AWS: multimodal biological foundation models accelerate drug discovery by 50 percent and diagnostics by 90 percent

Amazon Web Services published a detailed overview on April 23, 2026, on applying multimodal biological foundation models in therapeutics and patient care. The results point to significant improvements in both the drug discovery phase and clinical diagnostics.

What are multimodal biological foundation models?

Foundation models in biology are large neural models pretrained on vast amounts of biological data — analogous to the way GPT or Claude are pretrained on text. AlphaFold and RoseTTAFold paved the way for protein structures, and the new generation of models goes a step further.

Multimodality in this context means that a single model simultaneously processes multiple types of biomedical data: DNA and RNA sequences from genomics, medical images (MRI, CT, pathology slides), and clinical data from electronic health records (EHR). Instead of separate models per modality, the model learns a shared representation that connects all sources.

What numbers demonstrate the improvements?

AWS cites three key results from real-world applications. The first is a 4–7% improvement in AUC (area under curve) diagnostic accuracy when multimodal models are compared with unimodal ones on the same datasets.

The second result is up to 90% time savings in medical image analysis, where pathologists and radiologists receive prioritized sections to review instead of manually scanning entire images. The third is a reduction in costs of up to 50% in the early stage of drug development, where the model helps identify the most promising candidate molecules.

What does this mean for the pharmaceutical industry?

Traditional drug development takes 10–15 years and costs over a billion dollars, with most candidates failing in late-stage clinical trials. If foundation models can eliminate poor candidates early and identify promising drug targets, average development costs and timelines drop significantly.

Concrete applications include identifying new drug targets from genomic data, automated anomaly detection in pathology images, and linking genetic variants to clinical outcomes. Unlike AlphaFold, which solves one problem (protein structure), the multimodal approach targets the entire pipeline from discovery to application.