AlphaFold 3: Google DeepMind’s latest AI tech in drug discovery

AlphaFold 3: Google DeepMind’s latest AI tech in drug discovery

By Jorge Hurtado

Google DeepMind has introduced its latest AI tech, AlphaFold 3. It is an AI model capable of accurately predicting the 3D folded structure of proteins based only on their amino acid sequences.  AlphaFold can predict how proteins interact with one another and with other molecules, including DNA, RNA, and other small biomolecules.

What makes AlphaFold 3 unique?

AlphaFold 3 can accurately predict the intricate structures and interactions of biomolecules. This is a big step forward from AlphaFold 1 and AlphaFold 2, helping to understand proteins much better. 

AlphaFold 3 figures out how the protein interacts with other small molecules (ligands), DNA, RNA, and other proteins with remarkable accuracy rates compared to any other existing AI-developed model. Its true value lies in its ability to translate the shapes of these molecules into useful information about how they work and behave.

Google DeepMind has made AlphaFold Server openly accessible to make it easier for researchers worldwide to use AlphaFold 3. This approach is expected to drive breakthroughs in drug discovery, biotechnology, genomics, and our fundamental understanding of biological systems. However, unlike its previous models, DeepMind has not released the downloadable code of AlphaFold 3.

What are the strategic implications of AlphaFold 3, and can it influence the emergence of similar tools? Gain insights from our Technical Director, Dr. João Guerreiro.

The role of protein folding:

Proteins are made of long chains of amino acids, which fold and twist to form 3D structures. This process is known as protein folding. Each protein follows a unique pattern of folding. So, a “3D protein structure” refers to the specific shapes that proteins adopt through folding.

A protein’s 3D structure will determine how a protein will interact with other molecules and perform its function. If a protein folds incorrectly or adopts an abnormal shape, it will not function properly, leading to degenerative diseases.

A protein’s 3D structure is key to almost everything:

A protein’s 3D structure is the key to deciphering its molecular function and unraveling disease mechanisms. It also guides drug discovery efforts, explores evolutionary relationships, and enables protein engineering applications. Without protein’s structural information, understanding proteins and their roles in biological systems would be limited.

AlphaFold 3 can identify protein’s 3D structures without the meticulous trial and error that entails years of painstaking labor and costly financial investment. 

Advancing protein interaction modeling: 

AlphaFold 3’s capability to predict protein-ligand interactions is essential for drug discovery. It enables precise modeling of how potential drug molecules bind to their target proteins. This facilitates the identification and optimization of therapeutic candidates.

Understanding protein-nucleic acid interactions plays a relevant role in how genes are regulated. Accurate modeling of these interactions can provide essential insights into fundamental biological mechanisms and disease pathways.

AlphaFold 3 has significantly enhanced prediction accuracy for antibody-antigen interactions. Learning about these interactions is essential for applications in immunology and therapeutic antibody development. It helps to understand the exact binding between antibodies and antigens.

How AlphaFold 3 nails accurate predictions:

AlphaFold’s approach to predicting protein structures blends two key methods: bioinformatics and physics. The AI models use a physical and geometric inductive bias to learn from protein structure data without relying heavily on handcrafted features. The approach allows AlphaFold to efficiently learn from limited data while accommodating the complexity and diversity of structural information. 

AlphaFold can handle challenging cases like proteins with missing physical context or structures. It can even produce accurate models with underspecified structural conditions that were present in public repositories of protein data.

Database growth promotes innovations:

The AlphaFold Database has expanded its collection of predicted protein structures by over 714 times. It now includes over 214 million structures, up from the initial 300,000 in 2021.

The impact is enormous. AlphaFold can help find new drugs, understand diseases better, and improve biotechnology. Also, it can aid in making vaccines, fighting antibiotic resistance, and studying extinct species.

