This atlas asks how AI is changing the way basic science gets done. Four questions run through it: which problems have a verifier hard enough to trust, who is selling the picks and shovels, who is doing the mining, and which layer the value finally settles in. It has three parts: long-form research tracks with graded sources, a discipline map that follows the AI-method frontier in six fields, and a directory of the companies and models that matter.
The research tracks
The tracks are long, source-graded studies. Biology is by far the largest.
| Track | What it covers | Scale |
|---|---|---|
| AI for Bio | Challenges and validation, AI drug discovery, organoids and alternative methods, CROs/CDMOs, lab supply chains, the innovative-drug industry and its capital, and neuro and wetware computing. | 147 studies, 7 sections |
| Scientific simulation | A map of research simulation environments: three generations of simulation (first-principles, AI emulators, data-driven world models), three layers of sim2real gap, five ways simulation and AI relate, and the industry and capital picture. | 3 generations, 3 gap layers |
| AI benchmarks (related) | Who sets the questions, who grades them, and whether the questions themselves can be trusted. Includes Chinese science benchmarks, a summary of AI4S capabilities, and rubric-based evaluation. | Companion study |
The most recent additions (Oct 2026) are a six-part survey of global biopharma investors: a cross-region overview, China's tier-one and tier-two specialist VCs, state-owned funds, government guidance funds and pharma corporate VC, the newer generalist VCs with AI-plus-bio, synthetic biology and brain-computer-interface teams, and the A-share and Hong Kong public-market funds and managers.
From assistance to agency: the method spectrum
Several 2025 surveys and position papers converge on one framing. AI for Science is moving from partial assistance towards full scientific agency: systems that run hypothesis generation, experimental design, execution, analysis and refinement as one closed loop (Agentic Science survey). A NeurIPS 2025 position paper (OpenReview) splits the debate into two camps:
- Paradigm enhancement. AI speeds up traditional scientific tasks. This is today's mainstream, and it is overwhelmingly built on domain-specific or task-specific models.
- Paradigm transition. A general model fundamentally redefines how science is done. This is the frontier bet, and the evidence is thin. One claim, that foundation models turn materials science into a general-purpose, emergent AI system, failed adversarial verification in our Jun 2026 review.
So the accurate contrast is not "general versus specialised" but domain foundation model versus task-specific model. A single general model that works across the sciences does not really exist yet.
| Layer | Dominant method | Representative systems |
|---|---|---|
| Solving physical equations | PINNs and neural operators (FNO, AFNO, DeepONet) | NVIDIA PhysicsNeMo |
| Structure prediction and de novo design | Generative models (diffusion, flow matching) plus GNNs | AlphaFold 3, MatterGen, RFdiffusion |
| Atomistic simulation | Machine-learned interatomic potentials (MLIPs) | MatterSim, MACE-MP-0 |
| Discovery loops | GNN plus active learning plus first-principles (DFT) checks | GNoME |
| Reasoning and proof | LLM plus formal verification (Lean) plus search | AlphaProof, DeepSeek-Prover |
Two flagships that set the playbook
- AlphaFold 3 is a single diffusion model that jointly predicts complexes of proteins, nucleic acids, small molecules and ions, and it far outperforms classical docking (Nature, 2024). Notably it dropped AlphaFold 2's equivariant structure module in favour of diffusion over raw atom coordinates. That is a data-scale win for the "big generative model" side of the equivariance debate.
- GNoME combined graph networks, active learning and DFT verification to find about 2.2 million new crystals, roughly 380,000 of them highly stable. It lifted the stability hit rate on MatBench Discovery from about 50% to 80% (DeepMind).
Why domain models, not general LLMs
Science models reuse the same backbones as LLMs (Transformers, diffusion, GNNs), but they differ in four structural ways:
- The tokens are not text. They are amino acids, atoms, crystal lattices, or grid points in a physical field.
- Symmetry is built in. SE(3)/E(3)-equivariant networks hard-code invariance to rotation and translation, which a general LLM would have to learn from data.
- The objective is not next-token prediction. It is masked modelling (ESM), diffusion denoising (AlphaFold 3, MatterGen) or a physics-residual loss (PINNs, which can train with zero labelled data).
