{"id":11508,"date":"2026-11-10T12:44:47","date_gmt":"2026-11-10T07:14:47","guid":{"rendered":"https:\/\/chiralpedia.com\/blog\/?p=11508"},"modified":"2026-11-10T12:44:55","modified_gmt":"2026-11-10T07:14:55","slug":"from-generative-stereochemistry-to-autonomous-chiral-discovery","status":"publish","type":"post","link":"https:\/\/chiralpedia.com\/blog\/from-generative-stereochemistry-to-autonomous-chiral-discovery\/","title":{"rendered":"Episode 5: From Generative Stereochemistry to Autonomous Chiral Discovery"},"content":{"rendered":"\n

From Generative Stereochemistry to Autonomous Chiral Discovery<\/a><\/p>\n\n\n\n

Summary<\/strong><\/p>\n\n\n\n

A closed loop for chirality would design a stereochemically defined molecule, plan a route, execute it, measure enantiomeric excess, and decide what to run next. Every link now exists somewhere in the literature. None have been joined end to end for a stereochemical objective. The weakest links are not the ones usually blamed: generative models can be made to respect handedness, and robotic platforms can optimise enantioselective reactions. What is missing is reliable stereochemical specification in the foundation models most people actually use, automated measurement of enantiomeric excess inside the loop, and published failures to learn from. This closing article assesses those gaps, then returns to the question the series opened with: is AI becoming a driver of chirality-focused science, or a very good optimisation assistant?<\/p>\n\n\n\n

Key takeaways<\/strong><\/a><\/p>\n\n\n\n

    \n
  1. The loop is assembled in pieces.<\/strong> Design, planning, execution, measurement and decision each have demonstrations; no published system joins them for a stereochemical objective.<\/li>\n\n\n\n
  2. Many chemical language models are trained on SMILES with stereochemistry deliberately stripped out<\/strong> \u00a0a choice that makes the question of whether they \u201cunderstand\u201d chirality partly moot.<\/li>\n\n\n\n
  3. Chiral learning appears to be a distinct training phenomenon<\/strong>, not simply a capacity problem, according to recent mechanistic work.<\/li>\n\n\n\n
  4. Measurement is the underrated bottleneck.<\/strong> A loop optimising enantioselectivity needs an ee value automatically, and chiral analysis is slower than the chemistry. <\/li>\n\n\n\n
  5. The honest overall verdict is domain-dependent<\/strong>, and this article gives it plainly rather than splitting the difference.<\/li>\n<\/ol>\n\n\n\n

    Opening: components in search of a circuit<\/strong><\/a><\/p>\n\n\n\n

    Episode 4 catalogued real experimental successes: catalysts predicted and made, enzyme variants evolved under model guidance, a fourteen-step synthesis informed by a reaction model, resolving agents proposed for racemates the model had never seen.<\/p>\n\n\n\n

    Each is a component. In every case a human read the prediction, decided what to run, ran it, measured the result, and decided what to do next. The question for this episode is whether that human-mediated cycle can be closed  and whether it should be.<\/p>\n\n\n\n

    The central problem: five links, none of them joined<\/strong><\/a><\/p>\n\n\n\n

    A closed chiral discovery loop requires: (1) generating a molecule with specified stereochemistry; (2) planning a route that controls it; (3) executing the chemistry; (4) measuring the stereochemical outcome; (5) updating the model and choosing the next experiment.<\/p>\n\n\n\n

    Links 3 and 5 are mature and general. Link 2 works in narrow domains. Link 1 is where representation problems from Episode 1 resurface. Link 4 is the one nobody talks about and the one most likely to determine whether this happens.<\/p>\n\n\n\n

    \"\"
    Figure 1.<\/strong> The five links, and where each one breaks. Five links make a closed chiral discovery loop, and every one has a demonstration except measurement of enantiomeric excess inside the loop, which is why the loop is drawn open.
    The autonomy level definitions in the inset are marked TO CONFIRM, since the vocabulary is cited without a scale.<\/em><\/figcaption><\/figure>\n\n\n\n

    Can generative models specify stereochemistry?<\/strong><\/a><\/p>\n\n\n\n

    Episode 1 introduced the formal obstacle: Dumitrescu et al. (2025)<\/a> proved that E(3)-invariant models must disregard chirality, and showed empirically that E(3)-equivariant generative models produce the two enantiomers of a molecule with equal probability. A generator with this property is not making a stereochemical error occasionally  it is flipping a coin every time, silently. Their field-based alternative recovers configuration while remaining competitive on standard generation metrics, so this is a solved problem in principle.<\/p>\n\n\n\n

