What the field is writing about its own models.

Open source models are great but sometimes hard to run. This blog post provides tutorials to run ECMWF new AI weather forecast model, AIFS Single 2,0, on any GPU using Hugging Face Jobs or your own hardware.

Doctolib Research Lab released a pretrained medical encoder. Here is how to run it, and how to finetune it for the tasks you actually have.

A Web Data Recipe for Medical Encoder Pretraining Domain-specific encoders like medical ones are mostly pretrained on small, hand-curated corpora. What if we built one from the heterogeneous web instead?

A tour of tropical quiver representations and how their combinatorial structure connects to modern AI architectures.

A concrete construction toward the Hilbert–Pólya conjecture using surface orders and cyclic-time symmetry as a route to the Riemann Hypothesis.

A protein sequence design model fine-tuned with GFlowNets for thermostable and kinetically-aware enzyme engineering.

GPT-5.2 conjectured a compact formula for single-minus gluon tree amplitudes previously assumed to be zero for 40 years — a striking example of AI contributing to original theoretical physics.

A thorough primer on the state of AI for protein science — covering structure prediction, protein language models, generative design, and the full open-source model landscape.

How SandboxAQ's SAIR dataset of 1M+ protein–ligand structures is enabling AI-powered drug discovery with unprecedented structural coverage.

Boltz-2 outperforms AlphaFold3 on antibody-antigen interfaces and sets a new state of the art for protein-ligand binding affinity prediction.

BoltzDesign1 extends de novo binder design beyond protein targets to small molecules, RNA, DNA, and metal ions.

OpenFold3 is a fully open-source, commercially available AlphaFold3 alternative backed by the OpenFold Consortium — enabling unrestricted biomolecular structure prediction.

IntFold establishes a new state-of-the-art protocol for biomolecular complex structure prediction, setting records across standard benchmarks.

Chai-1r achieves AlphaFold3-level accuracy on protein-protein and antibody-antigen complexes with fully open weights and no usage restrictions.

A fully open reproduction of DeepSeek-R1's math reasoning training pipeline — data, code, and models — bringing transparent reasoning model training to the community.

A clear, intuitive walkthrough of how PINNs embed physical laws directly into neural network training — bridging traditional PDE-based modeling with data-driven deep learning.

Applying AI to understand and predict food allergies.

Exploring AI approaches to solving partial differential equations.

Examining the therapeutic potential of computationally designed miniproteins as a next-generation alternative to traditional antibody drugs.

A practical guide to designing antibody VHHs and scFvs de novo using RFdiffusion and ProteinMPNN, from target epitope to validated sequence.

A challenge for advancing fusion energy through AI.

Understanding drug interactions through AI-powered pharmacogenomics.

Generative models for protein design and engineering.

How to create and use antibody embeddings for therapeutic applications.

ML approaches for predicting key biophysical properties of therapeutic antibody candidates — stability, solubility, and immunogenicity — before wet-lab validation.

AI-powered promoter sequence design and analysis.

Introducing a large-scale dataset for SAR and optical remote sensing with language descriptions.

A blind challenge for predicting ADMET properties in drug discovery.

Introducing LeMaterial, a community effort to build the largest open database of materials and accelerate AI-driven discovery of new compounds and structures.

How computationally designed de novo miniproteins and minibinders are being developed as a new class of targeted therapeutics.

A practical guide to screening for protein–protein interactions (PPIs) as drug discovery targets using structure prediction and ML scoring.

Boltz-1 from MIT achieves AlphaFold3-level accuracy on protein and protein-ligand structure prediction with no restrictions on commercial use or input types.

Exploring generative AI approaches for designing multistep enzymatic pathways for biosynthesis and biocatalysis.

A comprehensive guide to improving protein function, stability, and diversity using generative AI and ESM-2.

Using RFDiffusion with custom guiding potentials to steer protein structure generation toward desired functional properties.

Using ESM-2 protein language model embeddings to score and predict the functional impact of point mutations.

Accelerating persistent homology alignment with ESM-2 embeddings and persistence landscapes for protein complex clustering.

Combining persistent homology with ESM-2 fine-tuning for protein–protein interaction network prediction and complex clustering.

Generating and optimising peptide binders for target proteins using ESM-2 embeddings and directed evolution.

Replacing traditional multiple sequence alignments with ESM-2 embeddings and persistent homology for structure-aware protein comparison.

Using ESM-2 and EvoProtGrad to simulate directed evolution in silico, optimising protein sequences for target properties.

Applying QLoRA fine-tuning to ESM-2 for accurate prediction of post-translational modification sites across protein sequences.

Measuring the intrinsic dimensionality of ESM-2 protein embeddings to understand the geometric structure of protein sequence space.

Using ESM-2 embeddings and linear sum assignment to predict protein–protein binding partners at scale.

Ensemble methods for ESMBind models to improve binding site prediction accuracy and robustness across protein families.

Fine-tuning ESM-2 with LoRA adapters to predict protein binding sites with high accuracy and parameter efficiency.

A continuously updated, near-comprehensive survey of ML techniques applied to experimental, phenomenological, and theoretical high-energy physics — maintained by the Inter-Experimental LHC ML Working Group.