TechGlobal.news

AlphaFold 3’s Nature study is now shaping early drug screens, with clear limits

Priya Nair ·

Google DeepMind and Isomorphic Labs extended AI protein structure prediction to full biomolecular complexes in a 2024 Nature paper, and those methods are now appearing in early discovery workflows, according to companies that have disclosed collaborations since. On Wednesday in London, the relevance is less about a single new result and more about how a peer reviewed advance is crossing into practice as large drug makers seek to shorten hit identification.

The shift matters because several pharmaceutical companies, including Novartis and Eli Lilly, signed multi year discovery collaborations with Isomorphic Labs after the study, citing the potential to speed up target and ligand assessment. The methods promise to reduce the number of physical assays needed before progressing candidates, which affects cost and timelines.

What the Nature study showed

The Nature paper, led by John Jumper and colleagues at Google DeepMind and Isomorphic Labs, reported a diffusion based architecture that predicts three dimensional structures of assemblies that include proteins, nucleic acids, and small molecules. The model generated atomic coordinates for entire complexes, and it returned per residue and per atom confidence scores to indicate likely accuracy.

The team trained and evaluated on public structural datasets, including entries from the Protein Data Bank. The study showed higher accuracy for many protein complexes than earlier tools and could place small molecules in plausible binding poses without explicit docking. It did not estimate binding affinities and it did not model full biophysical dynamics in solvent.

What the researchers did not claim

The authors did not claim that AI predictions obviate crystallography, cryo electron microscopy, or binding assays. They cautioned that training data bias and limited coverage of conformational change remain open issues. They also noted that predicted poses require experimental validation, particularly for novel chemotypes and allosteric sites.

Why this is showing up in pipelines now

Large partners have integrated structure predictions as a filter in virtual screening and as a guide for medicinal chemistry around a target pocket, company scientists have said in public R and D updates. The approach can rank candidates for follow up, then labs confirm hits with biophysical assays such as surface plasmon resonance or NMR.

Regulators are also preparing for increased use of AI in upstream discovery. Agencies in the United States and Europe have issued general guidance on machine learning in drug development, which encourages documentation of training data and validation plans. That aligns with how the Nature paper reported benchmarks and confidence metrics.

For technology providers, the implications are immediate. Cloud platforms, chemistry software vendors, and CROs are building services that wrap AI structure prediction with docking, simulation, and automated assays. The competitive edge will come from linking credible in silico rankings to high throughput, real world validation at scale.