Martin Weigt

This paper introduces ILMC, a sampling strategy for protein language models that improves diversity while steering generation toward desired properties.

This study investigates the reconstructability of evolutionary intermediates using generative sequence landscapes, finding that conditional sampling better captures plausible histories than maximum-likelihood prediction.

This paper introduces a coevolution-aware ancestral sequence reconstruction framework that improves reconstruction by incorporating residue-residue constraints from extant protein families.

The paper quantifies selection efficiency across Escherichia coli lineages by linking mutation scores to population genetics, showing significant variation in selection intensity between synonymous and non-synonymous mutations and pathogenic populations.

The paper uses probabilistic generative models to estimate the number of self-reproducing RNAs and finds that at least 10^39 such ribozymes may exist.

The paper improves generative models for biological sequences by integrating experimental feedback, increasing functional sequence generation from 6.7% to 63.7% in ribozyme design.

The paper develops a two-stage strategy for functional annotation and label-aware sequence generation in protein families using limited labels and phylogeny-aware evaluation.

The paper quantifies fluctuations in protein evolution by analyzing correlations between mutations and shows ancestral sequence properties influence evolutionary dynamics and reconstruction accuracy.

This paper introduces a data-driven epistatic model of protein evolution that reproduces sequence statistics across different time scales and uncovers the collective nature of epistasis.

TULIP is an unsupervised transformer model for predicting TCR epitope binding, outperforming existing methods on unseen epitopes by addressing data limitations and bias.

This paper introduces eaDCA, a sparse generative model for RNA families that efficiently predicts functional sequences and mutation effects, showing comparable performance to complex methods with fewer parameters.

The paper develops a method to generate protein domain sequences that interact with another domain, outperforming existing approaches and exploring fine-tuning and structure prediction.

The paper combines phylogeny and coevolution to improve the inference of interaction partners among paralogous proteins, showing significant performance gains in challenging scenarios.

The paper introduces simple autoregressive models for efficient generative modeling of protein sequences, showing comparable performance to complex models at lower computational cost.

This paper investigates how phylogenetic correlations affect coevolution-based contact prediction in proteins and finds that phylogeny-induced couplings are smaller than functional couplings and have limited impact on contact prediction.

The paper introduces DCAlign, a method for sequence alignment that incorporates coevolution and second-order interactions, improving upon profile models for proteins and RNA.

FilterDCA predicts inter-domain and inter-protein contacts using coevolutionary analysis and interpretable filters, achieving comparable accuracy to deep learning methods while being fully transparent.

The paper evaluates how dataset size and quality affect protein-protein interaction inference using synthetic data and shows quadratic scaling relationships.

The paper predicts unknown protein domain interactions using multi-scale coevolutionary analysis combining phylogenetic and residue-level coevolution data.

The paper shows that phylogenetic correlations alone can be used by DCA-based methods to infer protein interaction partners, performing as well as sequence similarity methods in some cases.

The paper introduces a method to reduce parameters in protein family models by selecting sequence motifs, leading to Hopfield-Potts models that capture residue conservation and covariation with fewer patterns.

This work improves pathogenicity prediction of missense mutations by incorporating sequence context, achieving accuracy comparable to state-of-the-art methods.

This study shows that coupling matrices from DCA contain quantitative physico-chemical information about amino-acid interactions, validated using synthetic data.

This paper introduces a method to infer protein-protein interaction networks from sequence co-evolution, achieving high true-positive rates in test cases.

The paper develops a method to infer mutational landscapes and assesses context-dependent mutation effects in beta-lactamase TEM-1, improving explanatory power by 40% over methods ignoring epistasis.

adabmDCA 2.0 provides a flexible and multi-language implementation of Direct Coupling Analysis for protein and RNA sequence data, enabling tasks like contact prediction and sequence design.