In 1972, Christian Anfinsen received the Nobel Prize in Chemistry for demonstrating that a protein’s amino acid sequence contains all the information necessary to determine its three-dimensional structure. For the following five decades, translating that principle into a computational method remained one of biology’s central unsolved problems — the protein folding problem. In November 2020, DeepMind’s AlphaFold 2 effectively solved it. Three and a half years later, AlphaFold 3 extended the framework to encompass virtually every class of biomolecule relevant to drug discovery.
The Protein Folding Problem: Why It Took 50 Years
Proteins fold into specific three-dimensional conformations determined by the complex interplay of hydrophobic interactions, hydrogen bonds, electrostatic forces, and entropy. A protein with just 100 amino acids has an astronomical number of possible conformational states — traditional computational approaches, including molecular dynamics simulation and homology modeling, could only reliably predict structures for proteins with close sequence similarity to experimentally solved structures.
The Critical Assessment of Protein Structure Prediction (CASP) competition, held biannually since 1994, benchmarks prediction accuracy using a metric called GDT_TS (Global Distance Test Total Score), where 100 represents a perfect prediction. Before AlphaFold, the best CASP13 results in 2018 achieved GDT scores around 60–70 for difficult targets. AlphaFold 2 at CASP14 in 2020 achieved a median GDT score above 90 across all targets — a margin of improvement so large that the CASP organizers described it as a solution to the problem.
AlphaFold 2: The Architecture That Changed Everything
Jumper et al., publishing in Nature in 2021, described the AlphaFold 2 architecture: a deep neural network combining multiple sequence alignments (MSAs), pairwise residue representations, and a novel Evoformer module that processes both sequence and spatial information simultaneously. The key innovation was representing both the 1D sequence information and the 2D inter-residue relationships in a joint attention mechanism — allowing the model to learn evolutionary constraints encoded across thousands of homologous sequences.
In 2022, DeepMind and EMBL’s European Bioinformatics Institute (EBI) released the AlphaFold Protein Structure Database, containing predicted structures for over 200 million proteins — essentially the entire UniProt database of known protein sequences. For the first time, structural biology had a near-complete catalogue of the proteome of essentially every organism that has been sequenced.
AlphaFold 3: Extending to the Full Drug Target Universe
Published in Nature in May 2024 by Abramson and colleagues, AlphaFold 3 extended the framework beyond proteins to include DNA, RNA, small molecules (ligands), ions, and covalent modifications in a single unified model. This is the step that makes AlphaFold directly relevant to drug discovery rather than just structural biology: the relevant question for medicinal chemists is not only what a protein looks like in isolation, but how it interacts with potential drug molecules.
AlphaFold 3 demonstrated substantially improved accuracy over AlphaFold 2 on protein-ligand binding prediction, as measured on the PoseBusters benchmark — achieving 76% of test cases within 2 Angstroms RMSD of the crystallographic pose, compared to 49% for the previous best method. For protein-nucleic acid complexes, AlphaFold 3 outperformed existing methods by roughly 50% on comparable benchmarks.
Specific Drug Targets and Pharmaceutical Response
The practical implications for drug discovery are substantial. Historically, structure-based drug design required X-ray crystallography or cryo-electron microscopy to obtain the protein structure — processes that can take months to years and fail entirely for proteins that resist crystallization. AlphaFold structures now provide starting points for structure-based design even when experimental structures are unavailable.
Several pharmaceutical companies have publicly described using AlphaFold structures in active programs:
- Insilico Medicine used AlphaFold-predicted structures of TNIK (Traf2 and NCK-interacting kinase) as a starting point for an AI-designed drug candidate that entered Phase II clinical trials for idiopathic pulmonary fibrosis in 2023 — one of the first cases of a fully AI-designed molecule advancing to clinical trials.
- Isomorphic Labs (DeepMind’s drug discovery subsidiary) reported using AlphaFold 3 structures for initial target engagement modeling in multiple undisclosed programs as of 2024.
- Academic groups have used the AlphaFold database to identify previously unknown binding pockets in KRAS mutant proteins and other historically “undruggable” targets.
Important Limitations
AlphaFold predicts static equilibrium structures; it does not predict protein dynamics, conformational changes upon binding, allosteric effects, or the behavior of intrinsically disordered regions — all of which are clinically relevant. The pLDDT (per-residue local distance difference test) confidence score provides a useful guide to prediction reliability, but regions with low pLDDT scores in flexible loops or termini should not be used for drug design without experimental validation.
Additionally, AlphaFold 3’s performance on protein-ligand complexes, while significantly better than previous methods, still shows notable failures for novel scaffold types and for binding sites with large induced-fit effects. Medicinal chemists continue to require experimental validation of predicted binding poses before committing to lead optimization campaigns.
Key Takeaway
AlphaFold 2 and 3 have genuinely transformed the starting conditions for structure-based drug discovery by providing high-quality structural predictions for virtually the entire proteome and extending to ligand-protein complexes. They compress the early target characterization phase from months to days. They do not eliminate the need for experimental validation, and significant limitations remain for disordered proteins and large induced-fit binding events.
Sources
1. Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596(7873):583–589. doi:10.1038/s41586-021-03819-2
2. Abramson J, Adler J, Dunger J, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024;630(8016):493–500. doi:10.1038/s41586-024-07487-w
3. Varadi M, Anyango S, Deshpande M, et al. AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high-accuracy models. Nucleic Acids Research. 2022;50(D1):D439–D444.
Medical Disclaimer: This article is for informational purposes only and does not constitute medical advice. Always consult a qualified healthcare professional for medical decisions.