Limitations of the AI model:

In only three years, AlphaFold has transformed protein structure predictions with accuracy. However, the AI model has yet to develop its total capacity. The main reason is that the availability of training data limits its learning capabilities. AlphaFold used publicly available datasets that were determined by researchers across the globe, such as the Protein Data Bank and UniProt

AlphaFold 3 showcases a remarkable accuracy of up to 50% in predicting biomolecular structures and interactions. It achieves approximately 76% accuracy in predicting protein-ligand interactions, 65% for protein-DNA interactions, and 62% for protein-protein interactions. However, while AlphaFold’s advancements are significant, it is key to exercise caution and not solely rely on its predictions. Ongoing experimental validation is still indispensable to ensure reliability.

AlphaFold’s impact on industries:

Undeniably, AlphaFold’s impact is already palpable across industries, prompting consideration of its potential applications.

Drug development and discovery

AlphaFold 3 uses precise protein structure predictions to expedite the identification of new drug targets and enhance the design of more effective therapeutics.

For example, AlphaFold significantly contributed to comprehending the structure of a critical protein for malaria vaccine development. Using the latest technology alone (X-ray crystallography and cryo-electron microscopy) offered low-resolution images. Therefore, the 3D structural models of the malaria protein were imprecise and incomplete. 

Combining the technology with AlphaFold predictive models, researchers at the University of Oxford were able to identify which was the critical malaria protein. They also identified key components for the vaccine. The vaccine quickly advanced from basic research to clinical development. 

AlphaFold’s structural prediction for a spike protein of a common cold virus accurately depicted how the virus protein interacts with antibodies and simple sugars. This provided a more precise match to the true virus protein structure. The discovery improved the understanding of the interactions between the immune system and the behavior of coronaviruses.

The AlphaFold AI model could greatly impact the pharmaceutical landscape by aiding drug discovery. It enhances drug development processes and leads to potential new treatments for various diseases.

These are examples of how the AlphaFold AI predictive model could impact the pharmaceutical landscape by improving collaborative drug discovery. It enhances drug development processes and leads to potential new treatments for various diseases.

Market disruption

While AlphaFold 3 broke with the open-source nature of the previous 2 models, it still offers free access for non-commercial use. DeepMind’s Isomorphic Labs is looking to handle the commercial licensing model for interested parties such as larger pharmaceutical companies. 

Before AlphaFold was released, private companies led most AI drug discovery efforts. This meant that many advanced AI algorithms, tools, and databases were patented or protected, limiting access to broader drug discovery research communities.

Increased competition

The accessibility of AlphaFold’s technology has attracted new entrants who have taken advantage of protein knowledge and applications. The increased competition will require all companies to set strategies to keep innovating in the long term.

Broader industry influence

AlphaFold has the potential to tackle other world problems, such as plastic pollutionIt has been used to design new types of enzymes that can more efficiently break down plastic waste so that they can be 100% recycled.

Other exciting business applications can be related to the production of biofuels and improving food production. The AI model can help to make their processes more efficient and economically viable. 

For example, AlphaFold 3 provided a structural prediction of a molecular complex. The AI model identified an enzyme protein, an ion, and simple sugars alongside the protein’s structure. The enzyme originates from a soil-borne fungus known for causing damage to crops. Interactions between this enzyme and plant cells could potentially aid in cultivating healthier crops.

What’s next?

As more data becomes available and AlphaFold AI models can gradually learn more, the continuous development of protein structures will pose opportunities and challenges for related industries and verticals. While AlphaFold presents immense potential for progress and innovation, it may also pose difficulties for traditional companies to respond as required. 

AlphaFold also brings serious concerns about the impact on funding, talent, and resources for experimental structural biology techniques if researchers overly rely on the AI model’s predictions without validating them experimentally.

By providing strategic insights and evaluating the implications of AlphaFold, PreScouter can help organizations bridge the gap between new developments and practical implementation strategies. This can enable companies to exploit the potential benefits of AlphaFold while remaining adaptable and competitive in a dynamic market.

Interested in generative AI for your business, but not sure where to start? Schedule a complimentary brainstorming session with our AI experts here or email us at

Never miss an insight

Get insights delivered right to your inbox

More of Our Insights & Work

Never miss an insight

Get insights delivered right to your inbox

You have successfully subscribed to our newsletter.

Too many subscribe attempts for this email address.