--- title: Why AlphaFold Didn't Solve Protein Folding — Pushmeet... description: A panel discussion between Google DeepMind's Pushmeet Kohli and Biohub's Sal Candido explores why AlphaFold's breakthrough, despite being celebrated as... canonical: https://daily.dev/posts/why-alphafold-didn-t-solve-protein-folding-pushmeet-kohli-google-deepmind-sal-candido-biohub-4s7gtspvz twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub | daily.dev og:description: A panel discussion between Google DeepMind's Pushmeet Kohli and Biohub's Sal Candido explores why AlphaFold's breakthrough, despite being celebrated as... og:url: https://daily.dev/posts/why-alphafold-didn-t-solve-protein-folding-pushmeet-kohli-google-deepmind-sal-candido-biohub-4s7gtspvz og:image: https://api.daily.dev/og/posts/4S7GtSPVZ.png og:image:alt: Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub **[Latent Space](https://daily.dev/sources/latentspace)** · 22 min read · 2 upvotes · 0 comments ## Summary A panel discussion between Google DeepMind's Pushmeet Kohli and Biohub's Sal Candido explores why AlphaFold's breakthrough, despite being celebrated as 'solving' protein folding, left major problems unsolved: protein dynamics, disorder, function, and design. They discuss the 'bitter lesson' as it applies to biological data (scaling only works once you find the right scaling law, and good data matters more than sheer volume), the tradeoffs between handcrafted scientific inductive biases and general-purpose scaling, the potential of cryo-EM micrographs to unlock richer structural information, the need to move from modeling individual proteins to whole biological systems ('building a virtual cell'), and questions about interpretability versus calibrated trustworthiness in models like AlphaFold 2. They close by discussing timelines for AI-driven breakthroughs in drug discovery and Biohub's mission of pursuing 10x rather than incremental improvements in curing disease. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Why do experts say AlphaFold didn't actually solve the protein folding problem? AlphaFold 2 replicates static structures deposited in the Protein Data Bank, but proteins are not rigid blocks; they are disordered, context-dependent, and their true distribution of shapes remains unknown. Protein dynamics, function, and design are still wide-open problems five years after AlphaFold 2's announcement, according to its co-creators at Google DeepMind. _Developers tracking how far AI models for biology have really come can follow this debate on daily.dev._ ### How did training on low-quality metagenomic sequences affect protein language model performance? Training a protein language model on metagenomic sequences, even though much of that data isn't even a complete real protein, improved the model's performance at designing real working proteins and understanding known proteins. This illustrates that abundant lower-quality data can still raise capability, though it can also tempt teams to scale up easy-to-generate data rather than the data a problem actually needs. _daily.dev helps developers weighing data quality versus data volume trade-offs follow real-world ML lessons._ ### What made AlphaFold 2's confidence score (pLDDT) important for trusting its predictions? Calibration of AlphaFold 2's uncertainty measure (pLDDT) was critical because a structure prediction with a GDT score around 90 is only useful if its confidence score accurately reflects correctness; an uncalibrated but highly confident wrong answer could mislead researchers into wasting a year of work. Trustworthy calibration mattered more than full mechanistic interpretability for safe practical use. _daily.dev keeps builders of AI-reliant workflows informed on why model calibration matters as much as accuracy._ ## Similar posts on daily.dev - [What’s next for AlphaFold: A conversation with a Google DeepMind Nobel laureate](https://daily.dev/posts/what-s-next-for-alphafold-a-conversation-with-a-google-deepmind-nobel-laureate-bka0wcmlt) · MIT Technology Review · 0 upvotes · 0 comments - [AlphaFold: Five years of impact](https://daily.dev/posts/alphafold-five-years-of-impact-tqbavrueh) · DeepMind · 0 upvotes · 0 comments - [Google shuts down Nobel Prize winning AlphaFold](https://daily.dev/posts/google-shuts-down-nobel-prize-winning-alphafold-h2gmeztef) · Hacker News · 5 upvotes · 2 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#machine-learning](https://daily.dev/tags/machine-learning) [View this post on daily.dev](https://daily.dev/posts/why-alphafold-didn-t-solve-protein-folding-pushmeet-kohli-google-deepmind-sal-candido-biohub-4s7gtspvz) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub","url":"https://daily.dev/posts/why-alphafold-didn-t-solve-protein-folding-pushmeet-kohli-google-deepmind-sal-candido-biohub-4s7gtspvz","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/why-alphafold-didn-t-solve-protein-folding-pushmeet-kohli-google-deepmind-sal-candido-biohub-4s7gtspvz"},"datePublished":"2026-10-10T00:37:31.265Z","dateModified":"2026-10-10T00:38:00.998Z","description":"A panel discussion between Google DeepMind's Pushmeet Kohli and Biohub's Sal Candido explores why AlphaFold's breakthrough, despite being celebrated as...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ccd2e1bb0c551a04f36af302d01e0b45?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ccd2e1bb0c551a04f36af302d01e0b45?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Latent Space","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Latent Space","logo":"https://media.daily.dev/image/upload/s--DnJ9laFj--/f_auto,q_auto/v1780213196/logos/latentspace?_a=BAMAMiWQ0","url":"https://daily.dev/sources/latentspace"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/why-alphafold-didn-t-solve-protein-folding-pushmeet-kohli-google-deepmind-sal-candido-biohub-4s7gtspvz","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,machine-learning","timeRequired":"PT22M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Latent Space","item":"https://daily.dev/sources/latentspace"},{"@type":"ListItem","position":3,"name":"Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/why-alphafold-didn-t-solve-protein-folding-pushmeet-kohli-google-deepmind-sal-candido-biohub-4s7gtspvz#faq","mainEntity":[{"@type":"Question","name":"Why do experts say AlphaFold didn't actually solve the protein folding problem?","acceptedAnswer":{"@type":"Answer","text":"AlphaFold 2 replicates static structures deposited in the Protein Data Bank, but proteins are not rigid blocks; they are disordered, context-dependent, and their true distribution of shapes remains unknown. Protein dynamics, function, and design are still wide-open problems five years after AlphaFold 2's announcement, according to its co-creators at Google DeepMind. Developers tracking how far AI models for biology have really come can follow this debate on daily.dev."}},{"@type":"Question","name":"How did training on low-quality metagenomic sequences affect protein language model performance?","acceptedAnswer":{"@type":"Answer","text":"Training a protein language model on metagenomic sequences, even though much of that data isn't even a complete real protein, improved the model's performance at designing real working proteins and understanding known proteins. This illustrates that abundant lower-quality data can still raise capability, though it can also tempt teams to scale up easy-to-generate data rather than the data a problem actually needs. daily.dev helps developers weighing data quality versus data volume trade-offs follow real-world ML lessons."}},{"@type":"Question","name":"What made AlphaFold 2's confidence score (pLDDT) important for trusting its predictions?","acceptedAnswer":{"@type":"Answer","text":"Calibration of AlphaFold 2's uncertainty measure (pLDDT) was critical because a structure prediction with a GDT score around 90 is only useful if its confidence score accurately reflects correctness; an uncalibrated but highly confident wrong answer could mislead researchers into wasting a year of work. Trustworthy calibration mattered more than full mechanistic interpretability for safe practical use. daily.dev keeps builders of AI-reliant workflows informed on why model calibration matters as much as accuracy."}}]} ```