adabmDCA: adaptive Boltzmann machine learning for biological sequences

Abstract

Background Boltzmann machines are energy-based models that have been shown to provide an accurate statistical description of domains of evolutionary-related protein and RNA families. They are parametrized in terms of local biases accounting for residue conservation, and pairwise terms to model epistatic coevolution between residues. From the model parameters, it is possible to extract an accurate prediction of the three-dimensional contact map of the target domain. More recently, the accuracy of these models has been also assessed in terms of their ability in predicting mutational effects and generating in silico functional sequences. Results Our adaptive implementation of Boltzmann machine learning, , can be generally applied to both protein and RNA families and accomplishes several learning set-ups, depending on the complexity of the input data and on the user requirements. The code is fully available at https://github.com/anna-pa-m/adabmDCA. As an example, we have performed the learning of three Boltzmann machines modeling the Kunitz and Beta-lactamase2 protein domains and TPP-riboswitch RNA domain. Conclusions The models learned by are comparable to those obtained by state-of-the-art techniques for this task, in terms of the quality of the inferred contact map as well as of the synthetically generated sequences. In addition, the code implements both equilibrium and out-of-equilibrium learning, which allows for an accurate and lossless training when the equilibrium one is prohibitive in terms of computational time, and allows for pruning irrelevant parameters using an information-based criterion.

Funneling modulatory peptide design with generative models: Discovery and characterization of disruptors of calcineurin protein-protein interactions

“…A protocol solely based on PSSMs, which neglects inter-dependencies between positions would yield a much lower success rate, as evidenced by (i) entropy calculations, which show that explicitly modeling inter-dependencies reduces the size of the space of candidate sequences by a factor of at least ~4000 (Methods) (ii) flexible docking simulations, which show that PSSM-generated sequences have worse docking energy than cRBM-generated ones on average and (iii) the microarray experiment, which did not identify any PSSM-generated sequence with high binding affinity to Cn. On the other hand, generative Potts models [ 64 , 26 , 65 ], which also integrate inter-dependencies, model equally well the distribution of natural binders ( Methods , S3A Fig ), and would likely perform comparably. Other machine learning generative models that account for inter-dependencies (see Wu.…”

GENERALIST: An efficient generative model for protein sequence families

“…We use GENERALIST to model sequence variability in proteins that span multiple kingdoms of life, alignment sizes, and sequence lengths. We compare the performance of GENERALIST with three other generative models, the Potts model (referred to as adabmDCA 9 ), the autoregressive . CC-BY-NC-ND 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.…”

“…However, there are significant issues with the Potts model. The associated numerical inference is computationally inefficient 9 , limiting their application to small proteins and protein domains ‫ܮ(‬ ‫‬ 100 residues). In comparison, median protein size in many organisms including humans is much larger ‫(‬ 350 residues) 10 .…”

“…We use GENERALIST to model sequence variability in proteins that span multiple kingdoms of life, alignment sizes, and sequence lengths. We compare the performance of GENERALIST with three other generative models, the Potts model (referred to as adabmDCA 9 ), the autoregressive DCA model (referred to as ArDCA 14 ), and a variational autoencoder-based model (referred to as VAE 19 ). We show that compared to these other models, GENERALIST captures higher order statistics of amino acid covariation across sequences.…”

DCAlign v1.0: Aligning biological sequences using co-evolution models and informative priors

“…The prior parameters (both for gaps and insertions) are extracted from the seed alignment in an unsupervised manner. Finally, to further speed-up the learning of the seed-based objective function, we obtain the parameters of the DCA model using pseudo-likelihood maximization [4] instead of Boltzmann Machine Learning [5,8]. DCAlign, is a computational pipeline that allow to compute the seed-model parameters in a few minutes, contrary to its original implementation which required at least a day of computation in the best scenario.…”