- The scale is one to two orders of magnitude smaller. Protein and materials foundation models run from millions to billions of parameters, not trillions. They trade scale for inductive bias.
Three questions decide it. Is there enough self-supervisable data to pretrain a foundation model? (Sequences, yes; PDEs, no.) Is the problem under hard physical constraints? (That favours specialised models with physics priors.) Is there a cheap verification loop, such as DFT or an experiment, that lets you trust generation at scale?
Biology: proteins, genomes and cells
Biology is arguably where AI has moved fastest from demonstration to infrastructure. The reason is structural. Life encodes itself in sequences: proteins are strings of amino acids, genomes strings of nucleotides. Decades of high-throughput sequencing have produced billions of these strings, which can be learned self-supervised, exactly the regime where large models excel.
Structure prediction
AlphaFold 2 (2021) used an Evoformer that reasoned jointly over a multiple-sequence alignment (MSA) and a pairwise residue representation, with an SE(3)-aware structure module emitting coordinates. AlphaFold 3 (2024) replaced that module with a diffusion network that denoises a cloud of atoms into a structure. Its decisive change is scope: one framework for complexes of proteins, nucleic acids, ligands, ions and modified residues. It reports far better protein–ligand accuracy than specialised docking tools, better protein–nucleic-acid accuracy than dedicated predictors, and better antibody–antigen accuracy than AlphaFold-Multimer v2.3.
ESMFold (Meta AI, 2022–23) folds from a single sequence, with no MSA or templates, by reading structure out of the ESM-2 protein language model (variants up to 15B parameters). It is reportedly about ten times faster than AlphaFold 2, which made the ESM Metagenomic Atlas of more than 600 million proteins possible. Accuracy is competitive where the language model is confident and weaker elsewhere.
All of these remain static predictors: dynamics, allostery and folding pathways are out of scope. Independent analyses report that AlphaFold 3 can hallucinate order in intrinsically disordered regions. Chirality and unusual stoichiometries remain weak spots, and accuracy drops for sequences with little evolutionary signal.
De novo design
- RFdiffusion (Baker Lab) runs SE(3)-equivariant diffusion on the RoseTTAFold backbone, treating each residue as a rigid frame, and generates backbones conditioned on motifs, symmetries or binding targets.
- ProteinMPNN solves the inverse-folding problem: given a backbone, design a sequence that folds into it. The canonical pipeline is RFdiffusion, then ProteinMPNN.
- AlphaProteo (DeepMind, 2024) targets de novo binders directly. On seven targets it reported 3× to 300× better binding affinities, including binders to VEGF-A and a SARS-CoV-2 protein, often after one round of medium-throughput screening.
- Flow-matching designers are the emerging alternative to score-based diffusion, offering straighter probability paths and faster sampling on the SE(3) manifold.
Sequence and genomic foundation models
- ESM3 (EvolutionaryScale, 2024) reasons jointly over sequence, structure and function and was trained, per the lab, on 2.78 billion proteins. It generated a novel fluorescent protein with about 58% identity to known GFPs, described as the equivalent of some 500 million years of evolution.
- Evo 2 (Arc Institute and NVIDIA, 2025) is a DNA model on the StripedHyena architecture, at 7B and 40B parameters, trained on about 9.3 trillion base pairs from more than 128,000 species, with context up to about one million nucleotides. It reports state-of-the-art zero-shot variant classification, for example on BRCA1.
- AlphaGenome (DeepMind, 2025) reads up to about 1 Mb of DNA and predicts thousands of functional tracks at up to single-base resolution. It matched or beat the best external models on 24 of 26 variant-effect evaluations.
- Nucleotide Transformer (InstaDeep, 2024) spans 50M to 2.5B parameters; a v3 line extends context towards 1 Mb.
The honest debate about cell models
scGPT and Geneformer apply the language-model recipe to single-cell transcriptomics, pretraining on tens of millions of cells. The candid 2025 finding: in zero-shot settings they can be matched or beaten by much simpler baselines (highly-variable-gene selection, scVI, Harmony) on batch integration and clustering, and more pretraining data did not reliably help. By 2026 the "virtual cell" had become its own funded field, with benchmarks such as the Virtual Cell Challenge and MVCBench and companies including GenBio AI, Xaira and the Arc Institute.
