This paper introduced AlphaFold 3, a unified deep learning model that predicts the joint structure of complexes containing proteins, nucleic acids, small-molecule ligands, ions, and modified residues. It replaces much of the prior architecture with a diffusion-based module that directly generates atomic coordinates. The model achieved substantially improved accuracy over specialized tools across many interaction types, including protein-ligand and protein-nucleic acid complexes.
This paper introduced RFdiffusion, a generative diffusion model built on the RoseTTAFold network for de novo protein design. It enables a range of design tasks, including unconditional generation, symmetric oligomer design, functional motif scaffolding, and binder design. Many designs were experimentally validated, with solved structures closely matching the intended models.
Kathryn Tunyasuvunakool, John Jumper and Demis Hassabis
This companion paper applied AlphaFold to predict structures for nearly the entire human proteome and 20 other key organisms, producing a large public database of predicted models. It assessed coverage and confidence across the human proteome, showing that a substantial fraction of residues could be modeled with high or very high confidence. The work created the AlphaFold Protein Structure Database, greatly expanding structural coverage beyond experimentally determined structures.
John Jumper, Richard Evans, Alexander Pritzel, David Silver, Oriol Vinyals and Demis Hassabis
The paper introduces AlphaFold2, a deep-learning system that predicts three-dimensional protein structures directly from amino-acid sequence with near-experimental accuracy. It combines a novel attention-based Evoformer over multiple sequence alignments and pairwise representations with an end-to-end structure module that produces atomic coordinates. AlphaFold won the CASP14 assessment by a wide margin, delivering atomic-level accuracy for the majority of targets.
This paper presented RoseTTAFold, a three-track neural network that simultaneously processes one-dimensional sequence, two-dimensional residue-pair distances, and three-dimensional atomic coordinate information, with information flowing between the tracks. The method achieved protein structure prediction accuracy approaching that of AlphaFold2 while being more computationally efficient. It also demonstrated rapid generation of accurate models for protein-protein complexes.
Longxing Cao, Inna Goreshnik, Brian Coventry and David Baker
The authors used computational de novo protein design to create small, stable miniproteins that bind the SARS-CoV-2 spike receptor-binding domain and block its interaction with ACE2. Two design strategies were used: incorporating the ACE2 helix into a designed scaffold, and building entirely new binders against the RBD. The best designs bound with picomolar affinity and neutralized the virus, with cryo-EM confirming the binding modes matched the computational models.
In this one-page report, Watson and Crick proposed a double-helical structure for the salt of deoxyribose nucleic acid (DNA), consisting of two right-handed helical polynucleotide chains coiled around a common axis and running in antiparallel directions. They proposed that the chains are held together by hydrogen bonding between specific complementary base pairs—adenine with thymine and guanine with cytosine—a feature dictated by the structure. They famously noted that this specific pairing immediately suggested a possible copying mechanism for the genetic material.