Few numbers in computational chemistry get repeated as often, or as loosely, as the estimate that there are roughly 10⁶⁰ possible drug-like molecules, a figure that traces back to a 1996 back-of-envelope calculation and has been revised up and down by other groups ever since, landing anywhere from 10²³ to 10⁶⁰ or higher depending on the assumptions used. Precise or not, it points at something real: chemists have synthesized and characterized on the order of 10⁸ compounds in total, a vanishing fraction of what could plausibly work as a drug. AI chemical space exploration is the industry's shorthand for the attempt to close that gap computationally, using artificial intelligence (AI) to propose, filter, and prioritize molecules that lie well outside the regions medicinal chemistry intuition and existing screening libraries would ever have suggested.
It sits within DDN's coverage of generative AI in molecular design, and follows directly from the discussion of diffusion models for drug design, which cover one specific way of generating candidate structures. This piece steps back to the broader question those methods are trying to answer: not just how to generate a molecule, but how to search a space too large to ever fully map, and how what gets generated actually becomes something a chemist can order and test. It connects to the guide to AI across drug discovery for the wider picture of where this fits in a drug program.
How large is chemical space?
The 10⁶⁰ figure comes from a 1996 estimate by Bohacek and colleagues, based on a maximum of 30 carbon, nitrogen, oxygen and sulfur atoms; other researchers have since produced their own estimates using different rules and constraints, some considerably lower, some higher still. The range itself is the useful part: however it is calculated, drug-like chemical space is large enough that no exhaustive enumeration will ever be practical, and the number matters mainly as an order-of-magnitude reminder of how much of it goes unexamined.
A more concrete anchor comes from actually enumerating a bounded slice of that space. The GDB-17 database systematically generated 166 billion organic small molecules containing up to 17 heavy atoms, a far smaller molecular size window than a typical drug but still large enough to illustrate the scale problem: even a rigorously enumerated, size-limited corner of chemical space runs into the hundreds of billions of structures. Commercial screening collections and corporate compound libraries, by comparison, typically hold in the tens of millions. Medicinal chemistry intuition, built from the compounds a chemist or a company has actually made before, draws from a smaller and more biased sample still.
What generative models actually explore
Generative AI does not search chemical space the way a database query searches a table. Most approaches instead learn some form of internal representation of molecular structure from a large training set of known compounds, then sample or optimize within that learned representation rather than enumerating candidates directly. An early and influential example, from Aspuru-Guzik's group, converted molecules into a continuous, differentiable representation that let a computer move smoothly between structures and use gradient-based optimization to move toward more desirable ones, an approach that later work extended with genetic-algorithm-style exploration methods built on more chemically robust molecular encodings.
Diffusion models, covered in detail in the companion article on 3D molecular generation, take a related but distinct approach: rather than optimizing within a learned latent space, they learn to reverse a noising process, gradually assembling a plausible 3D structure from randomness. Language-model-style approaches treat molecules as strings, usually in SMILES or SELFIES notation, and generate them the way a text model generates a sentence. All of these methods share the same underlying limitation: they explore only the regions of chemical space that resemble their training data closely enough to generalize, which means the training set itself quietly sets the boundaries of what counts as "beyond intuition."
A separate, increasingly discussed strategy skips the pure computation and closes the loop with real experiments: so-called self-driving laboratories pair a generative or optimization algorithm with automated synthesis and testing hardware, so each round of proposed molecules is actually made and measured, and those results feed directly back into the next round of proposals. This trades some of the speed of a purely in-silico search for grounding in real chemistry rather than a model's internal, and sometimes mistaken, sense of what should work. It remains far more capital-intensive to set up than a software-only generative pipeline, which is part of why most drug discovery programs still treat computational generation and physical testing as separate, sequential stages rather than a single continuous loop.
Drug-likeness constraints
An unconstrained generative model will happily produce structures that are chemically valid on paper but useless as drug candidates: too large, too greasy, too reactive, or simply impossible to make. Most drug discovery programs still lean on some version of Lipinski's rule of five (molecular weight under 500, calculated LogP under 5, no more than five hydrogen bond donors and ten acceptors) as a first-pass filter, alongside newer, more nuanced developability and synthetic accessibility scores.