The verdict. Biology is the field closest to genuine foundation models, because sequence data is huge, self-supervisable and full of evolutionary signal. Sequence space is conquered; the cell and the regulatory genome are not. Open problems: dynamics rather than snapshots, disorder and hallucination, experimental validation as the only ground truth for design, cell models that robustly beat baselines and predict perturbations, and leak-free benchmarks. The full library is in AI for Bio.
Chemistry: reactions, molecules and synthesis
AI for chemistry clusters into four loosely coupled problem families, plus an orchestration layer of LLM agents on top.
| Family | Core method | Flagship | Data | Maturity |
|---|---|---|---|---|
| Reaction prediction | Seq2seq Transformer, graph nets | Molecular Transformer | USPTO patents (millions, biased) | High (forward) |
| Retrosynthesis | Search plus template or neural policy | AiZynthFinder, Synthia, IBM RXN | Templates from patents | Medium–high |
| Generation | VAE, diffusion, GFlowNet | JT-VAE, 3D diffusion, GFlowNet | Property-labelled molecule sets | Medium (validity gap) |
| MLIPs | Equivariant GNN | MACE, ANI-1xBB, UMA, Allegro | DFT calculations (OMol25, OC20) | High near equilibrium, medium for reactions |
| Agents | Tool-using LLM | ChemCrow, Coscientist | Tools, literature, hardware | Early, demo stage |
- Reaction prediction. The Molecular Transformer (2019) recast forward prediction as machine translation over SMILES and added uncertainty calibration. Forward prediction is the most "solved" family, but graph-based models such as RXNGraphormer show the retrosynthetic direction is more sensitive to distribution shift.
- Synthesis planning is search over a single-step model. AiZynthFinder (AstraZeneca, open source) runs Monte Carlo tree search down to purchasable precursors, usually in well under a minute. Synthia (formerly Chematica) combines hand-coded rules with ML scoring and has produced routes to complex natural products that chemists judged plausible. ASKCOS (MIT) adds condition recommendation. RAscore, a cheap classifier predicting whether AiZynthFinder would find a route, now filters generative pipelines.
- Generation. 3D equivariant diffusion is the most active line; GFlowNets sample in proportion to reward, giving diverse candidate batches. The hard part is validity, synthesisability and property fidelity at once.
- MLIPs replace DFT in the loop at a fraction of the cost. The frontier is reactivity: ANI-1xBB (2025, 13M+ geometries) targets barrier heights and bond dissociation. Meta FAIR's OMol25 (more than 100M DFT calculations) and the Universal Model for Atoms (UMA) brought foundation scale in 2025.
- Agents. ChemCrow gave GPT-4 18 expert tools and planned real syntheses; Coscientist designed and ran palladium-catalysed cross-coupling on robotic hardware. Chemistry is one of the few fields where a closed hypothesise-execute-analyse loop has run on physical equipment.
A-Lab (LBNL with Google DeepMind) reported in Nature in Nov 2023 that an autonomous lab made 43 new inorganic materials in 17 days. Independent chemists argued its XRD phase identification was unreliable and that several "new" compounds already existed or were not convincingly characterised. Nature issued a correction, and critics say the core concerns remain unresolved as of early 2026. Autonomy claims must clear a higher characterisation bar than the headline metric.
Why chemistry trails biology. There is no AlphaFold of chemistry, and the gap is structural: fragmented data (MoleculeNet is a suite of small datasets, not one large corpus); reaction outcomes that depend on solvent, temperature, catalyst, concentration and order of addition, which public data omits; reporting bias, since patents over-represent popular reactions and almost never record failures; and multi-scale physics from electrons to bulk thermodynamics. SMILES foundation models, OMol25/UMA and agents push towards generality, but as of mid-2026 no one model dominates all four families. Open problems include condition and yield prediction, leaky USPTO splits that inflate top-k scores, reactive MLIPs, synthesisability of generated molecules, and reproducibility protocols for autonomous labs.
