AI in Chemistry: Molecular Simulation and the Future of Drug Design
Photo: Pavel Danilyuk / Pexels
In 2020, DeepMind's AlphaFold2 solved a problem that had stumped structural biologists for fifty years: predicting the three-dimensional shape of a protein from its amino acid sequence alone. Within two years the AlphaFold Protein Structure Database had released predicted structures for over 200 million proteins, essentially every protein encoded by every sequenced organism on Earth. That single event reshaped drug discovery, a field that depends on knowing a protein's shape well enough to design a molecule that fits into it. What used to take a graduate student a year of X-ray crystallography now often takes seconds of GPU time, and the ripple effects have reached molecular dynamics, virtual screening, and the design of entirely new drug candidates from scratch.
Why protein shape decides whether a drug works at all
Every protein is a chain of amino acids that folds into a specific three-dimensional shape, and that shape is almost entirely what determines its biological job. Two proteins built from identical sequences but folded differently would behave like two different machines built from the same parts bin. Biochemist Christian Anfinsen showed in the 1960s that, for most proteins, sequence alone determines the folded structure under normal physiological conditions, a principle now called Anfinsen's thermodynamic hypothesis. It won him the 1972 Nobel Prize in Chemistry and it is the reason a computer, in principle, should be able to predict a fold from a sequence without ever touching a test tube (Nobel Prize official site).
Determining protein structures experimentally, whether by X-ray crystallography, cryo-electron microscopy, or nuclear magnetic resonance (NMR) spectroscopy, takes months to years per protein and requires specialized instruments most labs cannot afford to run continuously. The Protein Data Bank (PDB), the field's central repository, accumulated roughly 180,000 to 200,000 experimentally solved structures across five decades of collective effort worldwide (RCSB PDB statistics). AlphaFold's 2022 database update released more than 200 million predicted structures in a single event, over a thousand times the entire experimental record, covering nearly every catalogued protein sequence known to science (EMBL-EBI AlphaFold DB announcement).
Drug discovery depends on this kind of structural knowledge because most drugs work by physically binding to a protein and either blocking or activating it. A cancer drug that inhibits a mutated kinase, for example, has to fit into a pocket on that kinase's surface the way a specific key fits a specific lock. Without a structure for the pocket, chemists are designing keys for a lock they have never seen. AlphaFold did not eliminate that problem, but it made a first-draft lock available for almost any protein a researcher might target, which is why pharmaceutical companies moved so quickly to build entire discovery pipelines around it.
How AlphaFold actually predicts a fold
AlphaFold2's core is a transformer-based neural network, the same broad architecture family behind large language models, but trained on a completely different kind of data: multiple sequence alignments (MSAs). An MSA is a stacked list of evolutionarily related protein sequences from different species, lined up so that corresponding positions sit in the same column. Positions that consistently mutate together across millions of years of evolution are usually physically close to each other in the folded structure, because a mutation at one site often forces a compensating mutation nearby to keep the protein stable. That co-evolution signal is a kind of fossil record written directly into the sequence data.
Think of an MSA the way a detective treats a stack of eyewitness accounts of the same event. No single witness has the full picture, but patterns that repeat across many independent accounts are usually true. AlphaFold's Evoformer module reads that stack of sequences with an iterative attention mechanism, repeatedly refining which positions are likely to be neighbors in 3D space, before a structure module converts those relationships directly into atomic coordinates using geometric reasoning rather than simple curve-fitting (Jumper et al., Nature, 2021).
AlphaFold3, published in Nature in 2024, extended the same underlying diffusion-based framework beyond proteins to DNA, RNA, small molecules, ions, and chemically modified residues, allowing it to predict the joint structure of a protein bound to a candidate drug molecule in a single pass. On the PoseBusters benchmark, a standard test for how well a model places a small molecule into a binding pocket, AlphaFold3 was roughly 50% more accurate than the best prior docking tools, without needing any experimental structural information as a starting point (Abramson et al., Nature, 2024).
Is AlphaFold's structure prediction accurate enough to use for drug design?
For well-studied protein families with deep evolutionary sequence data, AlphaFold predictions typically land within one to two Ångstroms of experimental structures, which is precise enough to guide medicinal chemistry. Accuracy drops for intrinsically disordered regions, membrane proteins, and complexes that undergo large conformational shifts on binding. Researchers treat predicted structures as high-quality starting hypotheses, cross-checking against any available experimental data before committing to a design (Abramson et al., Nature, 2024).
