MoCHI: neural networks to fit interpretable models and quantify energies, energetic couplings, epistasis and allostery from deep mutational scanning data

Abstract

The massively parallel nature of deep mutational scanning (DMS) allows the quantification of the phenotypic effects of thousands of perturbations in a single experiment. We have developed MoCHI, a software tool that allows the parameterisation of arbitrarily complex models using DMS data. MoCHI simplifies the task of building custom models from measurements of mutant effects on any number of phenotypes. It allows the inference of free energy changes, as well as pairwise and higher-order interaction terms (energetic couplings) for specified biophysical models. When a suitable user-specified mechanistic model is not available, global nonlinearities (epistasis) can be estimated directly from the data. MoCHI also builds upon and leverages theory on ensemble (or background-averaged) epistasis to learn sparse predictive models that can incorporate higher-order epistatic terms and are informative of the genetic architecture of the underlying biological system. The combination of DMS and MoCHI allows biophysical measurements to be performed at scale, including the construction of complete allosteric maps of proteins. MoCHI is freely available (https://github.com/lehner-lab/MoCHI) and implemented as an easy-to-use python package relying on the PyTorch machine learning framework.

A complete map of specificity encoding for a partially fuzzy protein interaction

“…Precisely quantifying the binding of PDZ3 to >100,000 peptides provides an opportunity to evaluate the extent to which binding to each of the four C-terminal residues is independent of sequence variation at the other three sites. We used MoCHI 31 to fit a two-state thermodynamic model to our binding data. The model accounts for the non-linear relationship between the Gibbs free energy of binding (dG) and the fraction of ligand bound to PDZ3 but otherwise assumes that the energetic effects of mutations (ddG) combine additively with no pairwise or higher order energetic couplings between mutations at different sites (Fig.…”

“…To translate the fitness scores, which capture the phenotypic effects of mutations, into free energy terms, we used the MoCHI package 31 to model the fitness with a two-state thermodynamic model for protein binding. Briefly, MoCHI takes as input amino acid sequences of each variant and predicts their fitness while correcting for global non-linearities (non-specific epistasis).…”

“…1 st and 2 nd order interactions for CRIPT N as determined by MoCHI 31 . a)-f) Clustered heatmap of 2 nd order interactions (dddGs) for CRIPT N combinatorial dataset with first order ddG terms for the individual mutations plotted on the x and y axes.…”

Site saturation mutagenesis of 500 human protein domains reveals the contribution of protein destabilization to genetic disease

“…We used MoCHI 63 to fit two-state thermodynamic models on a per-family basis (for families with at least 5 homologs and variants with a mean count >29). We specified a neural network architecture consisting of a single additive trait layer for shared folding energies across the family and a shared linear transformation layer.…”

Genetics, energetics and allostery during a billion years of hydrophobic protein core evolution

“…In this model, coefficients of the additive trait map ( f ) represent first-order ( ΔΔG f ) and pairwise (second-order) energetic coupling terms ( ΔΔΔG f ) applied to one-hot encoded sequence variants ( x ). An affine transformation h maps the molecular phenotype (fraction of molecules folded) to the measurement scale of the target variable (growth rate) 66 .…”