    In practice, it is often not solved because it is not attempted. Li et al. (2026, preprint)<\/a> note that many existing chemical language models are intentionally<\/em> trained on SMILES stripped of stereochemical information. Where that is the case, asking whether the model understands chirality is the wrong question; it was never shown any.<\/p>\n\n\n\n

    Do models that see stereochemistry learn it, or copy it?<\/strong><\/a><\/p>\n\n\n\n

    This is the question the protocol for this series posed most sharply, and there is now direct evidence bearing on it  though it is preprint evidence and should be read as such.<\/p>\n\n\n\n

    Li et al. built Pan-CORE, a family of transformer encoder\u2013decoder models for SMILES translation, and tracked chiral learning through high-resolution checkpoint analysis during training. They observe a reproducible pattern across model variants: chiral-token accuracy sits on a long plateau and then rises abruptly. Their interpretation is that this stagnation is not explained by model capacity alone, but reflects the complexity of chiral constraints  building on an earlier observation that transformers have specific difficulty with the @ and @@ tokens, and that confusing enantiomers prolongs training stagnation.<\/p>\n\n\n\n

    Their mechanistic analyses point to an encoder-centred transition, with chiral-token representations undergoing destabilisation and reconstruction, and targeted ablation identifying a small set of chiral-sensitive attention heads whose removal selectively degrades chiral-token accuracy even in a fully trained model.<\/p>\n\n\n\n

    If that finding replicates, the answer to \u201cnotation or understanding?\u201d is interesting: something more than surface pattern matching is happening, it is localised, and it emerges late. But this is a single preprint from one group, the mechanistic claims are correlational, and localisation of a capability is not the same as chemical comprehension. I would treat it as the most informative available evidence and not as settled.<\/p>\n\n\n\n

    Route planning with stereochemical control<\/strong><\/a><\/p>\n\n\n\n

    Pesciullesi et al.\u00a0(2020)<\/a> remains the clearest demonstration that a sequence model can be fine-tuned to handle regio- and stereoselectivity well enough to guide real synthetic decisions, validated in a fourteen-step synthesis. The important qualifier is domain: carbohydrates, with a small, curated transfer set. General stereochemically-aware retrosynthesis \u00a0where the planner reasons about configurational stability, protecting-group-dependent selectivity, and whether a proposed centre is even synthetically realistic \u00a0has not been demonstrated at that standard.<\/p>\n\n\n\n

    The measurement link<\/strong><\/a><\/p>\n\n\n\n

    Optimising a reaction for yield requires a number an instrument can produce quickly. Optimising for enantiomeric excess requires chiral analysis, which is slower, method-dependent, and historically developed by trial and error.<\/p>\n\n\n\n

    This is where Episode 2\u2019s material becomes load-bearing rather than incidental. Xu et al. (2023)<\/a> frame their output as a separation probability<\/em> under given conditions rather than a raw retention time  precisely the quantity an automated system needs to decide whether a proposed chiral method will resolve the peaks it must integrate. And Howarth and Goodman (2022)<\/a> provide the template for the decision side: a calibrated probability that a single proposed structure is correct, rather than a ranking among candidates.<\/p>\n\n\n\n

    An autonomous loop that reports ee without reporting confidence in the assignment is not doing science. The tools to avoid that exist; they have not been wired in.<\/p>\n\n\n\n

    Case study 1: how close is a self-driving lab?<\/a><\/p>\n\n\n\n

    Pilon et al. (2026)<\/a> report RoboChem-Flex, a low-cost modular self-driving laboratory combining in-house hardware with a Python framework integrating real-time device control and Bayesian optimisation, including multi-objective and transfer-learning workflows. It was validated across six case studies  photocatalysis, biocatalysis, thermal cross-couplings and enantioselective catalysis  spanning single- and multi-objective optimisation, and all code, firmware, GUI software and hardware design files are openly available.<\/p>\n\n\n\n

    Read carefully, this is both more and less than \u201cautonomous asymmetric synthesis\u201d. More, because it demonstrates that closed-loop optimisation of an enantioselective transformation is achievable on affordable hardware, and because openness makes it reproducible. Less, because enantioselective catalysis is one campaign among six, and because a central design feature is a sampling module that lets the platform interface with existing shared analytical equipment<\/em>  which in some configurations means human-in-the-loop rather than fully closed.<\/p>\n\n\n\n

    That is not a criticism. It is an accurate description of the current state, and the platform\u2019s flexibility about where the human sits is arguably its most practical feature.<\/p>\n\n\n\n