Molecular dynamics: watching atoms move instead of just where they sit
A static structure only tells part of the story. Proteins wobble, breathe, and change shape as they bind partners, and drugs have to work inside that motion, not against a frozen snapshot. Molecular dynamics (MD) simulation models how every atom in a system moves over time according to Newton's laws of motion, essentially running a physics simulation of a protein and its surrounding water molecules frame by frame.
The accuracy of an MD simulation depends entirely on its force field, the mathematical recipe describing how atoms attract, repel, and bond. Classical force fields such as AMBER, CHARMM, and GROMOS use simple analytical terms, harmonic springs for bond stretching, Coulomb's law for electrostatics, and Lennard-Jones potentials for van der Waals forces. These run fast, simulating millions of atomic movements per second on a single GPU, but they miss quantum mechanical effects like electron polarization and charge transfer that matter for tight binding interactions. Quantum mechanical methods such as density functional theory (DFT) capture those effects with high fidelity but become computationally impractical for systems larger than a few hundred atoms or simulation windows longer than picoseconds, which is billionths of a second, far too short to watch a drug actually find and settle into its binding site.
How do machine learning force fields close that gap?
Machine learning force fields (MLFFs) train neural networks on large datasets of DFT-computed energies and forces across thousands of atomic configurations, then use the trained network as a fast stand-in for the expensive quantum calculation. Models such as ANI, developed at the University of North Carolina, and equivariant graph networks like NequIP from MIT and MACE from the University of Cambridge, learn a smooth mathematical surface that approximates the true quantum energy landscape. Once trained, they evaluate new atomic configurations thousands of times faster than DFT while staying within one to two kilocalories per mole of the quantum reference, an error margin small enough for reliable free energy calculations.
MACE-OFF, a transferable force field released by the Cambridge group in 2024 and 2025, extended this approach specifically to organic molecules relevant to drug chemistry, and has been used to compute free energy surfaces in explicit solvent and simulate the folding dynamics of peptides and fully solvated proteins (Kovács et al., Journal of the American Chemical Society, 2025). This is a genuinely different tool from graph neural networks used purely for predicting a molecule's static properties, like solubility or toxicity from its structure alone; MLFFs are built to run a physically realistic simulation forward in time, frame by frame, the same way a flight simulator models an aircraft's actual aerodynamics rather than just estimating its top speed from a spec sheet. For a deeper look at how graph-based networks handle molecular property prediction specifically, see our explainer on graph neural networks in chemistry.
MLFFs are not without limits. A 2024 community benchmarking effort, the TEA Challenge, compared MACE, SO3krates, sGDML, and other architectures across molecules, surfaces, and periodic materials and found meaningful differences in stability and accuracy depending on the system and simulation length, with some models drifting or becoming unstable over long trajectories (TEA Challenge, npj Computational Materials, 2024). The field is advancing quickly, but "near-DFT accuracy" is a claim that has to be checked system by system, not assumed universally.
Virtual screening: testing billions of molecules without a single test tube
Traditional drug discovery often starts with high-throughput screening (HTS), physically synthesizing and testing hundreds of thousands of compounds in laboratory assays against a target protein. Each compound tested this way costs real money and real time, and the chemical space of possible drug-like molecules is estimated to contain more candidates than there are atoms in the observable universe, so even the largest physical screening campaigns sample only a sliver of what is theoretically possible.
Virtual screening replaces some of that physical testing with computation. A candidate molecule is computationally docked into a protein's binding site, and a scoring function estimates how well it fits and how strongly it would bind. Classic docking programs like AutoDock and Glide use empirically derived scoring functions built from decades of accumulated binding data. Machine learning scoring functions, trained on databases such as PDBbind, which contains roughly 20,000 experimentally measured protein-ligand binding affinities, have measurably improved on those classic methods by learning subtler patterns in how atomic-level interactions translate into binding strength (PDBbind database, Wang et al.).
Isomorphic Labs, DeepMind's drug discovery spinout, has built its pipeline directly on AlphaFold3-derived structures combined with proprietary scoring and design models. In February 2026 the company released IsoDDE, a unified drug-design engine that combines structure prediction, ligand binding prediction, affinity prediction, and antibody interaction modeling into a single pipeline, and reported that it roughly doubles AlphaFold3's accuracy on the hardest ligand-binding cases, those with less than 20% sequence similarity to anything in the training data (Isomorphic Labs). The company raised $600 million in its first outside financing round in March 2025 and has active drug-design partnerships with Eli Lilly and Novartis, with its first fully AI-designed candidates expected to enter first-in-human clinical trials by the end of 2026 (Clinical Trials Arena, 2026).
