[M2L 2025] 5.1 AI in Biology then and now - Kathryn Tunyasuvunakool

Value of the Information & Strength of the Argument

The talk provides valuable insights into the practical development of AI for science, particularly the AlphaFold series. The speaker’s firsthand experience lends credibility, and she effectively argues that domain-specific design choices, rather than a single breakthrough, were key to AlphaFold 2’s success. She also addresses the challenge of limited training data and the importance of confidence metrics for scientific use. The argumentation is coherent and well-structured, though it is primarily anecdotal and lacks a systematic comparison with other approaches.

Scientific Rigor, Source Quality, Title Accuracy

The speaker references the CASP competitions and the PDB as key resources, and mentions the AlphaFold papers and the Nobel Prize, but does not provide specific citations. The talk is based on her own work and experiences, which is appropriate for an expert opinion. The title accurately reflects the content, which traces the evolution of AI in biology. The talk is not a rigorous historical review, as she acknowledges, but it is scientifically sound and well-informed.

The title accurately reflects the content, which traces the evolution of AI in biology from early methods to current LLM-based approaches, using AlphaFold as a case study.

8/10

The speaker is a researcher at Google DeepMind with direct involvement in AlphaFold development, providing an insider perspective. The talk is based on personal experience and published work, with references to CASP and the PDB. However, it is not a systematic review and lacks external citations beyond the speaker's own projects.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 02:00Introduction: speaker's background and the context of AI for science at DeepMind.
• 05:00Discussion of the mid-2010s: machine learning rarely used in computational biology.
• 10:00Explanation of protein structure prediction problem and the importance of 3D structure.
• 15:00AlphaFold 1: using convolutional networks on MSAs and predicting distance maps.
• 20:00CASP evaluation and AlphaFold 1's performance.
• 25:00AlphaFold 2: end-to-end architecture, Evoformer, and structure module.
• 30:00AlphaFold 2's success and its impact on the field.
• 35:00Deployment of AlphaFold 2: open sourcing, AlphaFold DB, and confidence metrics.
• 40:00AlphaFold 3 and the shift to diffusion models.
• 45:00Broader context: LLMs in science and future directions.

• AlphaFold 2 paper — The speaker discusses the development and results of AlphaFold 2, referencing the paper.
• AlphaFold 3 paper — The speaker mentions the release of AlphaFold 3 and its diffusion module.
• CASP — The speaker references CASP as the evaluation for protein structure prediction.
• PDB — The speaker mentions the Protein Data Bank as a key resource for training data.

• AlphaFold 2 paper — The speaker's description of AlphaFold 2 aligns with the published paper.
• AlphaFold 3 paper — The speaker's description of AlphaFold 3 aligns with the published paper.