AlphaFold's Nobel Prize and GNoME's 2.2 million materials both got real headlines — but the honest story is about candidate generation, not discovery, and the gap between those two claims is where the hype lives.
Google DeepMind's GNoME project claimed to have discovered 2.2 million new crystal structures in 2023, of which some 380,000 were predicted to be stable enough to attempt synthesizing. It was framed as compressing 800 years of materials discovery into a matter of months. Then, in 2024 and 2025, independent researchers who actually checked the underlying crystallographic data — including a widely discussed critique from a materials science group at Princeton — found that a large share of the "novel" structures were duplicates, near-duplicates, or simple derivatives of compounds already in the training data, and that far fewer had genuinely new, experimentally useful chemistry than the headline number implied.
That pattern — a genuinely useful model, wrapped in a discovery number that doesn't survive contact with domain experts — is close to the median outcome for AI in science right now. The tools are real and some of them are already load-bearing in industrial pipelines. But "AI discovered X" almost always means "AI generated a large set of candidates, and a much smaller number survived expensive, slow, human-supervised verification." The interesting engineering question isn't whether the models work — it's where the candidate-generation step actually shortens the bottleneck, and where the bottleneck was never candidate generation in the first place.
Structural biology is the strongest case for AI in science, because it's the one place where the correctness of the model's output can be checked cheaply and definitively — you can crystallize the protein, or run cryo-EM, or increasingly just cross-validate against a second predictive model — and it worked well enough consistently that the field's own two AlphaFold-2 leads, Demis Hassabis and John Jumper, shared the 2024 Nobel Prize in Chemistry.
AlphaFold3 (2024) extended structure prediction from single proteins to protein-ligand, protein-nucleic acid, and protein-protein complexes, which is the version of the problem drug discovery actually cares about — a folded protein alone doesn't tell you whether a candidate small molecule will bind to it. DeepMind spun the commercial application out into Isomorphic Labs, which has already signed drug-discovery partnerships with Eli Lilly and Novartis worth billions in potential milestones. A separate lineage — Meta's ESM3 and the startup EvolutionaryScale it spawned, plus Chai Discovery's Chai-1 and Chai-2 models — treats protein sequence and structure as a joint language-model problem, generating novel protein designs rather than just predicting the structure of existing ones.
The caveat that matters here: AlphaFold-class models are excellent at predicting structure for sequences reasonably close to what's in the training distribution (the PDB, plus large sequence databases), and meaningfully worse on intrinsically disordered regions, on modeling the effect of a single point mutation on binding affinity, and on cases requiring real conformational dynamics rather than a single static structure. Wet-lab validation — expressing the protein, purifying it, running a binding assay — is still where most of the calendar time in a drug program goes. The model didn't remove that step; it removed months of trial-and-error candidate narrowing that used to precede it.
Materials science has the same shape as protein design — generate a candidate crystal structure, predict its formation energy and stability with a model, then try to synthesize it — but with a much harder verification step, because synthesizing an inorganic solid with a specific structure is slower and less standardized than expressing a protein.
GNoME used graph neural networks trained on density-functional-theory (DFT) calculations to screen candidate crystal structures for thermodynamic stability, and Berkeley Lab's "A-Lab" paired it with a robotic, largely autonomous synthesis pipeline that successfully made dozens of the predicted compounds. Microsoft's MatterGen takes a generative rather than screening approach — it produces candidate structures directly, conditioned on target properties like band gap or magnetic behavior, instead of filtering a pre-enumerated list. Both approaches report real synthesized compounds, which is more than most "AI discovery" claims can say.
The independent re-analysis of GNoME's output is the useful cautionary data point, though, and it's worth being specific about what it actually showed rather than repeating the discovery number uncritically: stability, as these models define it (a DFT-computed energy above the convex hull), is a necessary condition for a material to be synthesizable, not a sufficient one, and a meaningful fraction of the "new" structures turned out to be symmetry-equivalent or compositionally trivial variants of known phases rather than chemically novel materials. The DFT calculations used for training and screening are themselves an approximation with known systematic errors, so a model trained to reproduce DFT energies inherits DFT's blind spots. None of this means the pipeline is useless — A-Lab's synthesized compounds are real — it means the honest way to read a "2.2 million structures" headline is as "2.2 million candidates," with the actual discovery count being whatever fraction survives independent stability re-checks and eventual synthesis, which is orders of magnitude smaller.
