DeepMind’s New AI Just Cracked The Code Of Life

Summary
Dr. Károly Zsolnai-Fehér from Two Minute Papers interviews Dr. Pushmeet Kohli, Vice President of Research at Google DeepMind, regarding DeepMind’s breakthrough genomics foundation model, AlphaGenome, and the AlphaGenome Atlas. They discuss how the model decodes non-coding DNA and gene regulation, precomputes the impact of all possible single-letter genetic variations, and paves the way toward whole-cell simulation.

What is shown

• Interview & Technical Discussion: A studio interview between Dr. Károly Zsolnai-Fehér and Dr. Pushmeet Kohli exploring AlphaGenome’s architecture, mechanisms, and real-world implications [00:00–32:42].
• Explanation of Genome Decoding: Pushmeet explains the difference between the 20,000 protein-coding genes and the non-coding regulatory sequences governing expression [07:00–08:05].
• Cross-Species Generalization: Discussion of training on both human and mouse genomes to force the network to learn generalized biological mechanisms rather than memorizing human data [08:08–10:22].
• Resolution & Context Window: Pushmeet highlights advances over previous models like Enformer, scaling to a 1-million-base-pair context window while retaining single-base-pair resolution [10:23–12:43].
• AlphaGenome Atlas: Pushmeet explains how AlphaGenome precomputed the effects of all ~9 billion possible single nucleotide variants into a multi-petabyte database [15:08–17:36].
• AVIS Metric & Biophysical Validation: Explanation of the AlphaGenome Variant Impact Score (AVIS) and verifying mechanistic, biophysical predictions (such as splicing and expression changes) in laboratory wet-lab trials [18:15–23:58].
• Future Goals & 1,000,000x Speedup: Discussion of the next frontiers, including full virtual cell modeling, long-term climate predictions, and superconductor discovery [29:00–32:42].
• Sponsor Segment: Károly demonstrates running code, model inference, and generating AI outputs on Lambda cloud infrastructure [32:43–33:16].

Claims & numbers

• The presenter notes AlphaGenome matched or beat the strongest comparison baseline on 25 out of 26 variant effect benchmarks [12:44].
• Pushmeet Kohli states the human genome contains approximately 3 billion base pairs, encoding about 20,000 proteins [01:21, 07:14].
• Pushmeet Kohli explains that AlphaGenome Atlas precomputes all ~9 billion possible single nucleotide variations across the human genome (three possible mutations per position) [15:59].
• Pushmeet Kohli states AlphaGenome Atlas is 30 times larger than the AlphaFold database, requiring a petabyte-scale dataset [00:30, 17:22, 16:33].
• Pushmeet Kohli mentions the AlphaFold database contains 250 million protein structures and has been used by more than 4 million researchers across 190 countries [17:03, 17:14].
• Pushmeet Kohli states AlphaGenome operates across a 1-million-base-pair context window at single-base-pair resolution [12:09, 17:41].
• Károly notes the correlation coefficient ($\rho$) between predicted and observed effect sizes for mutations affecting gene activity is approximately 0.5 [17:55].
• Pushmeet Kohli confirms AlphaGenome and its model weights are available free for non-commercial and academic research [06:23].
• Károly states that AI training and inference have achieved an estimated 1,000,000× effective speedup over the previous 10 years [29:00].

Notable quotes

• Pushmeet Kohli [00:30]: “AlphaGenome Atlas is 30 times the size of AlphaFold database.”
• Dr. Károly Zsolnai-Fehér [12:44]: “I saw that it is so dominant, it matched or beat the strongest comparison model in 25 out of 26 variant effect benchmarks.”
• Pushmeet Kohli [29:33]: “Simulating even a cell, creating a virtual cell... if you can completely sort of really understand the workings of a cell, that will have amazing consequences for the fight against diseases like cancer...”

Assessment
This video is an in-depth interview and review between an independent science communicator and a principal research executive at Google DeepMind discussing a published research milestone. No synthetic demos or cherry-picked benchmark screens are presented on screen; the claims are delivered through direct discussion of published paper results and confirmed metrics.

Described by gemini-3.8-flash on 2026-10-08 from the video’s audio and frames.