Research
AlphaFold Was a Breakthrough, Not a Complete Account of Biology
A Latent Space panel with DeepMind and Biohub treats AlphaFold as a breakthrough, not the end of biological modeling. Here is what R&D teams can take from it.
AlphaFold was a breakthrough, but the Latent Space panel featuring Google DeepMind’s Pushmeet Kohli and Biohub’s Sal Candido frames protein structure prediction as only part of a larger challenge: modeling biology. Moderator Brandon Anderson leads a discussion about protein dynamics, disorder, biological systems, data, scaling, and trustworthiness. 1
For research and development teams using AI in biological research, the practical distinction is between what a model predicts and what a team hopes to conclude from that prediction. The panel’s topics suggest caution about treating a result on one biological modeling task as an answer to a broader one. 1
AlphaFold limitations: structure, dynamics, and biological systems
The source describes AlphaFold as a breakthrough while also identifying protein folding as unsolved. It names protein dynamics, disorder, and the limitations of static structure prediction as part of the discussion. 1 One useful reading is that a structural prediction and a claim about how a protein behaves are different kinds of claims; the latter should not be inferred automatically from the former.
The panel also discusses moving from models of individual proteins to whole biological systems and says building a “virtual cell” requires fundamentally different datasets. 1 That frames system-level modeling as more than extending a protein-level result. It does not, by itself, specify the data or methods needed to build such a model.
| Modeling question | What the source says | Practical reading |
|---|---|---|
| Protein structure | AlphaFold was a breakthrough, while protein folding remains unsolved. 1 | Be precise about what a structural prediction establishes. |
| Dynamics and disorder | These are named alongside the limitations of static structure prediction. 1 | Do not treat a static prediction as a complete account of protein behavior. |
| Whole biological systems | The panel discusses moving beyond individual proteins and says a “virtual cell” needs fundamentally different datasets. 1 | A protein-level result does not, on its own, establish a system-level outcome. |
These are practical implications drawn from the distinctions in the source, not a validation protocol presented by the panel. For an R&D team, a useful first question is: what is the model’s prediction target, and does it match the biological claim being made?
Scaling AI in computational biology: what more data or compute might change
Candido’s discussion, as summarized in the source, emphasizes that scaling compute and data alone will not solve biology. The source says that finding the right scaling law matters and that good data matters more than simply having more data. 1 This supports a measured takeaway: added scale is relevant when the data and task support it, not as a general guarantee of better biological predictions.
The panel raises whether low-quality metagenomic data could improve protein language models and whether those models may contain scientific knowledge that has not been uncovered. 1 The source presents these as questions for discussion, not as reported benchmark results. It also covers AlphaFold’s handcrafted architecture, scientific intuition, and tradeoffs involving specialized architectures and scaling. 1
AI drug discovery: a question, not a result
The source lists the question of when AI could deliver 10x–100x acceleration in drug discovery. 1 That is a topic framed as a question, not evidence in the supplied material that such an acceleration has been achieved.
Cryo-EM and richer representations
The panel discusses how cryo-EM micrographs could unlock richer biological representations. 1 The supplied source material does not report a measured gain or benchmark showing that adding micrographs improves a particular prediction task. 1 For teams assessing this direction, a richer input and demonstrated improvement on a relevant task are separate questions.
Trust and evaluation of biological AI
Trustworthiness and uncertainty calibration are identified as important topics, with the source framing them as more important than full interpretability. 1 The supplied material does not specify a measurement method, benchmark, or acceptance threshold for evaluating those properties. 1
A practical review can start with questions suggested by the panel’s topics:
- What is the prediction target: structure, dynamics, disorder, or a system-level outcome? 1
- Is the available data relevant to that target? The panel discusses the importance of finding the right data and scaling relationship. 1
- What uncertainty or trust evidence is available? The panel names these as priorities. 1
- What additional validation is needed before using the prediction to support a broader biological claim? This is an editorial question prompted by the distinctions in the discussion, not a protocol specified by the panel. 1
The source excerpt does not provide a ready evaluation method for these questions. 1 That limits what a reader can take from it as operational guidance; teams will need to establish their own criteria for a particular research decision.
The same habit of reading a score narrowly applies well beyond biology. Sakana AI's 73% core-claim result is a benchmark, not a deployment decision makes a comparable point about benchmark numbers, and AI Agent Governance: What Evidence Is Enough Before an Agent Can Take Action? takes up the related question of how much evidence is enough before acting on a model's output.
A decision rule for R&D teams
Treat the prediction target, supporting evidence, uncertainty, and intended claim as separate items to review. This checklist is an editorial synthesis of the panel’s discussion of protein structure, biological systems, data, trustworthiness, and uncertainty; it is not presented in the source as a formal framework. 1
The central question is not whether a biological AI tool sounds capable. It is whether the model’s output supports the specific conclusion a team wants to draw. The panel makes the case for thinking beyond individual proteins, while the supplied source material leaves implementation and evaluation details open. 1
Questions
- What are AlphaFold’s main limitations discussed in the Latent Space interview?
- The discussion identifies protein dynamics and disorder as limitations of static structure prediction and distinguishes individual protein modeling from whole biological-system modeling. [1]
- Does AlphaFold solve protein folding?
- No. The discussion presents AlphaFold as a major breakthrough but identifies protein folding as unsolved. [1]
- Does scaling compute guarantee better biological AI models?
- No. Candido’s point, as summarized in the source, is that scaling compute and data alone will not solve biology. Finding the right scaling law matters, and good data matters more than simply having more data. [1]
- Does the discussion provide a benchmark for evaluating biological AI tools?
- No. It raises trustworthiness and uncertainty calibration as important concerns but does not provide a concrete protocol, dataset, metric, or acceptance threshold. [1]
- What still needs experimental validation?
- The source does not identify specific predictions or prescribe experiments. It supports treating model outputs as limited predictions whose appropriate validation remains an open workflow question. [1]
Sources
- Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub · Latent Space · Oct 10, 2026
How this article was made: an AI writer persona researched and wrote it from the sources listed above; separate AI passes edited it, checked its claims against those sources and audited it; an editor at Yelno reviewed and approved it before publication. How the Yelno blog works.