Physics: PINNs, operators and simulation
AI for physics either solves the governing equations or replaces an expensive solver with a learned surrogate. It splits into two complementary families, plus a set of large flagship models that moved from demos to operations between 2024 and 2026.
| Axis | PINN | Neural operator |
|---|---|---|
| Output | One solution instance | A solution operator for a whole family |
| Supervision | The governing equation (can be data-free) | Usually data; PINO adds a physics loss |
| Reuse | Retrain for each new configuration | Amortised: one forward pass per instance |
| Strength | Accuracy, inverse problems, sparse data | Throughput across many instances |
| Weakness | Generalisation, training convergence | Out-of-distribution extrapolation |
PINNs put the PDE residual into the loss via automatic differentiation, so they can train with zero labelled data, but they solve a single instance and struggle with multi-scale and high-frequency behaviour. Neural operators learn mappings between function spaces: the Fourier Neural Operator (2021) works in the Fourier domain and can super-resolve zero-shot; AFNO turns that spectral mixing into a Vision Transformer token mixer and powers FourCastNet; DeepONet's branch/trunk design handles irregular geometries. NVIDIA PhysicsNeMo, the open-source successor to Modulus, packages all of these for industrial CFD, structural and electromagnetic work.
Weather and climate went operational
- GraphCast (Google DeepMind): GNN, 0.25° global medium-range forecasts matching or beating numerical weather prediction on many targets.
- Pangu-Weather (Huawei): 3D Vision Transformer, competitive with ECMWF's IFS, including cyclone tracks.
- FourCastNet (NVIDIA): an early AFNO model at 0.25°.
- GenCast (Google DeepMind): diffusion ensembles for 15 days, beating ECMWF's ENS on most targets, including extremes.
- Aurora (Microsoft): an Earth-system foundation model fine-tuned for air quality, ocean waves and cyclones (Nature, 2025).
- ECMWF AIFS became operational on Feb 25, 2025, alongside the physics-based IFS; the ensemble version followed later in 2025.
Beyond weather: DeepMind and EPFL used reinforcement learning to control the TCV tokamak's magnetic coils (Nature, 2022); learned closures and differentiable solvers accelerate turbulence and CFD; normalising flows speed up lattice QCD sampling; and emulators stand in for N-body cosmology runs. Open problems are out-of-distribution generalisation, long-horizon stability of autoregressive emulators, shocks and fine-scale turbulence, physical guarantees such as conservation, and how best to hybridise learned and physics-based components. The applied side, from first-principles codes to world models, is covered in Scientific simulation.
Mathematics: proving, formalising and conjecturing
Mathematics is the cleanest stress test for machine reasoning: a formal proof is mechanically checkable, so the verifier is free, exact and incorruptible. Two cultures are converging: informal LLM reasoning in natural language, and formal systems that emit machine-checked proofs in Lean, Isabelle or Coq. The winning pattern in both is neuro-symbolic.
| System | Approach | Result | Year |
|---|---|---|---|
| AlphaProof + AlphaGeometry 2 | LLM plus AlphaZero RL in Lean; symbolic geometry | IMO 2024 silver: 28/42, 4 of 6 solved | 2024 |
| AlphaGeometry 2 | LLM proposer plus symbolic deduction | 84% of IMO geometry 2000–2024 (up from 54%); IMO 2025 P2 in about 20 s | 2025 |
| DeepSeek-Prover-V2 (671B) | RL plus subgoal decomposition, Lean 4 | 88.9% on miniF2F-test; 49/658 on PutnamBench | 2025 |
| Kimina-Prover / Autoformalizer | RL formal reasoning; natural language to Lean | Open-weight prover and autoformaliser | 2025 |
| Gemini Deep Think | Informal LLM, parallel reasoning | IMO 2025 gold, 35/42 (human-graded) | 2025 |
| FunSearch | LLM plus evolutionary program search | New cap-set lower bound; bin-packing heuristics | 2023 |
| Ramanujan Machine | Numerical conjecture search | Continued-fraction conjectures for constants | 2021 |
| DeepMind × Oxford/Sydney | ML-guided intuition | Knot-theory theorem; Kazhdan–Lusztig progress | 2021 |
AlphaProof bootstrapped RL self-play by autoformalising about 80 million problem statements into Lean; it was the first AI to reach an IMO medal tier. OpenAI also reported 35/42 at IMO 2025, without official certification. Autoformalisation is the load-bearing bridge between the two cultures and the weakest link: frontier models struggle to emit well-typed Lean 4, sampling and filtering helps, and a statement can compile yet not mean what the English meant.
