Conference Presentation, Keynote
Nobel Laureate John Jumper: AI is Revolutionizing Scientific Discovery
- The speaker, formerly a physicist who transitioned to computational biology and machine learning, now leads AI research for science at Google DeepMind, focusing on accelerating discovery and enabling new medical applications.
- The core problem addressed is the "folding problem": while DNA provides the linear sequence for proteins, understanding the resulting 3D structure is critical for determining function, disease mechanisms, and drug design, yet experimental determination remains exceptionally difficult.
- Experimental protein structure determination via X-ray crystallography requires forcing proteins to form crystals—a process taking over a year in many cases involving thousands of failed attempts—followed by synchrotron data collection.
- By 2024, the Protein Data Bank (PDB) contains approximately 200,000 known structures, increasing by roughly 12,000 annually, which is vastly insufficient compared to the billions of protein sequences being discovered (a gap of 3,000x).
- AlphaFold 2 was developed to predict protein structures from amino acid sequences, aiming to bridge the gap where the cost of experimental determination (~$100,000 and years of work) exceeds practical limits.
- The success of AlphaFold 2 relied on three components: data (200,000 structures), compute (~128 TPU v3 cores for two weeks), and research, with the speaker arguing that research ideas were the most differentiating factor.
- A controlled experiment by the Al Qureshi lab demonstrated that AlphaFold 2 trained on only 1% of available data outperformed AlphaFold 1 trained on 100% of data, indicating that the architectural research advancements were worth approximately 100x more than raw data volume.
- The system's accuracy was not derived from a single breakthrough like "equivariance," but from the aggregation of many "mid-scale ideas," including the Invariant Point Attention (IPA) module, which contributed significantly to the 30 GDT score improvement over AlphaFold 1.
- In the CASP14 blind assessment (a biennial competition where structures are unknown until submission), AlphaFold 2 achieved roughly one-third of the error rate of the next-best competing systems, validating its transformative accuracy.
- To maximize scientific utility, DeepMind released the AlphaFold code open-source and simultaneously launched a database of predictions containing 300,000 structures initially, expanding to over 200 million predictions covering nearly every sequenced organism.
- Scientific trust in the tool shifted rapidly after the database release; while specialists were convinced by CASP results, experimental biologists only accepted the tool after independently verifying predictions against their own unpublished structures via word-of-mouth.
- Researchers discovered emergent capabilities in AlphaFold 2, such as predicting protein-protein interactions simply by "prompt engineering" inputting two sequences together, a use case not originally targeted by the developers.
- The Jang Lab at MIT utilized AlphaFold predictions to re-engineer a "molecular syringe" (contractile injection system) for targeted drug delivery, replacing specific protein legs with designed elements to target new cell types within a mouse.
- The speaker asserts that AI for science acts as an amplifier for experimentalists, allowing the field of structural biology to move approximately 5-10% faster, which generates exponential downstream value for drug development, vaccine design, and understanding biological processes like fertilization.
- Future AI for science is expected to evolve from narrow, foundational models (like AlphaFold) toward broader systems that internalize scientific rules to solve a wider range of biological and physical problems.