AI accelerates drug discovery with quantum mechanics by learning from it. Neural networks trained on quantum chemistry calculations, such as density functional theory (DFT), predict the energies and forces between atoms close to DFT accuracy but orders of magnitude faster, so simulations of how a drug binds its target can run at a scale exact calculations never could.
That matters because the numbers that decide whether a molecule becomes a drug candidate, such as how tightly it binds its target protein, are set by electrons, and computing electrons exactly is slow. This guide explains why drug design needs quantum mechanics, why DFT runs out of road, how neural network potentials such as ANI and MACE learn from open datasets like SPICE and Open Molecules 2025 (OMol25), where they sit next to docking, free energy calculations and AlphaFold-style structure prediction, what quantum computers may add later, and where all of it fails. How AI is used across the rest of the pipeline, from choosing targets to generative chemistry and clinical trials, is covered in the role of AI in accelerating drug discovery. This article is health information about research methods, not medical advice, and it recommends no product or treatment.
Why drug binding is a quantum-mechanical problem
A small-molecule drug works by fitting into a pocket on a target protein and staying there. Whether it stays is a question of energy: the binding free energy, which sums everything that changes when the molecule leaves water and settles into the pocket. Those contributions are electronic. Hydrogen bonds, the way a molecule's electron cloud shifts near a charged side chain (polarization), the energy needed to twist a bond into the shape the pocket demands (a torsion), and any chemistry involving metal ions or bonds that form and break all follow from how electrons arrange themselves. Quantum mechanics is the physics of that arrangement. Solving the electronic Schrödinger equation for a fixed set of atomic positions gives the energy, and the force on each atom follows from how that energy changes as the atoms move.
Routine simulations in drug discovery usually leave the electrons out. A classical force field, also called molecular mechanics (MM), treats atoms as balls with fixed partial charges, joined by spring-like bonds and angles whose parameters are fitted in advance. The MACE-OFF authors note that such empirical force fields have dominated biomolecular simulation for over 50 years and are widely used in drug discovery, but generally lack the accuracy and transferability of first-principles methods. The weak spots are specific: a 2020 preprint on free energy calculations names the modelling of a ligand's internal energy, especially its torsions, as a major limitation of MM force fields.
The stakes are small numbers. Chemists usually treat an error of about 1 kcal/mol as the bar for chemical accuracy, and the SPICE dataset paper uses the same standard. Experiments are not much tighter: when different groups measured the same protein-ligand affinities, their results differed by 0.77 to 0.95 kcal/mol (root-mean-square, depending on how the data was curated), according to a ChEMBL survey summarized in a 2023 benchmark paper. A method that is off by 2 or 3 kcal/mol cannot rank candidates that measurements themselves barely separate.
Some chemistry cannot be approximated cheaply at all. Cytochrome P450 enzymes carry out the oxidation that is the most common way the human body metabolizes drugs, and their iron-containing active site has what chemists call multireference character: no single simple picture of the electrons is adequate. A 2022 PNAS study chose it as a test case for quantum computers. We come back to it below.
Why exact quantum chemistry is too slow for drug design
Quantum chemistry offers a ladder of methods that trade accuracy for speed. At the top, coupled-cluster theory with single, double and perturbative triple excitations, CCSD(T), extrapolated to the complete basis set limit (CBS), systematically approaches the exact solution and is considered a gold standard. It is also, according to the ANI-1ccx authors, often impractical for systems of more than a dozen atoms.
Density functional theory is the workhorse one rung down. It is much faster than coupled cluster, but it needs an empirically chosen density functional, so its results depend on that choice. Its cost still grows roughly with the cube of the number of electrons, which, the OMol25 authors write, stands in the way of routine large-scale screening and severely limits long simulations and structures of more than a few hundred atoms.
Now compare that with what a binding simulation needs. A small protein in water is already large: the MACE-OFF team's test of the protein crambin in explicit water had about 18,000 atoms. Molecular dynamics advances time in steps of one to a few femtoseconds (MACE-OFF's liquid simulations used 1 fs, the OpenMM 8 benchmarks 4 fs), so one nanosecond of motion takes hundreds of thousands to a million steps, each needing energies and forces for every atom. Free energy calculations need many such simulations per compound. Calling DFT at every step is out of reach, which is why the field relied on force fields in the first place.