Three different claims hide behind the word: a machine-checked formal proof (strongest), a human-graded natural-language proof, or a correct final answer on a numeric benchmark, which says nothing about the reasoning. Contamination threatens the last most. MATH and AIME are largely saturated and widely published; PutnamBench and miniF2F use formal targets; FrontierMath (Epoch AI, 338 unpublished problems across four tiers) is the toughest standard benchmark. The AIMO Prize, funded by XTX Markets, awarded its first progress prize ($131,072) to Team Numina in 2024 and its second to NVIDIA's NemoSkills team.
The common skeleton: an LLM proposes a candidate, a formal verifier returns an exact reward, and a search or RL loop improves the proposer. Open problems are the cost of proof search, separating novel reasoning from memorisation, faithful autoformalisation, making full informal proofs machine-checkable, and discovery beyond narrow scorable niches.
AI for AI: automating ML research
Here AI is both the tool and the subject. The discipline splits into three tiers, and keeping them apart is the whole job: narrow algorithm discovery (real, verified wins), autonomous "AI scientists" (impressive pipelines of contested quality), and recursive self-improvement (mostly a thought experiment).
| System | What it found or did | Year | Verified? |
|---|---|---|---|
| AlphaTensor | Faster matrix-multiplication decompositions; 4×4 improvement on Strassen in GF(2); hardware-tuned schemes 10–20% faster | 2022 | Yes (Nature; checkable) |
| AlphaDev | Shorter sort-3/4/5 assembly; faster small hashing | 2023 | Yes (merged into LLVM libc++) |
| FunSearch | New cap-set constructions; online bin-packing heuristics | 2023 | Yes (Nature; verifiable) |
| AlphaEvolve | Recovered ~0.7% of Google's global compute; ~23% faster Gemini kernel; 50+ math problems (matched ~75%, improved ~20%); kissing number in 11D 592 → 593 | 2025 | Partly: infra wins verified, "56-year record" framing disputed |
| AI Scientist v2 (Sakana) | Fully generated ML paper cleared an ICLR 2025 workshop review | 2025 | Weakly: the paper later found to contain hallucinations |
AlphaDev's gains are large only for tiny fixed-size inputs (about 70%) and about 1.7% above 250,000 elements. AlphaEvolve's "48 multiplications for 4×4 complex matrices, first improvement over Strassen in 56 years" is the most over-quoted claim in the field: Winograd reached 48 in 1967, and Waksman's 1970 algorithm uses 46 over commutative rings with division by 2. It is a genuine improvement in a narrower setting, not a clean record.
Sakana's AI Scientist v1 wrote papers for about $15 each, but an independent evaluation found about 42% of experiments failed on coding errors, a median of about five often-outdated citations, and hallucinated numbers. The wider agent wave (Google's AI co-scientist, FutureHouse, open-source scaffolds) is consistently strong at ideation and boilerplate and weak at judging novelty and at not fooling itself.
| Benchmark | Task | Headline |
|---|---|---|
| MLE-bench (OpenAI, 2024) | 75 Kaggle competitions | Best setup reaches bronze or better in 16.9% |
| RE-Bench (METR, 2024) | 7 ML R&D environments vs 61 experts | Agents ~4× humans at 2 h; humans narrowly ahead at 8 h; ~2× ahead at 32 h |
| PaperBench (OpenAI, 2025) | Replicate 20 ICML 2024 papers (8,316 sub-tasks) | Best agent 21.0%; on a 3-paper subset ML PhDs 41.4% vs o1 26.6% |
Agents are fast and cheap but plateau at long horizons. Classical neural architecture search is mostly historical, overtaken by scaling laws. What is real in self-improvement is bounded: LLMs help with data curation, synthetic data, reward rubrics and training kernels. Recursive, unbounded self-improvement is unproven; models collapse when trained recursively on their own ungrounded outputs, and every working loop relies on a verifier or a human. Labs betting on the loop scaling include Recursive Superintelligence (a reported ~$4.65B valuation, pre-product), Safe Superintelligence (a reported ~$32B, no product) and Meta Superintelligence Labs (anchored by a $14.3B Scale AI deal), alongside Google DeepMind, OpenAI and xAI. Treat "ASI" as a stated goal and capital allocation, not a result.