Free energy perturbation (FEP), a rigorous physics-based method that uses MD simulation with enhanced sampling to compute relative binding strength between similar compounds, has become a workhorse in industry precisely because it is quantitatively accurate rather than just a ranking heuristic. Schrödinger's FEP+ implementation has demonstrated routine prospective accuracy of roughly 0.5 to 1 kilocalorie per mole for congeneric compound series, close enough to reliably rank which of several similar candidate molecules will bind tightest before any of them are synthesized (Schrödinger, Inc. published validation studies). Machine learning force fields are increasingly used to accelerate the underlying MD sampling in these calculations, cutting compute cost without sacrificing the physics-based rigor that makes FEP trustworthy to medicinal chemists in the first place.
Generative chemistry: designing a molecule instead of just searching for one
Screening, however large the library, is still fundamentally a search problem: you are picking the best candidate out of molecules someone already thought to make or catalog. Generative molecular design flips that around. A researcher specifies a target protein's binding pocket and desired properties, and a model generates novel molecular structures that were never in any database to begin with.
Graph-based generative models such as REINVENT, developed at AstraZeneca, and the Junction Tree Variational Autoencoder from MIT, learn a statistical distribution over valid, drug-like molecular graphs and then sample new structures from that distribution, nudged toward whatever property the chemist wants optimized. More recent diffusion-based approaches, including DiffSBDD, work directly in three dimensions: the model starts from noise scattered inside the shape of a target binding pocket and progressively denoises both the atom types and their spatial coordinates until a coherent, pocket-shaped molecule emerges. It is roughly analogous to how an image-generation diffusion model starts from static and refines it into a photograph, except the "photograph" here has to obey the rules of chemical valence and physically fit inside a specific three-dimensional cavity. Because the model operates in a rotation- and translation-invariant way, called SE(3)-equivariance, the same trained network can propose ligands for pockets it has never specifically seen before, and can also be steered to perform narrower tasks like growing a fragment, hopping between chemical scaffolds, or designing a linker between two known binding fragments (Schneuing et al., Nature Computational Science, 2024).
The most concrete proof of concept so far is rentosertib (also known as INS018_055), a TNIK inhibitor for idiopathic pulmonary fibrosis (IPF), a progressive and life-threatening scarring disease of the lungs. Insilico Medicine's generative AI platform, Pharma.AI, was used to both identify the TNIK target and design the molecule itself, taking the program from target identification to an Investigational New Drug (IND) filing in roughly 18 months, notably faster than the four-to-five-year timeline typical of conventional discovery programs. On June 3, 2025, Insilico and collaborators published Phase IIa results in Nature Medicine, the first peer-reviewed clinical validation of a drug where both target discovery and molecular design used generative AI. In the 71-patient, double-blind, placebo-controlled trial, patients on the 60 mg once-daily dose showed a mean improvement in forced vital capacity (FVC), a standard measure of lung function, of +98.4 milliliters over twelve weeks, compared with a mean decline of -20.3 milliliters in the placebo group, alongside dose-dependent reductions in profibrotic proteins including COL1A1 and MMP10 (Insilico Medicine / Nature Medicine, June 2025).
ADMET prediction, absorption, distribution, metabolism, excretion, and toxicity, is the other half of generative design that determines whether a molecule that binds well in a simulation can ever become a medicine. A compound might dock perfectly into a target's binding pocket and still fail as a drug if it is not absorbed through the gut, is broken down by liver enzymes before it reaches its target, or triggers cardiac toxicity by blocking the hERG potassium channel. Machine learning models trained on large proprietary and public pharmacology datasets now predict oral bioavailability, hERG inhibition risk, cytochrome P450 (CYP) enzyme interactions, and blood-brain barrier penetration early enough in the pipeline to steer medicinal chemists away from dead ends before a single milligram is synthesized.
Where the AI-accelerated pipeline actually speeds things up, and where it doesn't
It helps to be specific about which stages of drug discovery AI has genuinely compressed and which remain stubbornly slow, because marketing language around "AI drug discovery" often blurs the two together.
| Discovery stage | Traditional timeline | AI-accelerated approach | What AI changes |
|---|---|---|---|
| Target structure determination | Months to years per protein (X-ray, cryo-EM, NMR) | AlphaFold2/3 structure prediction | Seconds to minutes of compute; near-total coverage of known sequences |
| Hit identification | Months of high-throughput physical screening | ML-scored virtual screening of billions of compounds | Weeks instead of months; far larger search space per dollar |
| Lead optimization | Iterative synthesize-test-redesign cycles, months per cycle | Generative design + FEP-based affinity ranking | Fewer wasted synthesis cycles; better-prioritized candidates |
| ADMET profiling | Animal and in vitro studies, weeks to months | ML ADMET prediction models | Early triage before costly wet-lab testing |
| Preclinical to IND filing | Typically 4–5 years | Insilico's rentosertib program: approximately 18 months | Compressed timeline for at least one validated case |
| Clinical trials (Phase I–III) | 6–7 years on average | Largely unchanged | AI does not accelerate human biology or regulatory review |
| Regulatory approval | 10–15 years total, target to market | Still in progress for AI-native drugs | No fully AI-designed drug has reached approval as of mid-2026 |
Table: Where AI compresses drug discovery timelines, and where it doesn't (illustrative ranges compiled from cited sources above).
