Get your free personalized podcast brief

We scan new podcasts and send you the top 5 insights daily.

Younger says model makers evaluate on unique bespoke targets, so pharma buyers cannot tell which model is best. He is working with companies on a benchmark that standardizes target difficulty and characterizes tens of thousands of designs, hoping to get true apples-to-apples comparisons with tighter error bars.

Related Insights

The lack of comparable developability data is a major bottleneck. Natural Antibody's CEO suggests a 'walk before you can run' approach: instead of accounting for all variables, the industry should create a foundational dataset under a single condition. This focused dataset has proven transferable predictive power.

To demonstrate industrial-scale viability, Chai Discovery tested its antibody design model on 50 different targets. This focus on generalization, far beyond the typical 2-4 targets shown in academic research, is crucial for proving a model is not a statistical anomaly and is ready for real-world application.

The community standard of two-angstrom accuracy for protein-ligand predictions is insufficient. At that resolution, critical details like an aromatic ring's orientation can be wrong, rendering the model's output misleading for drug design. Genesis argues one-angstrom accuracy is the minimum for practical utility.

To avoid overfitting and prove true generalization, Bolts validates its protein design models by testing them across a wide array of targets from over 25 external academic and industry labs. This diverse, real-world testing is the ultimate benchmark of a model's utility in drug discovery.

Genesis's focus on sub-one-angstrom accuracy came from direct experience. When applying models to active drug discovery programs, it became 'pretty obvious' that the standard two-angstrom benchmark was inadequate. This highlights the gap between academic benchmarks and real-world utility.

Current AI for protein engineering relies on small public datasets like the PDB (~10,000 structures), causing models to "hallucinate" or default to known examples. This data bottleneck, orders of magnitude smaller than data used for LLMs, hinders the development of novel therapeutics.

Younger notes the PDB shows how proteins physically interact but not how strongly, and it has almost no examples of antibodies one or two mutations apart with measured affinities. Zero-shot optimization instead needs local affinity landscapes built from many single-point mutation variants across many targets.

When all pharma companies use similar AI models on similar data, competitive moats vanish. The next edge will come from creating superior training data. This means moving beyond raw data to datasets enriched with physical insights, such as a database of 'complexity hotspots' for all known proteins, teaching AI the underlying dynamics.

David Younger says his company's model training showed that solving truly hard protein design problems will require much more data than the field expects. He also warns public binding and affinity data are often not interoperable, lack proper controls, and lack annotations, making them hard to reuse.

Younger says the Protein Data Bank today holds about 10,000 antibody-antigen structures and adds roughly 1,000 yearly. Their internal scaling work suggests good generalizable de novo design models may need hundreds of thousands to millions of structures, so waiting on PDB growth is far too slow.