Deep generative models of genetic variation capture the effects of mutations

Abstract

The functions of proteins and RNAs are defined by the collective interactions of many residues, and yet most statistical models of biological sequences consider sites nearly independently. Recent approaches have demonstrated benefits of including interactions to capture pairwise covariation, but leave higher-order dependencies out of reach. Here we show how it is possible to capture higher-order, context-dependent constraints in biological sequences via latent variable models with nonlinear dependencies. We found that DeepSequence ( https://github.com/debbiemarkslab/DeepSequence ), a probabilistic model for sequence families, predicted the effects of mutations across a variety of deep mutational scanning experiments substantially better than existing methods based on the same evolutionary data. The model, learned in an unsupervised manner solely on the basis of sequence information, is grounded with biologically motivated priors, reveals the latent organization of sequence families, and can be used to explore new parts of sequence space.

Protein Design and Variant Prediction Using Autoregressive Generative Models

“…The autoregressive model is consistently able to predict the effects of mutations across a wide array of proteins and experimental assays (Figure 2a). When compared to other generative models trained on alignments of the same sequences, the autoregressive model is able to consistently match or outperform a model with only site-independent terms (30/40 datasets) and the EVmutation model with pairwise dependencies (30/40 datasets; Hopf et al, 2017); and it competitively matches the state-of-the-art results of DeepSequence (19/40 datasets; Riesselman et al, 2018).…”

“…Mutation effects, sequence families, and previous effect predictions for validation were curated from published work (Riesselman et al, 2018). The naïve llama immune repertoire was acquired from McCoy et al (2014).…”

“…can be used to estimate the plausibility of mutant sequence x (Mutant) relative to its wild-type, un-mutated counterpart, x (Wild-type) . This log-ratio has been shown to be predictive of mutation effects (Hopf et al, 2017;Riesselman et al, 2018). Importantly, this approach is fully unsupervised: rather than learning from experimental mutation effects, we can learn evolutionary constraints using only the space of natural sequences.…”

“…These sequences contain information about what contributes to a stable, functional protein. Generative models of protein sequences are powerful tools in learning these constraints to predict structure, mutation effects, and residue-residue interactions (Hopf et al, 2017;Mann et al, 2014;Riesselman et al, 2018). Yet these methods to date are dependent on alignments of homologous sequences, which impose a forced structure on the data that can introduce artifacts and exclude important information, particularly by alignment gaps caused by insertions and deletions (indels).…”

Expanding functional protein sequence space using generative adversarial networks

“…This imposes a great burden on experimental approaches aiming to design novel protein sequences, such as random mutagenesis 4 and recombination of naturally occurring homologous proteins 8,9 , as up to 70% of random amino acid substitutions typically result in a decline of protein activity and 50% are deleterious to protein function 4,[10][11][12][13][14][15][16] . On the other hand, Artificial intelligence (AI) is not limited by the amount of sequence variations it can process [17][18][19] and, instead of depending on a blind search process, is based on an inference-based one -it infers protein properties 18,20 and function 19,21 directly from training examples.…”