adabmDCA: adaptive Boltzmann machine learning for biological sequences.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Boltzmann machines model evolutionary relationships in protein and RNA families.
• They use local biases for residue conservation and pairwise terms for coevolution.
• Model parameters predict 3D contact maps, mutational effects, and functional sequences.

Purpose of the Study:

• To present adabmDCA, an adaptive implementation of Boltzmann machine learning.
• To enable general application to protein and RNA families with flexible learning setups.
• To provide a publicly available code for researchers.

Main Methods:

• Adaptive implementation of Boltzmann machine learning (adabmDCA).
• Application to diverse protein (Kunitz, Beta-lactamase2) and RNA (TPP-riboswitch) domains.
• Utilizes both equilibrium and out-of-equilibrium learning strategies.

Main Results:

• adabmDCA models demonstrate comparable performance to state-of-the-art methods.
• Accurate inference of protein/RNA contact maps.
• Successful generation of synthetic functional sequences.

Conclusions:

• adabmDCA offers a powerful and flexible tool for analyzing protein and RNA families.
• The implementation supports efficient training, including out-of-equilibrium methods.
• It enables parameter pruning for improved model interpretability.