The Curation Problem
Last month, I spent three hours trying to reconcile why the same reaction appeared in four different databases with four different yields: 72%, 85%, "good", and "not observed." This wasn't an edge case or an obscure transformation. This was a basic Suzuki coupling that any graduate student would recognize. The problem wasn't lack of data. The problem was that none of it could talk to the other pieces.
Raw chemistry data exists everywhere. PubChem contains over 100 million compounds. SciFinder indexes millions of reactions from decades of literature. Patent databases catalog industrial syntheses going back a century. But when I tried to train a model to predict reaction outcomes from this wealth of information, I discovered something frustrating: most chemistry data isn't actually useful for machine learning. It exists in incompatible formats, uses inconsistent nomenclature, lacks essential metadata, and provides no reliable ground truth for whether reactions actually work as claimed.
This is why we've been building what will become the largest structured organic chemistry dataset in the industry: 50 terabytes of reaction data, synthesis routes, molecular properties, spectra, literature extractions, and experimental outcomes. The number matters, but the word "structured" matters more.
What Structure Actually Means
Structuring chemistry data means solving problems that sound mundane but are technically brutal. It means standardizing molecular representations so that benzene written as "C6H6" in one paper, drawn as a ring in another, and encoded as "c1ccccc1" in a database all refer to the same molecule. It means tagging reaction conditions (solvent, temperature, catalyst, time, yield) with controlled vocabularies so that "rt" and "room temperature" and "25°C" can be properly compared. It means linking literature claims to experimental outcomes so you can distinguish between what chemists hoped would happen and what actually did happen in their hoods.
Most importantly, it means resolving contradictions across sources rather than ignoring them. When four papers report different yields for the same reaction, the answer isn't to pick one or average them. The answer is to understand why they disagree: different scales, different purification methods, different starting material purity, different definitions of success. This is where the real intelligence lies, and it's work that requires human expertise.
Contrast this with existing resources. SciFinder has impressive scale but limited structured reasoning data. You can search for reactions, but you can't easily train models on reaction patterns. Reaxys contains valuable reaction data, but it's locked behind expensive institutional walls and formatted for human browsing, not machine learning. Open databases like ChEMBL and PubChem are incredibly valuable for specific domains, but narrow in scope. No one has assembled structured, reasoning-ready chemistry data at this scale before.
The Models That Scale Unlocks
Once you have 50 terabytes of structured chemistry data, what do you actually train? The honest answer is that we're discovering this as we build it, but specific classes of models become viable that weren't feasible before.
Reaction prediction models require more data than most people realize. Given a set of reagents and conditions, predicting the major product sounds straightforward until you encounter the long tail of chemistry. Common reactions like amide formations and basic substitutions are well-represented in small datasets. But rare reaction classes, unusual substrates, and edge cases only appear with enough data volume. Small datasets overfit to the common cases and fail spectacularly on anything interesting. With sufficient scale, models start to recognize patterns in failures and exceptions, not just successes.
Retrosynthetic planning models are harder still. Given a target molecule, proposing viable disconnections and synthesis routes requires not just reaction knowledge but strategic judgment about what routes are practical. This is where failure data becomes crucial. A model trained only on published, successful syntheses learns an idealized version of chemistry. A model trained on millions of real synthesis attempts, including the ones that didn't work, learns something qualitatively different. It learns which disconnections look good on paper but fail in practice, which protecting groups cause problems downstream, which routes work at milligram scale but break at gram scale.
Property prediction models face a different challenge: sparsity and inconsistency. Predicting absorption, distribution, metabolism, excretion, and toxicity from molecular structure requires aggregating data from decades of pharmaceutical research, academic studies, and regulatory filings. The literature is full of contradictions. Different labs measure the same property on the same molecule and get different results. Different assays, different conditions, different definitions. A large, structured dataset allows models to learn from these contradictions rather than treating them as noise. The model learns that certain properties are inherently variable and adjusts its confidence accordingly.
Multi-modal models represent the most interesting frontier. Chemistry data isn't just structures and reactions. It includes spectra (NMR, IR, mass spec), crystallographic data, natural language descriptions from papers and patents, and images of reaction setups. Models that can reason across all of these simultaneously are qualitatively more useful than models trained on any single modality. But this requires exactly the kind of linked, structured dataset we're building, where a single reaction entry connects to its spectra, its literature description, its crystal structure, and its experimental conditions.
The ultimate goal is chemistry foundation models. Rather than training a model for a specific task, we want to train a general-purpose chemistry model on the full dataset and then fine-tune it for specific applications. This approach has revolutionized natural language processing, but it only becomes viable above a certain data threshold. Language models needed hundreds of gigabytes of text before the foundation model approach worked better than task-specific models. For chemistry, we believe that threshold is somewhere in the tens of terabytes, and our dataset crosses it.
The Encoding Problem
What interests me most about this project is what it reveals about chemistry itself. For a century, experts have encoded chemical knowledge in their own minds. You learned synthesis by working with someone who knew how to make molecules, and they learned from someone before them. The knowledge existed but it wasn't portable or scalable.
Building this dataset forces us to make that knowledge explicit. Every reaction condition, every substrate limitation, every practical trick that experienced chemists know intuitively must find expression in a form that machines can learn from. This isn't just about training better models. It's about creating the first truly comprehensive map of what we actually know about how molecules behave.
The gap between having the data and having the models is real. Data is necessary but not sufficient. But for the first time, we have the infrastructure to encode a century of chemical expertise at scale. What that unlocks will unfold in the years ahead.
Anatoly Chlenov, PhD is the founder of Molekula.ai. He completed his PhD at Princeton, his postdoc at Caltech under R.H. Grubbs, and spent twenty years in industry before founding Molekula in 2025. Beta access is available at molekula.ai.