Building those constraints into the generation process, rather than filtering afterward, is one of the more active areas of the field. Multi-parameter optimization sets up several competing property targets at once: potency, solubility, metabolic stability, selectivity, since a molecule that is optimal on one axis is often unusable on another. Constraining generation up front tends to produce a higher proportion of viable candidates than generating broadly and filtering hard afterward, though the tradeoff is a narrower, more conservative search of the space in exchange for a higher hit rate.
Synthetic accessibility adds a further constraint that property filters alone do not capture. A molecule can satisfy every drug-likeness rule and still have no practical route to a bench chemist's hood, whether because it requires an exotic reagent, an unstable intermediate, or a sequence of steps too long to be worth attempting. Synthetic accessibility scores attempt to estimate this directly from structure, usually by comparing a candidate against the kinds of substructures and reactions known to be routine, and are now commonly built into the scoring function a generative model optimizes against, alongside potency and selectivity, rather than treated as a separate downstream check.
Building AI-generated virtual libraries
Two distinct approaches have emerged for turning generative methods into libraries a chemist can actually work from. The first pairs generative algorithms directly with property prediction and scoring in a closed loop. Insilico Medicine's Chemistry42 platform, for instance, runs an ensemble of dozens of generative algorithms in parallel, scoring and filtering the output with both machine-learned and physics-based models before a chemist ever sees a shortlist; it has been licensed by more than 20 pharmaceutical companies for internal and external programs.
The second approach builds ultra-large libraries from validated chemistry rather than from free-form generation. Enamine's REAL Space, the best-known example, now spans more than 94 billion make-on-demand molecules assembled from roughly 200,000 in-stock building blocks combined through some 169 validated parallel-synthesis reactions, with a further, non-enumerated extension pushing into the trillions. It is not generative in the deep-learning sense, but it solves the same underlying problem from the other direction: instead of asking a model to imagine a molecule and then checking whether it can be made, it starts from chemistry known to work and combinatorially expands outward, at a scale AI-driven virtual screening tools are increasingly built to search directly.
From virtual library to synthesizable candidates
A molecule that exists only as a structure on screen is not yet useful. Closing that gap, from generated or selected structure to a physical sample in a well plate, has become its own AI application. PostEra's platform was built specifically around retrosynthesis prediction: given a target structure, it proposes a synthetic route and flags where that route runs into trouble, closing the loop between what a model wants to make and what a lab can actually produce that week. Enamine's make-on-demand model addresses the same gap by construction, since REAL Space compounds are synthesized to order at a reported success rate above 80% within three to four weeks, because the underlying reactions were validated before the library was ever enumerated.
This is where purely generative, library-unconstrained methods still run into trouble most often: a diffusion model or a language model can propose a structure with excellent predicted properties that turns out to have no known or economical synthetic route. The field's response has been to build synthesizability scoring, and increasingly retrosynthesis prediction itself, directly into the generation and ranking loop, rather than treating it as a filter applied only at the very end.
Notable programs
Insilico Medicine's own pipeline traces a direct line from its academic-style generative research to the clinic: a compound discovered using an early version of its generative platform, and later refined through Chemistry42, advanced into first-in-human trials for idiopathic pulmonary fibrosis, one of the more closely watched proof points that generative chemistry output can survive contact with a real drug program.
PostEra's most visible demonstration came through the COVID Moonshot, an open-science consortium that used crowdsourced medicinal chemistry submissions alongside PostEra's machine learning and retrosynthesis tools to triage designs against the SARS-CoV-2 main protease; the published campaign covered more than 18,000 compound designs and over 2,400 synthesized molecules, and the collaboration has since continued as a federally funded antiviral discovery platform. On the vendor-library side, Enamine's REAL Space has been paired with outside AI/ML platforms, including a 2025 collaboration with Recursion, to curate targeted screening libraries from its tens of billions of make-on-demand compounds.
Academic groups continue to supply much of the field's underlying methodology rather than its commercial infrastructure. Aspuru-Guzik's lab, split between Harvard and the University of Toronto over the course of this work, has been a recurring source of foundational generative and search methods that later commercial platforms have built on or adapted, a reminder that most of what looks like a product in this space started as a paper.
Key takeaways
|
This article was produced in accordance with Drug Discovery News’ Editorial Policies.