    For a wider frame, Tom et al. (2024)<\/a> survey the field and, usefully, define levels of autonomy by hardware and software capability. Adopting that vocabulary would prevent a great deal of overclaiming: most systems described as autonomous are automated at one level and supervised at another, and saying which is a small discipline with a large payoff.<\/p>\n\n\n\n

    Case study 2: the closest thing to a closed chiral loop<\/a><\/p>\n\n\n\n

    Elijo\u0161ius et al. (2025)<\/a> is, in structure, one full turn of the cycle: a model trained on historical data designed a resolution screen for six unseen racemates, the screen was run, three resolved, and the outcome was reported with a true-positive to false-negative ratio.<\/p>\n\n\n\n

    What makes it the closest approach is not automation  the experiments were human-executed  but that it completes the epistemic loop honestly: prediction, prospective test, published result including failures, and release of the underlying dataset of over 6,000 experiments. A loop that runs fast but never publishes what did not work is worth less than a slow loop that does.<\/p>\n\n\n\n

    Barriers<\/strong><\/a><\/p>\n\n\n\n

    Data scarcity and bias.<\/strong> Selectivity data cluster around catalysts that already worked.<\/p>\n\n\n\n

    Unpublished failures.<\/strong> The single most damaging structural problem. Models trained on the literature learn a censored view of chemistry, which is a direct cause of the extrapolation failures documented in Episode 4.<\/p>\n\n\n\n

    Inconsistent stereochemical labels<\/strong>, as Episode 1 established.<\/p>\n\n\n\n

    Absent negative controls and no stereoisomer-aware splitting<\/strong>, as Episode 3 established.<\/p>\n\n\n\n

    Opacity and reproducibility.<\/strong> Partly mitigated where code and data are released  increasingly the norm in this corner of the field, and worth insisting on.<\/p>\n\n\n\n

    Intellectual property.<\/strong> Resolution and process data are commercially sensitive; the Pfizer and Janssen releases are notable precisely because they are unusual.<\/p>\n\n\n\n

    Regulatory acceptance.<\/strong> An AI-assisted stereochemical assignment or process decision entering a submission needs provenance, applicability domain and uncertainty documented. Nothing in the technical literature prevents this; it simply has not been made routine.<\/p>\n\n\n\n

    Governance: what a defensible loop needs<\/strong><\/a><\/p>\n\n\n\n

    Four requirements, none of them exotic: a stated applicability domain<\/strong>, so a model\u2019s silence outside its training region is explicit; calibrated uncertainty<\/strong> on every stereochemical call; provenance and auditability<\/strong>, recording which model version proposed what, on which data, and what was measured; and human oversight positioned where the science is weakest<\/strong> \u00a0which, on the evidence of this series, means the biological consequences of configuration rather than the chemistry of making it.<\/p>\n\n\n\n

    \"\"
    Table 1.<\/strong> Due diligence on a stereochemical claim, and what a defensible loop needs. Three questions to put to any tool claiming stereochemical capability, four requirements for a loop you intend to run yourself, and a filled action column on every row, including the rows where no model exists.<\/figcaption><\/figure>\n\n\n\n

    Realistic expectations<\/strong><\/a><\/p>\n\n\n\n

    Two years.<\/strong> Chiral method development and resolving-agent screening move into routine industrial use as decision-support tools. More proprietary datasets are released. Closed-loop optimisation of enantioselective reactions is demonstrated by several more groups, still with human-mediated analysis.<\/p>\n\n\n\n

    Five years.<\/strong> In-line or at-line chiral analysis integrated into optimisation loops, making genuinely closed asymmetric optimisation ordinary for a defined class of reactions. Stereochemically-aware retrosynthesis extends beyond single domains. Chirality-aware ADME models begin to appear if<\/em> enantiomer-paired data are released  the technical work is not the obstacle.<\/p>\n\n\n\n

    Ten years.<\/strong> Multimodal models spanning structures, spectra, reactions, conditions and protein structures make stereochemical reasoning a default rather than a bolt-on. Whether this arrives depends far more on data-sharing norms than on architectures.<\/p>\n\n\n\n

    I would treat any claim of fully autonomous chiral drug discovery within this window with scepticism.<\/p>\n\n\n\n

    What AI Adds<\/strong><\/a><\/p>\n\n\n\n

    Across five episodes: correctness<\/strong> where representations previously failed silently; speed<\/strong> in measurement and assignment; screening reduction<\/strong> in separation and enzyme engineering; prospective proposals<\/strong> that survived experimental test; and, unexpectedly, datasets and audits<\/strong> \u00a0models have been used to find errors in the literature that trains them.<\/p>\n\n\n\n

    Current Limitations and Evidence Gaps<\/strong><\/a><\/p>\n\n\n\n