Sergey Ovchinnikov

The paper identifies candidate eukaryote homologs of Asgard archaea by analyzing structural clusters, revealing shared expansions of solenoid proteins.

This paper introduces Proton-PottsMPNN, a model that designs pH-sensitive binders by explicitly modeling protonation states, resulting in pH-dependent PD-L1 binders with transition pH values between 4.0 and 5.8.

The paper uses crosslinking mass spectrometry to map protein interactions in Toxoplasma gondii, identifying novel components of essential complexes and revealing structural diversity in mitochondrial machinery.

Trained on missing NMR amide assignments in the Biological Magnetic Resonance Data Bank, Dyna-1 (ESM-3 layer 22 embeddings) predicts micro-to-millisecond conformational exchange seen in relaxation data.

The paper measures protein stability for 1.8 million domains using a new experimental method and develops models that accurately predict stability with high precision.

This paper introduces ProteinEBM, a protein conformational energy model that outperforms existing methods in predicting mutation effects and protein structures.

This study presents a joint optimization approach for modeling protein conformational landscapes and sequence design, showing improved stability prediction and sequence generation.

This paper designs novel solenoid proteins using in silico evolution, achieving successful folding for alpha-solenoids and improved results for beta-solenoids with refinements.

CIRPIN identifies protein homologs related by circular permutation using a permutation-invariant graph neural network, uncovering thousands of novel pairs and revealing natural PDZ domain permutations.

This study shows that large protein language models can predict homo-oligomeric interface contacts, with performance improving with model size and alignment diversity.

The paper introduces SaprotHub, a platform that simplifies training and sharing of protein language models, enabling non-experts to build and collaborate on customized models.

Protein Hunter uses diffusion models to design proteins de novo by iteratively improving sequence and structure through hallucination and re-design.

This study designs a de novo peptide modulator, ELIXIR, that inhibits pathogenic late sodium current in NaV channels, reversing dysfunction in cardiac arrhythmias and epilepsy models.

This paper refutes claims that AF-Cluster fails to use local evolutionary couplings in its MSA clusters and shows they play an important role in predictions.

This paper introduces MSA Pairformer, a memory-efficient protein language model that outperforms existing methods in contact prediction and variant effect prediction.

This paper introduces RFpeptides, a deep learning framework for designing high-affinity macrocyclic peptides, achieving sub-nanomolar binding affinities and accurate structural predictions.

SoftAlign is an end-to-end protein structure alignment method that improves accuracy over existing tools while being compatible with other alignment strategies.

AF2χ predicts protein side-chain rotamer distributions using AlphaFold2, outperforming simulations and existing methods in speed and accuracy.

This study analyzes diatom diversity and abundance in the global ocean using Tara Oceans data, identifying key environmental drivers and community patterns.

BoltzDesign1 inverts AlphaFold3 to design protein binders for diverse targets without model finetuning, achieving high in silico success rates and structural diversity.

The paper uses generative models to predict absolute protein stability, achieving a mean error of 1.5 kcal/mol and a correlation coefficient of 0.7.

This paper introduces a protein design pipeline using AlphaFold2 and ProteinMPNN to refine large de novo protein backbones, improving structural integrity and designability.

The paper investigates how ESM-2, a protein language model, predicts protein contacts by analyzing coevolving residue statistics and local sequence context.

Hallucinating binders by backpropagating through AlphaFold2 weights gave, across 12 targets, binders in 10-100% of the designs tested per target, mostly at nanomolar affinity.

The Open MetaGenomic (OMG) corpus combines 3.1T base pairs and 3.3B protein coding sequences in mixed amino-acid and nucleotide format to pretrain the genomic language model gLM2.

This paper trains a genomic language model on metagenomic data to predict protein co-regulation and function by learning latent relationships in genomic context.

The paper predicts homo-oligomerization across four proteomes using AlphaFold2, finding that 50% of archaea, 45% of bacteria, and 20% of eukaryotes form homomers, with disease mutations enriched at interfaces.

AF2BIND predicts ligand-binding sites using AlphaFold2's pairwise representation without requiring a ligand, outperforming other models and linking bait amino acids to ligand properties.

This paper introduces a differentiable method for inferring evolutionary trees from leaf nodes, achieving low error rates on up to 128 leaves.

Fine-tuning RoseTTAFold as a diffusion denoiser gives RFdiffusion, a general protein backbone generator; a designed influenza haemagglutinin binder matched its model by cryo-EM.