Predictive design of crystallographic chiral separation

The efficient separation of chiral molecules is a fundamental challenge in the manufacture of pharmaceuticals and light-polarising materials. We developed an approach that combines machine learning with a physics-based representation to predict resolving agents for chiral molecules, using a transformer-based neural network. On historical data, our approach is 4-6 times more accurate than current practice. We further validate the model in a prospective experiment, where we use the model to design a resolution screen for six unseen racemates. We successfully resolved three of the six mixtures in a single round of experiments and obtained an overall 8-to-1 true positive to false negative ratio. Together with this study, we release a previously proprietary dataset of over 6,000 resolution experiments, the largest diastereomeric salt crystallisation dataset to date. More broadly, our approach and open crystallization data lay the foundation for accelerating and reducing the costs of chiral resolutions. Methods High-Throughput Experimentation reactions were set up inside of an INERT Inc. triple double sized glove box with O2 and H2O levels <20 ppm. Glass vials (0.7 mL, 8 x 30 mm) pre-equipped with stir bars were used for each reaction. The reactions were set up with the components and conditions described by each dataset entry at 0.04~mmol scale and 0.2 M final concentration with a 1:1 ratio of acid to base and 200 μL total volume. The reaction vials were sealed by crimp under the glove-box environment and placed in a metal Chemglass Optichem 96-well heating plate atop a general IKA stirrer heater plate with an external temperature probe to accurately and evenly control the plate. The vials were heated for 1 hour at 80 C before being cooled over 3 hours and left to stir for a further 15hrs at 25 C. Solubility observations were made after the hour at 80 C and after the 15hr stir period at 25 C. At the reaction end point, the vials were centrifuged to settle any precipitates and the liquors sampled and analysed by chiral SFC-MS for the determination of the liquor e.e. and the calculation of solids m.frac. and e.e. For any enriched liquor hits, the liquor was manually pipetted away from the centrifuged solid and the solid re-slurried in fresh solvent (using the same solvent choice as the reaction, half volume, 100 μL) to rinse off residual liquor. The vial was centrifuged a second time and the liquor (100 μL) again removed from the centrifuged solid and combined with the original. Both the combined liquors and the residual solid were brought to an equal volume of 500 μL using MeOH to ensure solution before being analysed by chiral SFC-MS. Measurement of the crystallised salt product % enantiomeric enrichment (ee) from isolated solids superseded the product % ee calculated based on SFC-MS analysis of uncrystallised material in the liquors. iChem explorer (Reaction Analytics, US) and Virscidian Analytical Studio™ software were used for data analysis. HPLC grade methanol, isopropanol, ammonium formate, ammonia (Fisher Scientific, Pittsburgh, PA, USA) and bulk grade carbon dioxide (AirGas West (Escondido, California, USA) were used in this study. The CO2 was purified and pressurised to 1500 psig using a custom booster and purifier system from FLW, Inc. (Huntington Beach, CA, USA). Analysis was performed using an Agilent 1260 SFC/MS system consisting of a binary pump, SFC control module, UV/DAD detector, and column compartment with an internal 6-position, 12-port valve, and 6120 MSD with an APCI source (Agilent, Inc., Santa Clara, CA, USA). A Gerstel MPS autosampler (Gerstel USA, MD, USA) was equipped with a 25 μL syringe and a 50 nL internal loop with a control method to vent CO2 from the loop prior to sample introduction. The effluent of the SFC is split to the MSD using a 3-way tee (Valco, Houston, TX, USA) and a 50 cm long, 50 μm i.d. PEEKsil capillary tubing (Trajan Scientific, NC, USA) located between the column outlet and BPR. All data was acquired using Agilent 64-bit ChemStation (Version C.01.10).手性分子（chiral molecules）的高效分离是制药与光偏振材料制造领域的核心难题。我们开发了一种结合机器学习与物理表征的方法，基于Transformer神经网络预测手性分子的拆分剂。在历史数据集上，我们的方法精度较当前主流方案提升4至6倍。我们进一步通过前瞻性实验验证了该模型：利用模型为6种未见外消旋混合物设计拆分筛选流程，仅一轮实验便成功拆分其中3种混合物，整体真阳性与假阴性比值达到8:1。本研究同步公开了此前专属的超过6000组拆分实验数据集，是目前规模最大的非对映体盐结晶数据集。从更广泛的视角来看，我们的方法与开放的结晶数据为加速手性拆分流程、降低其成本奠定了基础。 方法 高通量实验均在INERT Inc.生产的双规格三腔手套箱中进行，箱内氧气与水含量均低于20 ppm。所有反应均使用预装搅拌子的0.7 mL（8×30 mm）玻璃小瓶。反应体系按照数据集条目指定的组分与条件搭建，规模为0.04 mmol，终浓度0.2 M，酸碱比例1:1，总体积200 μL。反应小瓶在手套箱环境中采用压盖密封，随后置于金属材质的Chemglass Optichem 96孔加热板上，该加热板搭载通用IKA磁力搅拌加热台，并外接温度探头以实现精准均匀的控温。小瓶先于80 ℃加热1小时，随后以3小时梯度降温至25 ℃，并继续搅拌15小时。分别在80 ℃加热1小时后与25 ℃搅拌15小时后，观察体系的溶解情况。反应结束后，将小瓶离心以沉降沉淀物，取上清液通过手性超临界流体色谱-质谱（chiral SFC-MS）分析，测定上清液的对映体过量值（enantiomeric excess, e.e.），并计算固体的质量分数与e.e.值。 针对上清液中检出的富集产物，手动移去离心后的上清液，将固体用新鲜溶剂（与反应体系相同溶剂，体积减半即100 μL）重悬打浆，以洗去残留上清液。再次离心小瓶，移去第二次上清液（100 μL）并与原上清液合并。将合并后的上清液与残留固体均用甲醇（MeOH）定容至500 μL以确保完全溶解，随后通过手性超临界流体色谱-质谱分析。通过分离固体测定的结晶盐产物对映体富集百分比（ee值），优先于通过上清液未结晶物质的SFC-MS分析计算得到的产物ee值。数据分析采用iChem explorer（Reaction Analytics, 美国）与Virscidian Analytical Studio™软件完成。 本研究使用的试剂包括HPLC级甲醇、异丙醇、甲酸铵、氨（购自Fisher Scientific，美国宾夕法尼亚州匹兹堡市），以及工业级二氧化碳（购自AirGas West，美国加利福尼亚州埃斯孔迪多市）。二氧化碳通过FLW, Inc.（美国加利福尼亚州亨廷顿比奇市）定制的增压与纯化系统净化并加压至1500 psig。 分析测试采用Agilent 1260 SFC/MS系统，该系统包含二元泵、SFC控制模块、UV/DAD检测器与柱温箱，柱温箱内置6位置12通阀，搭配带有APCI源的6120 MSD（均来自Agilent, Inc.，美国加利福尼亚州圣克拉拉市）。实验搭载Gerstel MPS自动进样器（Gerstel USA，美国马里兰州），配备25 μL进样针与50 nL定量环，进样前可通过控制程序放空定量环内的二氧化碳。SFC流出液通过三通接头（Valco，美国德克萨斯州休斯顿市）分流至MSD，柱出口与反压调节器（BPR）之间连接一段长50 cm、内径50 μm的PEEKsil毛细管柱（Trajan Scientific，美国北卡罗来纳州）。所有数据通过Agilent 64位ChemStation软件（版本C.01.10）采集。