transformer / foundation model

Pallatom2 is a generative model for protein design that achieves high-affinity binder discovery and experimental success across multiple targets.

This paper designs de novo ligases that achieve site-specific ubiquitylation of a target protein with high selectivity and efficiency.

The paper introduces a protein language model pretrained on continuous biophysical fields, showing improved performance on thermostability prediction compared to sequence-only models.

RPFace predicts RNA-protein interfaces directly from sequence, outperforming existing methods on RNA interfaces and generalizing to in vivo data without pretraining.

FoldAbS repurposes protein folding models as foundation encoders to improve antibody screening by evaluating specificity, affinity, and developability in a unified framework.

This study evaluates AlphaFold 3's accuracy and reliability in predicting RNA structures and protein-RNA interfaces, finding that confidence scores do not always correlate with correct predictions.

The paper introduces AtomLM-9b, a language model using an all-atom structure tokenizer to predict protein structures with performance comparable to AlphaFold 3.

IDPForge generates all-atom ensembles of intrinsically disordered proteins and regions using a transformer diffusion model, showing good agreement with experimental data.

HighFold-Lasso predicts lasso peptide structures using topology-guided methods, achieving perfect topology success and low RMSD on a test set of 52 structures.

This paper shows that biological concepts can be represented as directions in foundation model embeddings, enabling interpretable and reusable analysis of diverse biological properties.

The paper uses AlphaFold2 to model class B β-lactamases without experimental structures, tests the BBL numbering scheme's validity, and identifies structural exceptions.

Reward-guided inference-time search over the Proteína-Complexa sequence-structure codesign model yields validated binders across 127 targets, kinases and a free carbohydrate, with codesigned sequences outperforming ProteinMPNN redesign.

Inferring tested-but-undetected pairs from prey lists of 5,530 published interaction screens yields 97 million tested human protein pairs and 18,982,508 non-interactions (≥3 tests) for calibrating screens and predictors.

This study uses ESM3 to generate knotted proteins, finding high success rates with guided decoding and robustness in real knotted proteins.

The paper structures and engineers deep-sea halophilic PET hydrolases to improve their salt tolerance and catalytic efficiency for industrial applications.

Web resource offering predicted monomeric and oligomeric structures for 1,603 M. leprae proteins, with pockets, epitope scores, essentiality and a target priority ranking, plus a nanopore resistance-calling module.

WaterFlow predicts ordered water molecule positions in protein structures with sub-angstrom accuracy, outperforming existing methods and improving drug design and crystallographic refinement.

AntiSite predicts antibody and nanobody paratopes using shared models with or without structural data, achieving high performance across benchmarks and generalizing to nanobodies.

This paper designs BMP-2 binders using AI pipelines to control bone regeneration, finding a binder with 2.07 nM affinity for BMP-2.

This study introduces DynGeo-Pheno, a geometric deep learning framework that predicts disease-associated phosphosites and missense mutations by integrating protein dynamics and structural information.

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.

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.

This paper introduces InterAb and InterAb-Opt, computational frameworks that predict and optimize broadly neutralizing antibodies using all-atom modeling and pre-trained language models, showing significant improvements in binding affinity and neutralization activity.

MolX is a geometric foundation model that jointly learns representations of protein pockets and ligands from 3D structural data, achieving state-of-the-art performance in predicting molecular interactions.

This paper introduces Boltz2ESI, a framework predicting enzyme-substrate interactions using a foundation model, outperforming existing methods by capturing active-site plasticity and evolutionary context.

This paper evaluates limitations of deep learning in biomolecular structure prediction, highlighting the benefits of explicit ion modeling for improving conformational accuracy.

A reasoning LLM combining ESM3 embeddings with GO-GPT term predictions produces protein function annotations that experts rated tie-or-better than UniProt in 79% of cases (SFT model).

ProtBLIP2-SST generates open-ended protein function captions by integrating sequence, structure, and text using a two-stage framework.

This paper adapts co-folding models for structure-based protein-protein docking by replacing their template module with a docking module and training with a flow-matching objective, showing competitive performance on specific targets.

ConformFlow generates protein conformational ensembles using a scalable normalizing flow framework, outperforming diffusion-based methods in speed and generalization.