Operations research: learned optimisation
OR covers routing, scheduling, resource allocation and combinatorial optimisation, with decades of mature machinery: branch-and-bound, cutting planes, metaheuristics and MIP solvers. The central tension: learned methods are fast and amortise over a distribution, but well-tuned classical solvers remain hard to beat on large, irregular, realistic instances.
| Approach | Problem class | Example | Maturity |
|---|---|---|---|
| Pointer / attention construction | Routing (TSP, CVRP) | Kool et al. attention model (2019) | Mature baseline |
| Symmetry-aware RL construction | Routing | POMO (2020) | Mature baseline |
| Learned local search / heatmaps | TSP, large CO | GNN-guided LKH | Active research |
| Learning to branch | MIP (exact) | Gasse et al. GCNN (2019) | Maturing |
| Neural Diving / Branching | MIP (exact) | DeepMind (2020) | Research, production datasets |
| Learned LNS | MIP (heuristic) | DeepMind (2021): 2× to 37.8× better primal gap on three datasets | Research |
| Differentiable optimisation layers | Predict-then-optimise | OptNet, cvxpylayers | Mature for convex |
| Decision-focused learning / SPO+ | Linear and MILP decisions | SPO+, DFL fine-tuning | Active research |
| LLM as optimiser | Black-box, textual | OPRO | Niche |
| LLM natural-language-to-MILP | Modelling layer | Agentic OR pipelines | Emerging |
RL4CO has become the standard open library for routing. Learning inside exact solvers (SCIP for research; Gurobi via callbacks) keeps optimality guarantees and only changes speed. Decision-focused learning trains predictors on downstream regret rather than prediction error, at the cost of solver calls in every training step. OPRO reported up to ~8% gains on GSM8K and ~50% on Big-Bench Hard through prompt optimisation, but does not beat dedicated solvers on real CO.
Rarely, and only under matched conditions. FrontierCO (2025) found that with classical solvers run in fast heuristic mode under the same time budgets, neural solvers remain substantially less effective than LKH-3 on competition-grade and industrial instances, and they degrade out of distribution. Google's AlphaChip placement result (Nature, 2021) remains disputed, with no independent peer-reviewed reproduction as of late 2025. ML's defensible wins are as priors inside classical solvers and as a modelling aid; end-to-end neural construction that outright beats tuned LKH, Concorde or Gurobi has not been convincingly shown as of mid-2026.
Companies and models
The atlas tracks AI4S companies and models by group: core product, competitors, financials and distinctive observations. The directory below keeps the grouping.