| Method | What it models | Cost | Typical role in drug discovery |
|---|---|---|---|
| Coupled cluster, CCSD(T) | Electrons, close to exactly | Often impractical beyond about a dozen atoms at the CBS limit | Reference energies for small fragments |
| Density functional theory | Electrons, through an approximate functional | Grows roughly with the cube of the electron count | Training data, fragment and reaction energies |
| Machine-learned potential | Energies learned from DFT or coupled cluster | Grows linearly with atoms; far below DFT, above force fields | Ligand energetics and torsions, ML/MM simulation |
| Classical force field (MM) | Fixed charges, springs and fitted parameters | The cheapest physics-based option | Dynamics and free energy of whole proteins |
| Docking scoring function | A fast empirical score for one pose | Cheap enough to screen 170 million compounds in a 2019 study | Virtual screening of very large libraries |
How machine-learned force fields learn quantum chemistry
A machine-learned interatomic potential (MLIP), also called a neural network potential or machine learning force field, replaces the quantum calculation with a trained function. It takes the elements and 3D positions of the atoms and returns the energy. Forces come from the derivative of that energy with respect to the positions, so the model drops into molecular dynamics code wherever a force field would. What makes it scale is locality: the total energy is written as a sum of atomic contributions, each computed from the atoms within a cutoff radius of a few ångströms. Cost then grows linearly with the number of atoms, where the most efficient quantum methods grow with its cube, as the ANI-1ccx paper points out. For machine-learned potentials in crystals and other materials, see AI in materials science.
Training is the expensive, one-time part. Researchers run DFT, or a more accurate method, on a large set of small molecules and molecular fragments in many distorted shapes, and the network learns to reproduce their energies and forces. Because each atom only sees its neighbourhood, a model trained on small fragments can then be applied to much larger systems, as long as the local environments look like the ones it was trained on.
Three lines of work stand out:
- ANI. ANI-1, first posted in October 2016, built on a modified version of the symmetry functions of Behler and Parrinello. It was trained on molecules of up to 8 heavy atoms from the GDB databases, using hydrogen, carbon, nitrogen and oxygen, and matched reference DFT on molecules of up to 54 atoms with root-mean-square errors as low as 0.56 kcal/mol. ANI-1ccx (2019) was trained on DFT data first, then retrained by transfer learning on a set of CCSD(T)/CBS calculations chosen to span chemical space; it approaches coupled-cluster accuracy on benchmarks that include drug-like torsions and runs billions of times faster than CCSD(T)/CBS. ANI-2x (2020) added sulfur, fluorine and chlorine. Its authors note that those seven elements make up about 90% of drug-like molecules, and that it predicts DFT energies with a speed-up of about a million times.
- MACE. MACE (2022) is an equivariant message-passing network: atoms exchange information with their neighbours in a way that respects rotation, and its higher-order, four-body messages need only two message-passing layers, which keeps it fast. MACE-OFF, first posted in December 2023, applies it to organic molecules. It was trained on a neutral subset of the SPICE dataset covering ten elements, reproduces torsion scans of molecules it has not seen, and describes organic liquids and molecular crystals. Its authors also ran nanosecond simulations of the solvated protein crambin, at about 300,000 steps per day on one NVIDIA A100 GPU.
- Models trained on OMol25. The OMol25 release in 2025 came with baseline models trained on the new data, including eSEN, GemNet-OC, MACE and UMA, and published results for each on the same tests.
The speed-up is relative. The OpenMM 8 paper notes that ML potentials, although much faster than high-level quantum chemistry, are still orders of magnitude slower than most classical force fields, which often makes them impractical for whole macromolecules on their own. That gap is why the hybrid ML/MM simulations described below matter.
The datasets behind the models: SPICE and OMol25
An ML potential is only as good as its reference data, and for years that data was small or narrow. Two open datasets now set the baseline for drug-like chemistry.
| Dataset | Published | Size | Reference method | Coverage |
|---|---|---|---|---|
| SPICE | January 2023, Scientific Data | Over 1.1 million conformations | DFT, ωB97M-D3(BJ)/def2-TZVPPD | Drug-like small molecules, dimers, dipeptides, solvated amino acids; 15 elements, charged and neutral |
| OMol25 | May 2025, arXiv preprint | More than 100 million DFT calculations, systems up to 350 atoms | DFT, ωB97M-V/def2-TZVPD | 83 elements: small molecules, biomolecules, metal complexes, electrolytes; varied charge and spin |
SPICE was designed for one job: training potentials that simulate drug-like small molecules interacting with proteins. It stores forces as well as energies, plus multipole moments and bond orders, and its authors showed that models trained on it reach chemical accuracy across a broad region of chemical space.
