DeepMind’s AlphaFold 2 can now predict structures of diverse molecules beyond proteins

DeepMind, the artificial intelligence company behind AlphaGo and AlphaZero, has announced a significant update to its protein structure prediction system, AlphaFold 2. The new version of the system can now generate accurate 3D models of diverse molecules beyond proteins, such as DNA, RNA, and small organic compounds. This breakthrough could have profound implications for various fields of science, such as disease research, drug development, synthetic biology, and more.

AlphaFold 2 is a deep learning system that uses neural networks to predict the 3D shape of a protein from its amino acid sequence. The system was first introduced in 2018 and achieved remarkable results in the Critical Assessment of Protein Structure Prediction (CASP) competition, a biennial challenge for researchers to test the accuracy of their predictions against real experimental data. In 2020, AlphaFold 2 improved its performance even further, reaching near-atomic accuracy for most of the proteins in the CASP14 test set.

However, proteins are not the only molecules that have complex 3D structures and play important roles in biological processes. Other biomolecules, such as DNA and RNA, also have intricate shapes that determine their functions and interactions. Moreover, small organic compounds, such as drugs and metabolites, can bind to proteins and modulate their activities. Therefore, being able to predict the structures of these molecules as well as proteins could greatly enhance our understanding of how life works at the molecular level.

In a blog post published on October 31, 2023, DeepMind revealed that they have extended AlphaFold 2 to be able to handle a wider range of molecular types. They achieved this by using a more general representation of atoms and bonds, and by incorporating additional information from databases and literature. They also trained the system on a larger and more diverse dataset of over 400,000 known structures from the Protein Data Bank (PDB), which contains not only proteins but also nucleic acids and small molecules.

The results are impressive: AlphaFold 2 can now generate predictions for nearly all molecules in the PDB, frequently reaching atomic accuracy. The system can also handle multimeric complexes, such as protein-DNA or protein-drug interactions, and predict how they change their shapes upon binding. Furthermore, AlphaFold 2 can model novel molecules that are not in the PDB, such as synthetic peptides or drug candidates.

Google DeepMind AI just revealed a significant AlphaFold update to revolutionize future drug development.

DeepMind claims that this is the first time that a single system can predict structures of diverse molecules with such high accuracy and speed. They also state that they are committed to making AlphaFold 2 accessible to the scientific community and plan to release the code and models in the near future. They hope that their system will accelerate scientific discovery and enable new applications in various domains.

Some examples of potential applications include:

• Disease research: By predicting the structures of disease-related proteins and their interactions with DNA, RNA, or drugs, AlphaFold 2 could help identify new targets for therapeutic interventions or design novel molecules with desired effects.
• Drug development: By predicting the structures of drug candidates and their binding modes to proteins, AlphaFold 2 could help optimize drug properties or discover new drug leads.
• Synthetic biology: By predicting the structures of synthetic peptides or proteins and their interactions with other molecules, AlphaFold 2 could help engineer new biological systems or functions.
• Biochemistry: By predicting the structures of enzymes or metabolites and their catalytic mechanisms, AlphaFold 2 could help elucidate biochemical pathways or reactions.

The update on AlphaFold 2 is a remarkable achievement that demonstrates the power and potential of artificial intelligence for advancing science. It also opens up new possibilities for exploring the molecular world and understanding its complexity. As DeepMind’s CEO Demis Hassabis said, “We believe this is just the beginning.”