The pattern in that table is consistent: AI collapses the computational bottlenecks, structure determination, virtual screening, and molecular design, dramatically, because those steps are fundamentally information-processing problems that scale with compute. It has made comparatively little difference to the biological bottlenecks, because a Phase III clinical trial still needs the same number of human patients, the same follow-up period to observe outcomes, and the same regulatory review regardless of how the drug was designed. The rate-limiting step in drug discovery has shifted from "can we find a plausible candidate" to "can we synthesize, test, and clinically validate it fast enough to keep up."
What people get wrong about "AI-designed" drugs
The phrase "AI-designed drug" gets used loosely, and the loose usage creates real confusion about how much of the process AI actually touched. There is a meaningful difference between a drug where AI proposed the exact molecular structure from scratch, a drug where AI merely helped prioritize which of several human-proposed candidates to synthesize first, and a drug where AI was used only to predict a target's structure while a chemist designed the actual molecule by hand. Companies have commercial incentives to describe all three as "AI-discovered," and press coverage rarely draws the distinction.
Has any fully AI-designed drug been approved for patients?
As of mid-2026, no drug whose molecular structure was originally generated by an AI model has received full regulatory approval from the FDA or EMA. Several are in active clinical trials: Insilico Medicine's rentosertib has completed Phase IIa for idiopathic pulmonary fibrosis, and Isomorphic Labs expects its first AlphaFold3-derived oncology and immunology candidates to enter first-in-human trials by the end of 2026. Approval, if it comes, is realistically still years away given standard Phase II and Phase III timelines (Nature Medicine, June 2025; Clinical Trials Arena, 2026).
A second common misconception is that AlphaFold "solved" drug discovery the way it solved protein folding. Structure prediction answers "what does the target look like," but binding affinity, selectivity against related off-target proteins, pharmacokinetics, and toxicity are separate problems that a static structure alone cannot answer. AlphaFold3 helps with some of these, particularly protein-ligand complex prediction, but the overall discovery pipeline still requires molecular dynamics, generative design, and ADMET prediction as distinct downstream steps, each with its own error rates and failure modes.
A third misconception treats simulation accuracy as a solved problem across the board. Machine learning force fields report near-DFT accuracy, but that accuracy is typically demonstrated on chemical systems similar to their training data. Benchmark efforts like the TEA Challenge have shown that different MLFF architectures can diverge meaningfully in stability and accuracy once pushed outside familiar territory, which is exactly the territory novel drug candidates tend to occupy by design (TEA Challenge, npj Computational Materials, 2024). Extrapolation, not interpolation, is where these models are weakest, and drug discovery is inherently an extrapolation problem: the whole point is to find a molecule that does not already exist.
FAQ
Q: Is AlphaFold's protein structure prediction accurate enough to use for drug design? For globular proteins with well-represented sequence families, AlphaFold predictions are highly accurate, typically within one to two Ångstroms RMSD of experimental structures. For intrinsically disordered regions, membrane proteins, and protein complexes with large conformational changes on binding, accuracy is lower. Researchers validate AlphaFold structures against any available experimental data and treat predicted structures as starting points rather than ground truth (Abramson et al., Nature, 2024).
Q: What is ADMET and why does it matter in drug design? ADMET stands for absorption, distribution, metabolism, excretion, and toxicity. A compound can bind its target perfectly in a simulation and still be useless as a medicine if it is not orally absorbed, is metabolized before reaching the target, or is toxic to the liver or heart. Predicting ADMET properties early in the pipeline saves substantial time and cost by filtering out doomed candidates before synthesis.
Q: How do machine learning force fields maintain near-quantum accuracy? They train a neural network on datasets of density functional theory (DFT)-computed energies and forces across thousands of atomic configurations, learning a smooth interpolation of the underlying quantum energy surface. Evaluated on configurations similar to the training data, models like MACE-OFF reproduce DFT-level results within roughly one to two kilocalories per mole, precise enough for free energy calculations (Kovács et al., JACS, 2025).