Open Molecules 2025 is roughly a hundred times larger. Posted on 13 May 2025 by Meta's FAIR research lab with academic and national-laboratory collaborators, it holds more than 100 million DFT calculations (over 140 million in the paper's March 2026 revision) at one consistent, high level of theory, and took about 6.6 billion CPU core-hours by the authors' count. Its biomolecular part is aimed squarely at binding. Protein pockets with their bound ligands were cut out of experimental structures, and random drug-like molecules from the ZINC20, ChEMBL and GEOM databases were docked into pockets to add more. The authors also recomputed existing datasets, such as ANI-2x and SPICE2, at the same level of theory, so models are not trained on a patchwork of methods. The evaluations that ship with it include protein-ligand interaction energies and ligand strain: the energy difference between a ligand's bound shape and its most stable shape.
Where ML potentials fit: docking, free energy and ML/MM
In a structure-based project, computational methods work as a funnel, and each stage is slower and more accurate than the one before.
Docking places each molecule in the pocket and scores the pose with a fast function. It is cheap enough for enormous libraries. In a 2019 Nature study, a team docked 170 million make-on-demand compounds against two targets, the enzyme AmpC β-lactamase and the D4 dopamine receptor, then synthesized and tested 44 and 549 of the top-ranked molecules. It worked: it found new families of inhibitors, and crystal structures confirmed the predicted poses for AmpC. The same study also shows what a docking score is: for the D4 receptor, hit rates fell almost steadily as scores got worse, so a score shifts the odds rather than measuring an affinity.
Free energy calculations do the precise part. Free energy perturbation (FEP) and related alchemical methods simulate one ligand being gradually transformed into another, in the pocket and in water, to compute their difference in binding free energy. The 2023 benchmark paper, by scientists at Schrödinger testing FEP+ from the company's software suite, calls FEP the most consistently accurate way to predict relative binding affinities. Across 13,732 predictions, 57.5% were within 1 kcal/mol of experiment and 12.9% were off by more than 2 kcal/mol. The authors conclude that with careful preparation of protein and ligand structures, FEP can match the reproducibility of experiments. With standard MM force fields, these methods deliver 1 to 2 kcal/mol accuracy in well-behaved protein-ligand systems, according to the 2020 preprint.
ML/MM is where machine-learned potentials enter the workflow. The idea mirrors QM/MM: model the ligand, where force fields are weakest, with an ML potential, and the protein and water with a classical force field. OpenMM 8, the open-source molecular dynamics toolkit (published in December 2023), made this a few lines of code. Any PyTorch model can supply forces, and a mixed system is built from an ordinary force field plus a list of the atoms to treat with ML. At publication, it offered ANI-1ccx and ANI-2x as ready-made potentials.
The payoff shows up in the numbers. In the 2020 preprint, researchers corrected standard MM free energy calculations with an ANI-2x/MM potential in a post-processing step. On a benchmark series of inhibitors of the kinase TYK2, the error in absolute binding free energies fell from 0.97 to 0.47 kcal/mol, for a small additional computational cost. It is one target and one ligand series, but it shows where the gain comes from: better energetics for the ligand itself.
Where AlphaFold-style structure prediction fits
Every method above starts from a 3D structure of the target and a plausible pose for the ligand. When no experimental structure exists, deep learning now supplies one. AlphaFold 3, published in Nature on 8 May 2024, predicts the joint structure of complexes of proteins, nucleic acids, small molecules, ions and modified residues. On the PoseBusters benchmark, its authors report far greater accuracy for protein-ligand interactions than state-of-the-art docking tools. AlphaFold's wider place in AI-driven science is covered in AI-enabled scientific discoveries.
Two limits keep physics in the loop. AlphaFold 3 predicts static structures like those in the Protein Data Bank, not how a complex moves in solution, and like other generative models it can invent plausible-looking structure in disordered regions. Newer co-folding models reach toward energies too. The open-source Boltz-2 (2025, MIT licence) predicts binding affinity alongside structure, and its developers describe it as the first deep learning model to approach the accuracy of FEP while running 1,000 times faster. That is the developers' own claim, on their own benchmarks.
In practice the tools stack rather than compete. An experimental or predicted structure provides the pocket and a starting pose; docking or co-folding narrows the list; FEP or ML/MM simulation tests the energetics and motion of the short list before anything is made.
Will quantum computers speed up drug discovery?
