Toward machine-guided design of proteins

Abstract

Proteins-molecular machines that underpin all biological life-are of significant therapeutic and industrial value. Directed evolution is a high-throughput experimental approach for improving protein function, but has difficulty escaping local maxima in the fitness landscape. Here, we investigate how supervised learning in a closed loop with DNA synthesis and high-throughput screening can be used to improve protein design. Using the green fluorescent protein (GFP) as an illustrative example, we demonstrate the opportunities and challenges of generating training datasets conducive to selecting strongly generalizing models. With prospectively designed wet lab experiments, we then validate that these models can generalize to unseen regions of the fitness landscape, even when constrained to explore combinations of non-trivial mutations. Taken together, this suggests a hybrid optimization strategy for protein design in which a predictive model is used to explore difficultto-access but promising regions of the fitness landscape that directed evolution can then exploit at scale.

Unified rational protein engineering with sequence-only deep representation learning

“…Unlike other approaches to representing proteins, namely as one-hot-encoded matrices as in Biswas et. al 2018 3 , RNNs produce fixed-length representations for arbitrary-length proteins by extracting the hidden state passed forward along a sequence. While padding to the maximum sequence length can in principle mitigate the problem of variable length sequences in a one hot encoding, it is ad-hoc, can add artifacts to training, wastes computation processing padding characters, and provides no additional information to a top model besides the naive sequence.…”

“…Traditional approaches to protein engineering rely on random variation and screening/selection without modelling the relationship between sequence and function 1,2 . In contrast, rational engineering approaches seek to build quantitative models of protein properties, and use these models to more efficiently traverse the fitness landscape to overcome the challenges of directed evolution [3][4][5][6][7][8][9] . Such rational design requires a holistic and predictive understanding of structural stability and quantitative molecular function that has not been consolidated in a generalizable framework to date.…”

“…Recently, this flexibility was demonstrated in protein structure prediction, replacing complex informatics pipelines with models that can predict structure directly from sequence 16 . Additionally, deep learning has shown success in sub-problems of protein informatics; for example: variant effect prediction 15 , function annotation 17,18 , semantic search 18 , and model-guided protein engineering 3,4 . While exciting advances, these methods are domain-specific or constrained by data scarcity due to the high cost of protein characterization.…”

Minimum epistasis interpolation for sequence-function relationships

“…Another strategy has been to assume that the observed phenotype is a simple non-linear function of some underlying nonepistatic trait [32,40], a pattern of epistasis known as univariate [8,24], non-specific [31] or global [40,41] epistasis, which appears to be well suited-primarily to sequence-function relationships that are essentially noised versions of single-peaked landscapes. Finally, a variety of machine-learning techniques [8,12,[42][43][44][45] have been employed that can fit more complex forms of epistasis than global epistasis or pairwise interaction models. However, these require substantial tuning and the resulting models exhibit behavior that is difficult to interpret.…”

Deep exploration networks for rapid engineering of functional DNA sequences

“…These models are now beginning to be combined with search heuristics and high-throughput assays to forward-engineer DNA and protein sequences (Rocklin et. al., 2017, Biswas et. al., 2018Sample et.…”