The AI Models Doing Science No One Talks About: AlphaFold, ESM, and the Discovery Engines

No chatbot discovered 2.2 million new crystal structures or folded a billion proteins. A different family of AI models did that — quietly, without a subscription plan or a press cycle, and most of the public has never heard their names. This is the third article in LIWARSE’s five-part series on the AI landscape by type of use.

Scientific foundation models are built differently from general-purpose assistants: narrower in scope, trained on structured scientific data rather than internet text, and judged by whether their predictions hold up in a lab, not by how fluently they converse. LIWARSE’s physician-founder perspective makes this category especially relevant — much of what follows either already touches patient care or will within a few years.

Structural Biology: AlphaFold 3 and the ESM Family

Google DeepMind’s AlphaFold line — recognized with the 2024 Nobel Prize in Chemistry for its earlier AlphaFold2 — has evolved from predicting a single protein’s shape to predicting how proteins interact with DNA, RNA, and drug-like small molecules. The freely accessible AlphaFold Protein Structure Database now covers more than 200 million predicted structures. DeepMind’s own drug-discovery spin-off, Isomorphic Labs, has since built a proprietary successor it calls IsoDDE — unlike AlphaFold, kept entirely in-house, a notable retreat from the open-publication norm that made AlphaFold so widely useful.

Running in parallel is the ESM family, most recently extended into an open-source atlas — built with ESMFold2 — that has predicted the shape of roughly a billion proteins, vastly expanding the known protein universe beyond what AlphaFold alone had mapped. Where AlphaFold started from structure, ESM’s language-model approach treats protein sequences the way an LLM treats text, giving it a complementary strength in generative protein design rather than pure structure prediction.

---

Materials, Weather, and the Physical Sciences

Beyond biology, Google DeepMind’s GNoME has computationally identified 2.2 million candidate stable crystal structures, of which hundreds have already been synthesized by outside researchers — a materials-science discovery rate no human-only laboratory pipeline could match. GraphCast and its successors have shown that a learned model can rival, and in some measures beat, traditional numerical weather simulators at a fraction of the compute cost, with direct implications for disaster preparedness and, by extension, public safety.

A newer and more structurally significant development is the emergence of AI “co-scientist” systems — multi-agent tools, including one built on Google’s Gemini, that propose and debate biomedical hypotheses for a human researcher to evaluate, rather than simply answering a query. These systems do not replace the scientist who designs the experiment and judges the result; they compress the space of candidate hypotheses a human has to sift through first.

---

Risks and Benefits Through the LIWARSE Lens

Benefits

• These models compress research timescales that once took human teams years into days or weeks — directly serving LIWARSE’s mission of advancing human and all life on Earth through faster discovery of treatments, materials, and climate tools.
• Open databases like the AlphaFold Protein Structure Database and the ESMFold2 atlas democratize access: a lab in a resource-poor region can query the same structural predictions as a lab at a major research university.
• Faster, cheaper weather prediction from models like GraphCast has direct, near-term public-safety value for disaster response.

Risks

• Generative structural models are explicitly acknowledged, including by AlphaFold 3’s own developers, to be prone to hallucinating plausible-looking structure in poorly-determined regions — a failure mode with real consequences if a predicted structure guides a drug trial without adequate confidence screening.
• The shift from AlphaFold’s open publication model to Isomorphic Labs’ proprietary IsoDDE signals that the most capable next-generation tools in this category may not stay freely available, concentrating advanced discovery capability in whichever lab can afford to keep it closed.
• AI co-scientist systems risk being treated as autonomous discoverers rather than hypothesis generators if the human-judgment step at the end of the loop is skipped under time or funding pressure.

The LIWARSE Assessment

Scientific foundation models are among the clearest embodiments of LIWARSE’s mission in the entire AI landscape — tools that measurably advance human and biological life, with a track record already validated in peer-reviewed literature rather than benchmark claims alone. The open-versus-closed tension that runs through every other article in this series appears here too, in a quieter form: whether the next generation of discovery tools follows AlphaFold’s open-publication precedent or Isomorphic Labs’ proprietary one will shape who gets to build on these breakthroughs next.

---

Under the 3 Absolute Laws, a model that helps discover a life-saving material or medicine serves the first law directly — provided its predictions are verified, not simply trusted, before they reach a patient or a production line.

— The LIWARSE Movement | liwarse.org
Safety of Life · Advancement of Life · Together.