The other cluster of tools treats the scientific literature itself as the bottleneck: a working biologist cannot read the ~1.5 million papers published in the life sciences each year, and standard keyword search misses papers that are relevant but don't share vocabulary with the query.
Elicit and Consensus do literature-grounded question answering — retrieve papers, extract claims, and synthesize an answer with citations, essentially a domain-restricted RAG system with extra steps around claim extraction. FutureHouse, a nonprofit spun out of former OpenAI and Google researchers, built PaperQA2 specifically to reduce hallucinated citations (a known failure mode of naive literature-QA systems) by forcing every claim to be traceable to a retrieved passage, and followed it with Robin, an agent that chains literature search, hypothesis generation, and even proposes analyses against public datasets with fairly minimal human steering. Google's "AI co-scientist" system, built on Gemini, uses a multi-agent debate-and-tournament structure — generation, reflection, ranking, and evolution agents critiquing each other's hypotheses — and was tested on a narrow, cherry-picked-by-design set of problems: it reportedly reproduced an unpublished experimental finding about drug repurposing candidates for a form of leukemia that a Stanford lab had already validated independently, which is a genuinely interesting result but is also exactly one anecdote, not a controlled study of the system's hit rate.
This is also the corner of AI-for-science that looks most like the tool-calling agent stack this blog usually covers — these systems are, mechanically, an LLM orchestrating calls to a literature search API, a code execution sandbox, and sometimes a lab-automation API, which is the same tool-calling pattern MCP servers expose to general-purpose coding agents, just pointed at PubMed and a data repository instead of a filesystem. The hard part isn't the orchestration; it's that "novel hypothesis" is a much fuzzier success criterion than "tests pass," so evaluating whether these systems are actually generating good science, versus plausible-sounding recombinations of the literature, is still mostly done by domain experts reading the output by hand.
The most literal version of "AI does science" is a robotic lab that runs its own experiments in a closed loop — propose a candidate, synthesize or test it, feed the result back into the model, repeat without a human in the loop for each cycle. Berkeley's A-Lab and Emerald Cloud Lab (a commercial, remotely-operated lab used by academic and biotech groups) are the most cited working examples, and Carnegie Mellon's "Coscientist" project showed an LLM planning and executing a multi-step chemistry synthesis using lab automation APIs, including looking up a procedure and correcting its own errors mid-run.
These systems work today for reactions and measurements the lab was specifically built and calibrated to run — a narrower slice of "science" than the framing suggests. They're closer to a very well-instrumented, self-optimizing production line for a defined class of experiments (crystal synthesis, well-characterized organic reactions) than to a general-purpose robotic scientist. Scaling to new reaction classes still requires new hardware calibration and new safety validation, which is slow and expensive in exactly the way that resists a software-only speedup.
| Application area | What the model does | Maturity / adoption | Where it currently breaks down |
|---|---|---|---|
| Protein/molecule structure prediction | Predicts 3D structure and complexes from sequence | High — production use in pharma (Isomorphic Labs, Recursion) | Disordered regions, point-mutation effects, dynamics |
| Materials discovery | Screens or generates candidate crystal structures for stability | Medium — real synthesized compounds, but discovery counts are commonly overstated | DFT-derived training data inherits DFT's errors; symmetry-duplicate candidates |
| Literature synthesis / hypothesis generation | Retrieves, summarizes, and recombines published findings | Medium — genuinely useful for search, unproven at scale for novel hypotheses | Citation hallucination (partially mitigated), hard to evaluate "novelty" objectively |
| Self-driving labs | Closed-loop propose→synthesize→measure→retrain | Low-medium — works for narrow, pre-calibrated reaction classes | Doesn't generalize to new reaction types without new hardware/safety work |
Across all four areas, the honest version of "AI accelerated this" is almost always "AI cheaply generated a much larger candidate pool, and an expensive, slow, human-or-robot verification step got to be more selective as a result." That's a real and valuable compression of the discovery pipeline — it's just a different claim than "AI discovered a new drug" or "AI discovered 380,000 materials," and the gap between those two claims is where most of the hype-versus-reality argument in AI-for-science actually lives. When you read a discovery announcement, the question worth asking isn't "did the model do this," it's "what fraction of the model's output survived the same verification standard the field used before the model existed" — and that number is almost never in the press release.