Debora S. Marks

The paper introduces a hybrid framework to isolate stability-independent activity effects in enzymes using dual-graph deconvolution, validated with experimental data.

PG-LLM benchmarks language models for protein variant ranking, showing they outperform many sequence-based predictors but fall short of specialized models.

The paper uses machine learning to guide cell-free expression of carbonic anhydrases, identifying variants with improved stability under high temperatures.

The paper develops a synthetic Fab yeast display library integrated with machine learning to discover antibodies against cell surface antigens, yielding hundreds of validated antibodies and a public dataset.

The paper evaluates how protein fitness models predict and design variants with altered specificity, finding that models biased against such changes can be leveraged by using weighted differences in model scores.

The paper benchmarks sequence- and structure-based protein models using a unified framework, showing structure-based models outperform others in generating functional variants.

The paper uses AlphaFold-Multimer to predict nanobody binders to MRGPRX2, demonstrating in silico antibody discovery for a GPCR target.

The paper investigates the determinants of metal import and specificity in the bacterial transporter DraNramp by analyzing sequence variants and their effects on Mn2+ and Mg2+ import.

evedesign provides a unified framework for conditional biosequence design, enabling multiobjective workflows and experimental integration across various applications.

The paper models SARS-CoV-2 evolution during the pandemic, finding that immune escape became more important than infectivity as variants evolved.

This paper uses machine learning to predict high-affinity nanobody binders from sparse sequencing data, outperforming deep learning and semi-supervised methods with linear models.

This paper introduces popEVE, a deep generative model for proteome-wide disease variant prediction, identifying 442 genes in a developmental disorder cohort, including 123 novel candidates.

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.

This study develops a computational pipeline for analyzing spatial proteomics data to identify tissue states associated with chemotherapy response in TNBC, revealing distinct spatial patterns and predictive markers.

This paper evaluates variant effect prediction models on viruses, showing that state-of-the-art models underperform on viral proteins compared to alignment-based models, especially in low-sequence regimes.

The paper introduces a protein engineering competition to benchmark computational models for protein design and characterization, with results from a pilot tournament involving seven teams.

RNAGym provides a benchmarking framework for RNA fitness, secondary structure, and tertiary structure prediction using diverse datasets and standardized assays.

The paper measures catalytic constants for diverse Adenylate Kinase sequences and mutants, revealing distinct catalytic peaks and challenging existing hypotheses about thermophilic enzyme kinetics.

LOL-EVE predicts promoter variant effects using a transformer model trained on mammalian promoter sequences, showing improved performance in indel prediction and haplotype analysis.

This paper introduces a method to rapidly evolve genes in the lab, revealing new insights into enzyme adaptation that machine learning models cannot predict.

This study uses evolutionary models to design highly divergent protein variants with enhanced functional properties while maintaining structure.

The paper identifies a relationship between protein structure and glycan structure, introducing a glycoimpact matrix to predict glycosylation patterns and validate it through disease variants and engineered antibodies.

The paper harmonizes single-cell perturbation datasets and introduces E-statistics for analyzing perturbation effects across diverse data sources.

ProteinGym unifies over 250 deep mutational scans and ~65k clinical variants to benchmark over 70 models in zero-shot and supervised settings; TranceptEVE ranks best overall.

This paper introduces a de-biasing method for sequence models of protein fitness to improve predictions of functional variants far from natural sequences.

This paper introduces ProteinNPT, a non-parametric transformer for protein property prediction and design, showing improved performance in label-scarce and multi-task settings.

Blending an alignment-based family model with an autoregressive protein language model, weighted by how deep the alignment is, reportedly improves mutation-effect prediction against experimental assays and ClinVar.

The paper develops a high-content assay to identify viral genes that suppress innate immunity and validates known and novel immune suppressors from various viral families.

The paper identifies extremotolerant proteins that form condensates and protect human cells from apoptosis, including a synthetic protein and a region of ApoE.

MaveDB v2 curates over three million variant effects from MAVE studies to facilitate analysis and dissemination of functional assay data.