AI News

Google DeepMind and AlphaFold 3: AI revolutionizes biology

AlphaFold 3 is no longer limited to proteins: the model now predicts interactions between DNA, RNA, ligands and molecules. A turning point for pharma and biotechs.

5 min read
GoogleDeepMindScienceInnovationResearch
⚡ The news in 30 seconds

AlphaFold 3 models all the molecular interactions of living organisms

Google DeepMind released AlphaFold 3 in May 2024, a model capable of predicting the structure and interactions of all biological molecules: proteins, DNA, RNA, ligands and small molecules. Accuracy reaches 76% on protein-ligand complexes, a 50% leap compared to previous methods. The pharmaceutical sector sees it as a major accelerator of drug discovery.

For life sciences companies, AlphaFold 3 can shorten the upstream phases of R&D by 6 to 18 months.

Opportunities for businesses

AlphaFold 3's ability to model molecular interactions creates concrete opportunities well beyond fundamental research. Three sectors are particularly concerned.

💊

Pharma & drug discovery

Identify drug candidates in a few days instead of several months. Sanofi estimates a 40% gain on the cost of preclinical phases thanks to in silico modeling.

🧬

Biotech & diagnostics

Design custom antibodies and biosensors. Startups use AlphaFold 3 to accelerate the design of diagnostic tools by 3 to 5 times.

🌱

Food & agriculture

Improve industrial enzymes for food fermentation and preservation. BASF announced a 25% reduction in enzyme optimization cycles.

Risks to anticipate

⚠️

Technological dependence

AlphaFold 3 is controlled by Google via Isomorphic Labs. The predictions are reliable but opaque: it is impossible to fully audit the model. Companies must plan for a systematic experimental validation strategy.

💡

Intellectual property

AI-generated structures raise questions of patentability. The EPO (European Patent Office) has not yet ruled clearly on the protection of molecules designed by AI. Consult an IP advisor before filing a patent based on these predictions.

Our recommendations

1

Assess your internal use cases

Identify the R&D projects where molecular modeling represents a bottleneck. Prioritize cases where the time savings exceed 3 months.

2

Launch a POC on an existing pipeline

Integrate AlphaFold 3 into an ongoing project to compare the results with your current methods. Indicative budget: 15,000 to 30,000 EUR for a 6-week POC.

3

Train your teams and secure your IP

Organize training for your researchers on using the AlphaFold server. In parallel, have your patent filing processes audited by a firm specialized in AI-IP.

Key takeaways

Accuracy
76% on protein-ligand complexes
R&D gain
6 to 18 months on upstream phases
POC budget
15,000 - 30,000 EUR
Key sectors
Pharma, biotech, agri-food

Frequently asked questions

Is AlphaFold 3 accessible to life sciences SMBs?

Yes. Google offers free access via the AlphaFold server for academic research. For commercial use, a license via Isomorphic Labs is required, but the costs remain lower than those of a conventional screening campaign.

What is the difference between AlphaFold 2 and AlphaFold 3?

AlphaFold 2 predicted only the 3D structure of proteins. AlphaFold 3 also models the interactions between proteins, DNA, RNA, ligands and small molecules, which paves the way for AI-assisted drug design.

Do you need bioinformatics skills to use AlphaFold 3?

The web interface is accessible to researchers without technical expertise. However, integrating it into a drug discovery pipeline requires bioinformatics skills or the support of a specialized partner.

AlphaFold 3 tech sheet

AlphaFold 3

Molecular structure prediction model based on a diffusion architecture. Trained on the PDB and enriched with proprietary data.

Pricing

Academic research Free
Commercial license On quote (Isomorphic Labs)

Comparison

Criterion AlphaFold 3 RoseTTAFold All-Atom ESMFold
Multi-molecule interactions Yes Partial No
Accuracy (LDDT) 0.89 0.82 0.73
Free cloud access Yes Yes Yes
On-premise deployment Restricted Yes Yes

Related articles