AI4S Insight | Overcoming Post-Processing Challenges: Ou Shisheng Completes the "Last Mile" of Automated Organic Synthesis

The development of automated organic synthesis has historically focused on reaction units. Automated dosing, temperature control, and stirring are reliably achieved through machinery and programming, and multi-channel parallel systems have significantly sped up the screening of reaction conditions. However, a complete synthetic route doesn't conclude when the reaction finishes; essential post-processing steps, including separation and purification, are still required. As depicted in the figure below, separation and purification form a distinct, critical module within the AI4S end-to-end synthesis framework. In practice, however, automating these post-processing procedures still presents many difficulties. Considering the practical limitations of mg-scale synthetic R&D, this article explores ways to achieve automated post-processing.

 End-to-End Framework for Chemical Synthesis Development (Ruan, Y. et al., Nature Communications 15, 10160, 2024)

(a) The workflow comprises six stages: literature search → substrate and condition screening → kinetic study → condition optimization → scale-up → purification → product delivery. The modules involving LLM agents in this work are screening, kinetic study, optimization, scale-up, and purification (highlighted in orange).

(b) Six LLM agents are assigned distinct roles: Literature Scouter, Experiment Designer, Hardware Executor, Spectrum Analyzer, Separation Instructor (responsible for separation condition decisions), and Result Interpreter.

(c) The web-based interactive interface allows chemists to describe tasks in natural language. The agents then propose recommendations or generate executable operations within the interface. Work proceeds after chemists review, approve, or revise these proposals until the task is completed. The interface also displays parameters such as agent configuration, token consumption, correctness, and level of detail. This empowers users without programming expertise to operate the automated experimental platform and interpret the results.

This framework treats purification as a separate, essential step and designates a specific decision-making agent for it. This aligns with the core perspective of this article, which emphasizes that "focus lies after the reaction."

 I. Structural Constraints in Organic Synthesis R&D: Post-Reaction Bottlenecks

Typical R&D workflows in medicinal chemistry and general organic synthesis involve several steps: reagent dosing, temperature-controlled reactions, solid-liquid separation, thin-layer chromatography (TLC) interpretation, column chromatography purification, concentration, product confirmation, and packaging. The primary efficiency bottlenecks are not necessarily in performing the reaction itself, but rather in a series of judgment and handover steps that occur afterward.

· Inconsistent Microscale Operations: Weighing and pipetting small amounts of reagents during dosing can easily lead to errors in milligram-scale operations. Similarly, the volume and placement of spots for thin-layer chromatography rely heavily on manual manipulation. As a result, the same formulation can produce varied results across different batches, reducing data reproducibility.

· Sample Loss During Transfer and Separation: Transferring reaction mixtures and performing solid-liquid separation after the reaction are particularly prone to sample loss in milligram-scale systems. Target compounds can be retained on filter membranes, filter cakes, and vessel walls, directly lowering the recovery yield.

· Experience-Dependent TLC Interpretation and Elution Setup: Traditional workflows require chemists to visually inspect TLC plates under UV lamps and manually estimate the retention factor (Rf). The elution gradient for preparative chromatography is then determined empirically. Both the reproducibility of spotting and the accuracy of Rf estimation are difficult to guarantee. Furthermore, the logic for deriving gradients from visual interpretation is challenging to learn from written documentation.

· Challenges in Purification Consistency and Decision-Making: The consistency of silica gel packing for dry loading directly impacts column efficiency. Incorrect gradient settings often necessitate repeat separations. Peak collection and fraction assignment typically require manual judgment based on multiple signals, and the collected fractions frequently need separate testing for structural confirmation.

· Automation Gaps in Concentration and Sample Collection: Concentrating multiple batches sequentially is time-consuming and carries a risk of cross-contamination. The endpoint for concentration is often determined by visual inspection, leading to over- or under-evaporation, which can affect purity. Collection and labeling remain manual processes, prone to omissions and errors, with poor traceability.