Q: Has any fully AI-designed drug been approved by regulators? As of mid-2026, no fully AI-designed drug has received full regulatory approval, though several candidates are in active clinical trials. Insilico Medicine's rentosertib has published positive Phase IIa results for idiopathic pulmonary fibrosis, and Isomorphic Labs' first AlphaFold3-derived candidates are expected to enter first-in-human trials by the end of 2026.
Q: What is the difference between molecular dynamics simulation and virtual screening? Molecular dynamics simulates how atoms physically move over time according to the laws of motion, useful for studying how a drug binds, unbinds, or how a protein changes shape. Virtual screening is a faster, more approximate process that scores how well large numbers of candidate molecules might fit a target's binding site, used to narrow millions or billions of candidates down to a shortlist worth simulating or synthesizing in detail.
Q: Why can't machine learning force fields simply replace all classical force fields? Machine learning force fields are typically trained on specific classes of molecules and can lose accuracy or become numerically unstable when applied to chemistry outside their training distribution, exactly the kind of novel chemical space drug discovery is trying to explore. Classical force fields remain useful because their behavior, even when less accurate, is predictable and well understood across a much broader range of systems.
Q: How is generative molecular design different from graph neural network property prediction? Graph neural networks for molecular property prediction take an existing molecule and estimate a property, such as toxicity or solubility. Generative models work in the opposite direction: starting from a target's binding pocket or a desired property profile, they construct new molecular structures atom by atom that did not previously exist in any database. Our explainer on graph neural networks in chemistry covers the property-prediction side of this relationship in more depth.
Q: What open-source tools are available for AI-based drug design? RDKit (cheminformatics), OpenMM (molecular dynamics), the Open Drug Discovery Toolkit (ODDT), DeepChem (machine learning for drug discovery), and Meta's ESM (Evolutionary Scale Modeling) protein language models are all freely available. AlphaFold's model weights and code are open-source and hosted publicly on GitHub by DeepMind and EMBL-EBI.
Q: Why does AI speed up early drug discovery but not clinical trials? Structure prediction, virtual screening, and molecular design are computational problems that scale with available compute, so faster hardware and better models directly shorten them. Clinical trials measure how a drug behaves in human bodies over months or years, a biological and regulatory process that cannot be sped up by additional computing power, which is why overall time from target discovery to market approval remains dominated by the clinical and regulatory stages regardless of how fast the design phase becomes.
The takeaway
AI has not shortened the drug discovery pipeline evenly. It has collapsed the computational front end, structure prediction, virtual screening, and generative molecular design, from a process measured in years to one measured in weeks or less, while leaving the biological and regulatory back end, clinical trials and approval, essentially untouched. Rentosertib's Phase IIa results are the first hard clinical evidence that an AI-generated molecule against an AI-identified target can move a human lung disease biomarker in the right direction, which is a genuinely different claim than simply optimizing a docking score. The next few years will show whether that result generalizes, or whether idiopathic pulmonary fibrosis turned out to be an unusually favorable test case for a still-young technology.
Read next:
Follow new articles
Add SciHub101 to your feed reader to see new science and technology stories.
About the Author
SciHub101 is an independent science and technology publication founded and edited by Buğra Sözeri, a software engineer and app developer based in Türkiye. AI tools assist with drafting and research. Source documentation in the existing archive is uneven, and articles are being reviewed and corrected. The Editorial Team byline identifies editorial responsibility; it does not certify specialist or independent human review. Where we compare products, anything built by our own team is labelled as such — see the editorial guidelines and About page for how corrections and disclosures work.
Related Articles
Food Allergies vs. Food Intolerance: The Science Explained
They both make you feel terrible after eating certain foods. But a food allergy and a food intolerance are different biological mechanisms, and confusing them can be dangerous.
How Preservatives Keep Food Safe: The Chemistry Explained
Before refrigerators, salt and vinegar kept people alive. Today, a few molecules in the parts-per-million range prevent millions of cases of food poisoning a year. Here's how.
Graph Neural Networks in Chemistry: Reading Molecules as Math
Molecules are naturally graphs—atoms as nodes, bonds as edges. Graph neural networks exploit this structure to predict chemical properties with striking accuracy.
How Machine Learning Is Accelerating Materials Discovery
AI is compressing decades of trial-and-error chemistry into weeks, helping scientists discover new compounds with unprecedented precision.
AI-Driven Scientific Discovery: Can Machines Generate Hypotheses?
Beyond automating experiments, AI systems are beginning to propose novel scientific hypotheses—a capability that could fundamentally change how science is done.
Deep Learning and Microscopy: Seeing Materials at the Atomic Scale
AI is giving electron microscopes superhuman vision—resolving atomic structures, identifying defects, and automating analyses that once took weeks.