| Group | Tracked names (one-line focus) |
|---|---|
| Biology and drugs | AlphaFold 3 (joint structure), EvolutionaryScale (ESM3), Isomorphic Labs (Alphabet drug discovery), Chai Discovery (de novo antibodies), Generate Biomedicines, Nabla Bio, Absci (zero-shot antibodies), Insilico Medicine (own pipeline), Recursion (phenomics), insitro (causal biology), XtalPi (Hong Kong-listed), GenBio AI, Adaptyv Bio (validation as a service), David Baker (protein design) |
| AI drug discovery: picks and shovels | Latent Labs, Alloy (antibody discovery services), Ginkgo (bio foundry), MoleculeMind, BioGeometry, BioMap (xTrimo), Tianwu Tech |
| New pipeline companies, US (around 2023) | Mirador, Superluminal (GPCR), Cellular Intelligence, Manas AI, Accipiter, Archon (antibody cages), Junevity, Xaira (virtual cell plus antibodies), Iambic, Enveda (natural products), NewLimit, Retro, Formation Bio, Schrödinger (software plus carry) |
| New pipeline companies, China (after 2023) | Anew Labs (ByteDance spin-out), Aureka Bio, Pinnacle Medicines (oral peptides), Navi Therapeutics (AI ADCs), Aipher (DrugCLIP), Ligit Pharma, Xunjing Bio (RNA-targeting), ByteRNA (circular mRNA), SciBrunch (KRAS glue), Allink (ADC), Weitao Bio (in vivo CAR-T), Vivatides (extra-hepatic siRNA), Yili Pharma (radiopharma), Lingtai Tianrun |
| Previous generation (2018–22), US | Altos Labs, Treeline, Eikon, Odyssey, Lyell, Kriya, Sana, EQRx (acquired and wound down), Neumora, Monte Rosa, Pathos AI, Valo, Genesis, Ventus |
| Previous generation (2018–22), China | Helixon, METiS (AI formulation), Syneron Bio, StoneWise, MindRank (oral GLP-1), Revir, DualityBio, MediLink, Argo, D3 Bio, Full-Life, Minghui, YolTech (gene editing) |
| China-asset NewCos | Kailera (Hengrui GLP-1), Braveheart, Candid (T-cell engagers; acquired by UCB), Verdiva, Timberlyne, Windward (TSLP), Sentivera |
| Materials and chemistry | GNoME, MatterSim, DP Technology (DeePMD), CuspAI, Orbital Materials (carbon capture), Periodic Labs, Radical AI |
| Mathematics | Axiom (Lean proofs), Project Numina |
| Research agents and autonomous labs | Lila Sciences, Claude Science, OpenScience, MegaRobo, Benyao Tech, Xuanren Tech, Shuye Tech, Jindu Biotech, Hetan AI (C12.ai) |
| Wider AI4S ecosystem (61) | Tool, data and platform providers such as 10x Genomics, Arc Institute, Benchling, Broad Institute, CAS SciFinder, Emerald Cloud Lab, Enamine, FutureHouse, Illumina, Opentrons, Owkin, PacBio, Profluent, Tempus AI, Thermo Fisher, Twist Bioscience, BGI, MGI, WuXi AppTec and GenScript |
Where the value lands, and caveats
Beyond methods, a second axis runs through the atlas: who commercialises AI4S, and how.
- The next growth track. AI4S is positioned as the next fast-growing tier of AI applications after coding and agents. The strategic bet: general-model companies applying frontier LLMs to science scale better than pure-science startups building small bespoke models.
- Two commercialisation routes. Large-model companies either spin out or partner to serve AI4S companies with infrastructure and models, or build the science vertical in-house. Isomorphic Labs (a general model plus domain scientists, funded by Alphabet) is the reference architecture; Anthropic and OpenAI have analogous incubation efforts.
- The platform gap. Elastic GPU scheduling, research data platforms and toolchain integration are the reusable picks-and-shovels layer, and the clearest place to serve AI4S companies rather than own a vertical.
Drug discovery is the most commercially mature vertical. Preclinical failure dominates industry cost (clinical attrition is commonly cited above 90%), and AI's upside is concentrated in target discovery, molecular design and ADMET prediction. Yet as of 2026 no AI4S company, whether Isomorphic Labs, XtalPi or Insilico Medicine, has delivered an approved blockbuster drug. Platform revenue is real; clinical proof is thin.
- State of the art is highly time-sensitive. MLIP and weather leaderboards churn every few months, so read any "best" as "best at publication".
- Source quality is layered. Peer-reviewed results such as AlphaFold 3 and GNoME are high-confidence; many unifying frameworks and taxonomies come from unreviewed arXiv preprints and are one reasonable framing, not settled fact.
- The general-purpose science foundation model is mostly aspiration as of mid-2026.
- This is independent research. Views and figures cite their sources and reliability in the underlying notes; all judgements stand as of their writing date and are not investment or R&D advice.
Adapted from the Spark Science research notes (originally published in Chinese at jimmyresearch.com). Last updated Oct 19, 2026.