The presentation featured a foundational primer on protein folding.
In the first keynote "Programmable Biology: Generative AI for Molecular Design" from @saakohl
starting by showing us the exponential we're getting accustomed to in AI/ML + giving us a primer on "what is a protein, what should we care, what's up with this folding stuff, ..."
Users are excited about Saakohl's keynote on generative AI for molecular design at CVPR 2026 because of its optimistic scaling laws and bold predictions for self-improving biology systems.
Drugs take incredibly long time to discover, trial, approve: 10 years to patient, $2B per drug, and 9/10 drugs fail in the end #cvpr2026
conclusions and (extremely bold) predictions:
Self-improving design -> Whole cell prediction / sim. -> Personalized design /genome cond. -> Predicting human drug response
-> *All diseases solved* (by 2036!?)
wild confidence, but if he is right... what a world @CVPR #CVPR2026
The big hope put simply #cvpr2026
How does AlphaFold work?
"transformerish" stuff + explicit pairwise modeling via outer product over sequence to give the model a "scratch pad" to allow "reasoning" over pairwise interactions (when you ablate this it gets much worse). also some templates (not as important)
Open problems
Key Point / Analogy: changing one "pixel" breaks the whole structure
AlphaFold was a "unilateral" and discontinuous advance in the field
@Latent_Labs Of course, they have an agent Latent-Y that can do all sorts of tasks: antibody design and interaction with Latent-X2, fetching info from literature. It gives up to 50x speedups #cvpr2026
So with structure prediction "solved" how do we design drugs?
"Predictors Enable Generative Models"
Protein Data Bank -> AlphaFold -> protein generative models is analogous to ImageNet -> AlexNet -> GANs (and then diffusion)
train a denoising diffusion model treating atoms like a point cloud and conditioning on the binding target
benchmarks and simulation are not enough - wet-lab verification is essential
peptides and mini-binders
Models are able to generate molecules that bind to unseen targets
the generated molecules also have nice properties
Pass@10 matches or exceed brute force screening over trillion scale molecule database.
also works for ion channels (classically hard to drug); in the past people started with eg tarantula venom
can also give the generative model to a specialized agent as a tool call
The AlphaFold2 architecture in a nutshell: a beautiful mix of AI and domain experience poured there, eg. the evoformer block uses a tweaked attention to take into consideration the structure and interactions in proteins #cvpr2026
Protein mutation analogy to computer vision: as if a single wrong pixel could break an image #cvpr2026
The decades-long protein folding problem and how crazy the AlphaFold series of models performed #cvpr2026
The presentation featured a foundational primer on protein folding.
In the first keynote "Programmable Biology: Generative AI for Molecular Design" from @saakohl
starting by showing us the exponential we're getting accustomed to in AI/ML + giving us a primer on "what is a protein, what should we care, what's up with this folding stuff, ..."