· Disconnected Data Across Units: Each unit in the workflow is typically controlled by standalone software. Workflows are manually linked, and process records are largely handwritten. Result data become scattered across various media and formats without unified process context, hindering reuse and modeling efforts.

mg-Scale High-Throughput Synthesis Workstation from Ou Shisheng

II. End-to-End Automation: A mg-Scale High-Throughput Synthesis Workstation

To overcome challenges such as numerous process steps, significant decision-making demands, and the need for high precision in microscale operations, Ou Shisheng's automated milligram-scale high-throughput synthesis platform is designed with end-to-end automation of the synthetic workflow as its primary goal. Integrated within a single instrument, the platform manages solution preparation, temperature-controlled reactions, positive-pressure solid-liquid separation, automatic thin-layer chromatography detection, preparative chromatographic purification, mass spectrometry confirmation, concentration, and automatic capping. All operations are orchestrated by transfer rails and robotic arms. For instance, consider the platform delivered to a university in China:

Multi-channel parallelization to replace serial trial-and-error. The system supports 24 parallel reaction channels, each equipped with a syringe pump for automatic liquid dosing with ±1% accuracy and 0.29% repeatability (RSD). Heated magnetic stirring is integrated, allowing independent stirring from 0 to 1200 rpm across the 24 channels and temperature control from ambient +5 °C to 150 °C. A robotic arm automatically handles magnetic stir bars, and reaction vials are transferred automatically throughout the workflow.

Positive-pressure membrane-based solid-liquid separation reduces transfer loss. Positive pressure is applied via connecting rods to push the sample liquid through an integrated membrane filter, retaining particulate impurities. This method achieves solid-liquid separation under positive pressure. Subsequent backwashing with solvent redissolves solids on the membrane, minimizing target compound retention on the filter. The multi-channel solid-liquid separator includes an automatic filter tip replacement mechanism, eliminating exposure risks and efficiency losses associated with manual filter changes.

Automatic Rf interpretation via image recognition and auto-generation of elution gradients. The liquid-handling unit automatically spots samples. Silica plates are fed from a magazine, and the chromatography chamber opens and closes automatically. Two-dimensional vision combined with UV imaging automatically calculates Rf values. This Rf data is directly imported into the preparative chromatography system to generate and execute gradient elution tables, transforming the traditionally experience-dependent interpretation step into method parameters driven by analytical results.

Integrated purification and fraction identification. A 12-channel automatic dry-loading unit ensures uniform silica adsorption and packing, compatible with various specifications and operating in time-sharing mode to enhance packing consistency. Dual medium-pressure preparative chromatography units perform binary gradients with dual-wavelength UV detection. They support multiple collection modes, including full collection, threshold collection, slope-based collection, and peak collection. Upon UV detection of a peak, a single quadrupole mass spectrometer is automatically triggered to confirm fractions and collect samples on demand, eliminating the wait time for offline testing.

Concentration, vial sealing, and full-process data traceability. 25-channel parallel rotary vacuum distillation is used for solvent removal. Automatic barcode scanning and vial sealing are implemented. Panoramic cameras facilitate comprehensive material tracking. Process parameters and results are automatically recorded and archived in a structured format, with experimental reports exportable in PDF.

III. Data Value from Automation: Enabling Structured Data Generation for AI4S

For AI4S, automated equipment provides value by acting as a standardized data production pipeline. When setpoints and measured values from each step—including dosing, reaction, separation, interpretation, purification, and concentration—are automatically correlated and archived by unified scheduling software, each sample enters the database with complete information on its "origin, process conditions, and results." This forms the minimum data unit required for modeling. This is particularly crucial for post-processing: if the parameters of post-processing steps are not automatically recorded, interpretable intermediate links between reaction conditions and final results will be lost, making the data difficult to reuse.

IV. Standardization and Reusability of Empirical Operations

Separation is reliably performed using the positive-pressure membrane structure. Rf values are calculated through thin-layer chromatographic image recognition, and elution gradients are automatically adjusted accordingly. Mass spectrometry verifies collected fractions in real time, while concentration and sample collection are executed programmatically with automatic sample ID assignment. Steps in organic synthesis that previously depended heavily on individual experimental experience are transformed into standardized operational capabilities that can be invoked repeatedly. Consequently, R&D work reduces its reliance on personal hands-on experience and is increasingly driven by reproducible data and standardized workflows.

From deployed catalytic instruments to the milligram-scale high-throughput synthesis workstation adapted for general organic synthesis workflows, Ou Shisheng continues to advance its AI4S research infrastructure. Automation capability is no longer confined to specific reaction scenarios; it has been extended to the broader post-processing chain of organic synthesis. Guided by the philosophy of “empowering scientists’ dreams through an intelligent instrument ecosystem,” Ou Shisheng is dedicated to developing a stable, reliable, traceable, and collaborative intelligent instrument system that encompasses the entire organic synthesis workflow, from reagent dosing to sample collection.

Appendix: Key Technical Specifications