Michael Heinzinger

Translating chemical reactions into enzyme sequences enables formate dehydrogenase engineering

This paper introduces REXzyme, a model translating chemical reactions into enzyme sequences, and demonstrates its ability to generate active formate dehydrogenases for CO2 reduction.

Protein Language Models: Learning From Evolution, Designing Beyond It

This paper argues for evaluating protein language models through prospective experimental success rather than retrospective predictive accuracy.

ContrasTED: contrastive domain embeddings for scalable remote homology classification

ContrasTED improves homology classification by using contrastive learning to assign domains to superfamilies with high accuracy, outperforming existing methods.

Architectonic Spandrels in the Origin of Enzymatic Function

The paper investigates whether functional frustration in proteins is an evolutionary adaptation or a byproduct of fold architecture, finding that frustration hotspots are inherent physical constraints co-opted for function.

FrustrAI-Seq: Scaling Local Energetic Frustration to the Protein Sequence Space

A fine-tuned ProtT5 protein language model predicts per-residue energetic frustration from sequence (Pearson r 0.86 on unseen CATH topologies) and profiles the human proteome in ~17 minutes on one GPU.

On the state of protein function prediction: a report on the fourth CAFA challenge

This paper reports on the fourth CAFA challenge, showing improvements in Gene Ontology term prediction but highlighting challenges in new evaluation settings.

DeePEn - A Depth sensitive benchmark for Protein Engineering

DeePEn evaluates protein engineering models' ability to predict fitness of variants with increasing mutational distance from wildtype.

Pseudoperplexity Probes Memorization in Protein Language Models

This study uses pseudoperplexity to assess whether ProtT5 memorizes its training data, finding detectable but limited memorization.

Protein Language Modeling beyond static folds reveals sequence-encoded flexibility

This paper introduces ProtProfileMD, a model that predicts protein dynamics from sequences, improving homology detection by capturing conformational changes through 3Di tokens.

ProFam: Open-Source Protein Family Language Modelling for Fitness Prediction and Design

ProFam is an open-source protein family language model that improves fitness prediction and design by leveraging evolutionary context through homologous sequences.

Protein Structure Informed Bacteriophage Genome Annotation with Phold

Phold pairs ProstT5-predicted 3Di tokens with Foldseek searches against 1,363,704 curated phage structures, annotating 49.4% of CDS in 1,419 INPHARED genomes versus 34.7% for Pharokka MMseqs2.

Democratizing Protein Language Model Training, Sharing and Collaboration

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.

TMVisDB: Annotation and 3D-visualization of transmembrane proteins

TMVisDB provides a database and visualization tool for transmembrane proteins, enabling large-scale analysis of predicted structures and topologies without requiring costly AlphaFold3-style predictions.

ProtSpace: a tool for visualizing protein space

ProtSpace visualizes protein embeddings in 2D/3D to reveal functional patterns and challenges existing classifications in phage and venom protein analyses.

Bilingual Language Model for Protein Sequence and Structure

ProstT5, a ProtT5 fine-tuned to translate between amino acids and Foldseek 3Di tokens, predicts structure strings for remote homology search and performs inverse folding.

Fine-tuning protein language models boosts predictions across diverse tasks

This study shows that fine-tuning protein language models improves prediction performance across tasks and can be done efficiently, especially for small datasets.

Protein Embeddings Predict Binding Residues in Disordered Regions

The paper introduces a machine learning model, IDBindT5, to predict binding residues in intrinsically disordered protein regions, achieving balanced accuracy comparable to state-of-the-art methods.

AlphaFold2 reveals commonalities and novelties in protein structure space for 21 model organisms

The paper uses AlphaFold2 to analyze protein structures from 21 organisms, identifying 26 novel superfamilies and expanding structural diversity insights.

CATHe: Detection of remote homologues for CATH superfamilies using embeddings from protein language models

The paper introduces CATHe, a method using protein language model embeddings to detect remote homologues for CATH superfamilies, achieving high accuracy and validating predictions with AlphaFold structures.

Domain loss enabled evolution of novel functions in a gene superfamily

The paper reconstructs the evolutionary history of three-finger toxins (3FTxs) and identifies their non-toxin ancestor as a non-secretory Ly6 in squamate reptiles.

LambdaPP: Fast and accessible protein-specific phenotype predictions

LambdaPP provides fast, accessible protein feature predictions including structure and function, using AI models and embeddings from protein Language Models.

TMvisDB: resource for transmembrane protein annotation and 3D visualization

TMvisDB provides a resource for browsing and visualizing predicted transmembrane proteins with topological annotations, using AlphaFold DB and TMbed predictions.

From sequence to function through structure: deep learning for protein design

This paper reviews recent advances in deep learning for protein design, presents a practical pipeline for sequence-to-property analysis, and suggests a generated protein sequence for biosynthetic applications.

Nearest neighbor search on embeddings rapidly identifies distant protein relations

The paper explores using protein embeddings for nearest neighbor search to detect remote homology, highlighting challenges with multi-domain proteins and proposing solutions.

Ultra-fast protein structure prediction to capture effects of sequence variation in mutation movies

EMBER3D predicts protein 3D structures from single sequences rapidly, enabling real-time analysis of mutation effects and structural changes.

SETH predicts nuances of residue disorder from protein embeddings

SETH predicts residue disorder from protein embeddings using ProtT5, outperforming complex methods while being faster and capturing nuanced disorder variations.

Protein language model embeddings for fast, accurate, alignment-free protein structure prediction

This paper presents a method using protein language model embeddings for alignment-free protein structure prediction, achieving performance comparable to co-evolution-based methods at lower computational cost.

Contrastive learning on protein embeddings enlightens midnight zone

This paper introduces ProtTucker, a contrastive learning method using protein embeddings from pLMs to improve recognition of distant homologous relationships in the midnight zone of protein similarity.

Protein embeddings and deep learning predict binding residues for various ligand classes

The paper introduces bindEmbed21, a deep learning method using protein embeddings to predict binding residues for metal ions, nucleic acids, and small molecules, achieving high accuracy without using MSAs or structural data.

ProtTrans: Towards Cracking the Language of Life's Code Through Self-Supervised Deep Learning and High Performance Computing

This paper trains and evaluates language models on protein sequences to predict structural and localization features, achieving state-of-the-art results without evolutionary information.