitive proteins, viral proteins to large multidomain ATPase transmembrane transporters. SEEC, alternatively, provides a way to model sequence evolution with realistic statistical features observed in evolutionary data and permits the generation of novel functional trajectories that can expand our knowledge of compensatory effects on fitness, entrenchment, evolvability, and the possibility to test these evolutionary trajectories experimentally. We focus on the family of β-lactamases and their role in antibiotic resistance. Together these two generative approaches to sequence space can be combined to tap into functional and fitness landscapes of protein systems where data is available and can guide concrete and interpretable hypotheses on sequence-function relationships that can be experimentally validated.
Faruck Morcos (Sun,) studied this question.