Predictability failures in AlphaFold 3 drug discovery pipelines

The October trial failure with Novartis exposes specific vulnerabilities in AI-driven drug discovery. AlphaFold 3 predicts interactions involving proteins, DNA, RNA, ligands, and ions, yet the model produces overlapping atoms in certain predictions that can interfere with accurate docking studies for medicinal chemists. These errors occur in protein-nucleic acid complexes where both components exceed 100 nucleotides or 2,000 residues in total. The system also generates spurious alpha helices within disordered regions that show low confidence scores. Because accuracy varies across different molecule types and interface types, the correlation between model confidence and true prediction accuracy remains inconsistent. Even with the 50% improvement in accuracy for protein interactions compared to existing methods, the system struggles with specific biological complexities. In CASP14, AlphaFold achieved top ranking by producing high accuracy predictions, but researchers noted that accuracy fell short for a third of its predictions. AlphaFold 3 improves upon this by modeling chemical modifications, yet it still occasionally produces homomers where entire chains overlap. This instability persists despite the structural accuracy improvements seen since the 2021 methods paper was published. Earlier versions like AlphaFold 2 achieved a median error of less than 1 Angstrom, yet the current model still produces errors in complex interactions.

Isomorphic Labs and the competition for talent

Isomorphic Labs creates a competitive environment for DeepMind through its independent drug design engine, IsoDDE. The massive influx of capital into Isomorphic Labs through a $600 million funding round led by Thrive Capital in March 2025 creates a significant resource gap between it and the research teams at DeepMind. This financial strength challenges DeepMind’s ability to maintain its research dominance in the medical AI sector. Following the April 2023 merger of Google Brain and DeepMind, which responded to competition from OpenAI’s ChatGPT, the combined entity aims to lead generative AI, but internal competition persists. Google Brain focused on scale and produced models like PaLM and BERT, while DeepMind emphasized scientific questions and reinforcement learning. DeepMind originally entered the market after Google acquired it in 2014 for between $400 million and $650 million. Top researchers moved between Google and competitors during 2025 and 2026, a movement that complicates long-term project stability. You should observe how these talent shifts impact the continuity of the AlphaFold development cycle. Demis Hassabis and John Jumper, who received the Nobel Prize in Chemistry in 2024, must manage these shifting loyalties to maintain research cohesion. Will the separation of commercial drug design from core AI research eventually erode the collaborative benefits of the original DeepMind mission?

Technical constraints for enterprise users

Commercial partnerships face strict technical boundaries when using the AlphaFold Server for large-scale projects. The platform restricts users to 30 jobs per day and imposes a 5,000 token limit on the total size of any single job. Every protein or nucleotide chain requires a minimum of 4 amino acids or nucleotides to generate a valid prediction. Researchers using the AlphaFold Protein Structure Database can access over 200 million entries, which provide coverage of UniProt. This database includes predictions for the 30 WHO priority global health proteomes to support research on pathogens of major global public health importance. Access to this data follows a CC-BY-4.0 licence, allowing for academic and commercial use. The server provides confidence metrics, including pLDDT and Predicted Aligned Error (PAE), though users must use Python or another language to analyze PAE JSON files. The latest database release integrates AI-derived residue-level annotations from scientific literature to provide biological context. This allows users to navigate from residues in AlphaFold predictions to publications describing function or interactions. For the most reliable results, users must ensure that input queries contain at least 20 amino acids during sequence searches. Users must also cite the Abramson et al. (2024) paper when using predictions from the AlphaFold Server.