Not yet. First, a distinction: everything above runs on ordinary computers. Quantum mechanics in drug discovery today means quantum chemistry software on CPUs and GPUs. Quantum computers are a separate technology that may one day run some of those calculations.
Their promise is the part classical methods handle worst: molecules whose electrons are strongly correlated, such as metal-containing active sites. The 2022 PNAS study, by Google Quantum AI with Boehringer Ingelheim and others, took cytochrome P450 as its test case. It estimated that computing the ground-state energy of the largest model of a key P450 intermediate would need about 4.6 million physical qubits and 73 hours at a physical error rate of 0.1%, or about 500,000 qubits and 25 hours if error rates fell to 0.001%. Its conclusion was careful: simulating these enzymes at a large enough scale has the potential to be a problem where quantum computers beat classical ones.
Set that against the hardware as announced:
- December 2024: Google's Willow chip, with 105 qubits, showed errors falling as more qubits were used for error correction.
- June 2025: IBM set out a plan to deliver Starling, a fault-tolerant machine running 100 million gates on 200 logical qubits, by 2029.
- October 2025: Google reported Quantum Echoes on Willow, which it calls the first verifiable quantum advantage, running 13,000 times faster than the best classical algorithm on a leading supercomputer. In a separate proof of principle with UC Berkeley, it studied two molecules of 15 and 28 atoms, and the results matched traditional NMR. Google suggests quantum-enhanced NMR could one day help show how drug candidates bind their targets.
Note
Treat quantum computing for drug discovery as research. The P450 estimate needs millions of physical qubits; Willow has 105. None of these sources reports a drug candidate found or optimized on a quantum computer.
The limits: training data, reference methods and the lab
Every stage in this chain can be wrong in a way the next stage does not see, and the failure modes are well documented.
A local model is blind beyond its cutoff. Many ML potentials let atoms interact directly only within a cutoff radius, typically around 6 Å, while the interaction between two charges fades slowly with distance, as 1/r. OMol25 tests this explicitly, with separate short-range and long-range evaluations, and the MACE-OFF authors state that without explicit long-range interactions their model applies only to neutral, non-radical and non-reactive systems.
- A model is capped by its reference. An ML potential trained on DFT inherits DFT's errors. In the MACE-OFF paper, DFT torsion barriers differed from coupled-cluster values by 0.2 kcal/mol on average, which the authors describe as the best their model could possibly reach on that test.
- Accuracy drops outside the training data. ANI-2x covers seven elements and ANI-1ccx four. On OMol25's out-of-distribution test, models did better on neutral organic molecules and biomolecules than on electrolytes and metal complexes.
- Preparation matters. In the 2023 FEP benchmark, adding both protonation states of the ligands for one target (TNKS2) cut the error from 2.10 to 1.60 kcal/mol, with no change to the method.
- Structures are predictions. AlphaFold 3's poses are static, and it can hallucinate structure in disordered regions.
- Experiments disagree too. With independent measurements differing by 0.77 to 0.95 kcal/mol, no method can be judged more precisely than the data it is compared with.
Important
Only the lab decides. Every result in the 2019 docking study is a compound that was synthesized and tested, and for AmpC, crystal structures confirmed the predicted poses. Everything upstream narrows what gets made; it does not replace making it.
Before trusting an ML potential on a project, work through these checks:
- Check the domain. Compare the elements, charge states, metals and any reactive chemistry in your molecules with what the model was trained on.
- Spot-check against the reference. Recompute a sample of your own structures at the training level of theory and compare energies and forces.
- Benchmark retrospectively. Run the full workflow on a compound series you have already measured before using it to choose what to make.
- Record provenance. Store the model version, training data and reference method with every prediction, so a result can be traced and repeated.
The data engineering around the models
Much of the work in this field is data work. OMol25 alone took about 6.6 billion CPU core-hours, and its value depends on bookkeeping: one level of theory throughout, quality checks on every calculation, and test splits held out by composition so a model's error on unfamiliar chemistry is measured rather than assumed. Any team building on data like this faces the problems of every large machine learning dataset: versioned data, recorded provenance and evaluations that run before a new model is trusted.
Those are general data and machine learning engineering problems, not chemistry. We do not work in drug discovery or molecular modelling. Our data platform service builds pipelines with tests, lineage and access control, and serves governed datasets for ML and AI. For teams putting language models into their own workflows, our AI and automation service tests every model, prompt or source change against an evaluation set of real cases before it goes live. More on the subject is under the AI topic.
