AI Retrosynthesis: Why Reaction Data Decides the Route

AI Retrosynthesis: Good Routes Start With Good Data

AI retrosynthesis has matured from an academic curiosity into a routine first step in many synthesis labs. Give a tool a target molecule and it works backward, breaking bonds step by step until it reaches starting materials you can buy, then presents a ranked list of routes in seconds. The algorithms have improved steadily, but the most important development lately has been about data rather than models. In late September, CAS announced a collaboration with Novartis to combine Novartis's proprietary experimental reaction records with more than 160 million curated reactions from the CAS Content Collection, standardising data currently scattered across lab notebooks, reports, and shared drives. Coverage described it as building the data layer AI needs, not a new model. That framing is exactly right, because a retrosynthesis tool is only as good as the reactions it learned from.

Types of Synthesis Planning Tools

Each approach makes a different trade between coverage, reliability, and how well it explains itself.

Why Reaction Data Is Suddenly the Story

Most organisations that run synthesis have enormous amounts of reaction data, and most of it is hard to use. Electronic lab notebooks record reactions in inconsistent formats. Conditions live in free text. Yields are reported differently by different chemists. Failed experiments are often logged briefly or not at all. Publicly available reaction data, from journals and patents, has its own problems, with incomplete conditions and a strong bias toward reactions that worked.

The CAS and Novartis work reflects a wider realisation in the industry: the competitive advantage in AI-assisted chemistry is shifting toward whoever has clean, standardised, well-annotated reaction data, including their own internal history. Better data improves route suggestions, condition predictions, and any future language model or agent built on top of it.

How AI Retrosynthesis Works

The basic idea dates back decades: work backward from the target, choosing a disconnection that corresponds to a known reaction, and repeat until every branch reaches available materials. AI makes that search far broader and faster. Models score possible disconnections by how likely they are to work, search the resulting tree of options, and rank complete routes by factors such as step count and starting material availability.

Where AI Retrosynthesis Routes Break

The failure modes are familiar to chemists who use these tools:

• Selectivity problems. A step looks fine on paper but the molecule has another reactive site that competes.
• Missing conditions. A route proposes a transformation without workable temperature, solvent, catalyst, or protecting group strategy.
• Precedent mismatch. A reaction worked on a simple substrate in the literature but fails on a complex, functionalised one.
• Hazardous steps. Routes can include reagents or intermediates that are dangerous, especially at scale.
• Starting material assumptions. A "purchasable" building block may be unavailable, expensive, or slow to source.
• Invented chemistry. General language models in particular can describe reactions that do not exist or cite papers incorrectly.

The Missing Negative Results

The single biggest weakness in reaction data is what is not there. Journals publish successes. Patents describe what worked well enough to claim. Failed reactions, low yields, and unexpected side products rarely make it into the public record, which means models learn a distorted picture where most attempted reactions appear to succeed. A model trained that way will be overconfident about exactly the kinds of steps chemists know to be tricky.

Internal data can correct this, because organisations do record failures, even if messily. In our view, the most valuable thing a chemistry organisation can do for its AI future is boring: record failed and low-yield reactions consistently, with conditions, in a structured way. That data is far scarcer and more valuable than another million successful reactions from the literature.

Vetting a Proposed Route

1. Check each disconnection against real precedent. Find examples on similar substrates, not just the reaction type in general.
2. Look for competing reactive sites. Ask what else in the molecule could react under the proposed conditions.
3. Confirm conditions are specified and realistic. A route without workable conditions is a sketch, not a plan.
4. Screen every step for hazards. Reagents, intermediates, exotherms, and gas evolution all need review.
5. Verify starting material supply. Check availability, price, lead time, and purity from real suppliers.
6. Compare more than one route. The top-ranked route is not always the best once practical factors are weighed.
7. Run the riskiest step first. Test the step most likely to fail before investing in the rest of the sequence.

Pros and Cons of AI Retrosynthesis

• Pro: broader exploration. Tools consider many more disconnections than a person would sketch by hand.
• Pro: faster starting points. Chemists spend less time on initial planning and more on testing.
• Pro: fresh ideas. Less obvious routes can open shorter or cheaper paths.
• Con: overconfidence from biased data. Missing negative results make difficult steps look easier than they are.
• Con: thin on conditions and safety. Many suggestions need significant work before they are runnable.
• Con: skills erosion. Junior chemists who rely on tools may develop weaker intuition for what will fail.

Real Scenarios Worth Thinking Through

These scenarios are illustrative, showing how AI retrosynthesis plays out in practice rather than presented as verified case studies.

A medicinal chemistry team adopts a route ranked first by a planning tool. The third step relies on a coupling that works well in published examples on simple aryl substrates but fails repeatedly on their densely functionalised intermediate. The second-ranked route, slightly longer, works first time. The team now runs the riskiest step of the top two routes before choosing.

A process chemist asks a general language model for conditions on an unusual transformation and receives a confident procedure with a journal reference. The reference does not exist and the conditions are borrowed from a different reaction. A planning tool grounded in curated data proposes something more modest and more realistic.

A company standardises its electronic lab notebook data, including failed reactions, and retrains an internal model. Its route suggestions become noticeably more conservative on steps the company has historically struggled with, which is exactly what experienced chemists had wanted.

“After testing multiple AI models on coding, research, and business prompts, combined outputs produced more reliable results than any single model.” Internal multi-model evaluation, Talkory research team.

From Bench Route to Plant Route

A route that works on a few hundred milligrams is not necessarily a route anyone would run at scale. Process chemistry weighs different factors: reagent cost, solvent volumes, waste, heat management, filtration, crystallisation, and above all safety. Retrosynthesis tools are improving at incorporating some of these, but plant-ready route design still depends heavily on process chemists and safety specialists.

The same gap between screening and scale shows up in materials, which we discussed in AI materials discovery. And when AI-generated answers about chemistry are wrong, the consequences can be serious, a risk we explored in AI in the chemical industry.

Why Talkory Wins

Talkory is not a retrosynthesis engine and does not replace one. Where it helps is the reasoning around a proposed route. It runs the same question across GPT, Claude, Gemini, Grok, Perplexity Sonar, and Kimi K3 in one pass, so a chemist can ask whether a transformation is plausible on a given substrate, what competing reactions to expect, or what hazards a reagent carries, and see whether six independent models agree. Agreement suggests well-established chemistry. Disagreement is a useful flag that a step deserves a literature check or an early experiment before the route is committed.

Final Verdict

AI retrosynthesis is now a genuinely useful starting point for synthesis planning, and it will keep improving. But the quality of its suggestions depends far more on reaction data than on algorithms, and most reaction data is biased toward successes and short on conditions. The organisations that benefit most will invest in clean, standardised internal data, including failures, and treat every suggested route as a hypothesis: check precedent on similar substrates, look for competing reactivity, specify conditions, screen for hazards, verify supply, and test the riskiest step first. The tool proposes. The chemist, and the flask, decide.