shotgun 方法实现高度多重化哺乳动物代谢工程

Main

对细胞行为施加精确且可预测的控制，是合成生物学的核心目标。在代谢工程中，这种控制通常通过协调多个转录单元（TU）的活性来实现，每个转录单元由一个基因及其调控元件组成。合成代谢途径常常通过将来自不同物种的编码序列（CDS）移植到正交宿主中来构建，从而赋予宿主新的代谢能力。这种方法已取得了显著的成功，包括开发出基于细胞的工业制造工艺替代方案、稳定此前依赖季节性植物来源的制药供应链，以及捕获大气中的二氧化碳以生成具有工业价值的生物质^1,2,3。尽管这些策略已在微生物底盘中得到广泛验证，但在哺乳动物细胞中，这种途径规模的代谢工程实例仍然很少^4,5，尽管人们普遍认识到，改善对哺乳动物细胞功能的控制可以增强细胞疗法、生产日益复杂的生物制品，并改善病毒载体的可规模化制造^6,7,8。现代DNA合成技术如今使得整合来自整个系统发育树的序列或从头设计的序列成为可能，极大地扩展了哺乳动物细胞可及的功能范围。

尽管前景广阔，但哺乳动物代谢工程实例有限，部分原因在于缺乏指导代谢途径设计所需的基础知识。给定一个目标代谢途径和一个目标宿主细胞类型，人们面临一系列复杂的设计决策，包括CDS的选择、调控元件的配置、每个编码蛋白的细胞内定位，以及实现最佳功能所需的每个TU的相对表达水平和化学计量比（图 1a）。这些变量构成了巨大的组合设计空间，然而目前几乎没有机制性理解来指导最优选择。

在快速增长、遗传可操作的微生物系统中，例如Escherichia coli（20分钟倍增时间）或Saccharomyces cerevisiae（90分钟倍增时间），机制知识的缺乏通常可以通过设计-构建-测试（DBT）框架来缓解，在该框架中，代谢途径被快速构建、测试并迭代优化，直至达到令人满意的性能^9,10（图 1b）。在哺乳动物细胞中应用同样复杂的代谢工程策略则要困难得多，主要原因是其倍增时间要长得多（通常超过20小时），这极大地减缓了迭代的速度。

随着途径复杂性的增加，或者当多种功能被整合到单个细胞中时，这一限制变得愈发严重。虽然此类复杂功能的最终实现可能依赖于近期在向哺乳动物细胞合成和递送大于100 kb的大型DNA构建体方面的进展（参考文献 ¹¹），但这些方法并不适合高通量筛选，因为将大型DNA构建体递送到哺乳动物细胞中的效率极低。模块化组合策略，例如CLASSIC，提供了一种解决方案，能够在约10 kb的尺度上筛选约10⁵个构建体¹²（2–3个转录单元）；然而，DNA大小与递送成功率之间存在强烈的反比关系；递送更大的构建体变得指数级困难，使得对复杂合成设计进行暴力筛选变得不切实际¹³。

为了克服哺乳动物系统中实验筛选的局限性，一些代谢工程方法在实验室测试之前，先对候选干预进行计算机筛选^12,14。针对常用的哺乳动物细胞类型，已经开发了复杂的转录组数据库和代谢模型，能够识别上位性代谢功能，并设计出复杂的多重敲除细胞系，从而改善生物工艺和抗体的可制造性^15,16,17。然而，这些模型通常是在自然进化塑造的观测数据上训练的，并且是针对生物制造中常用的狭窄范围的物种和细胞类型进行优化的。因此，它们不适合生成式设计，包括涉及引入异源或完全合成的转录单元的场景。包含合成代谢途径的转录单元阵列的设计，又增加了一层复杂性，超出了这些模型目前能够处理的范围。因此，需要合成数据——系统性的、大规模的代谢扰动，可用于训练能够进行预测和面向设计应用的机器学习模型¹⁸。

在此，我们介绍鸟枪法基因工程（SGE），这是一种用于哺乳动物代谢工程的多重方法，能够快速发现复杂的通路解决方案，并创建用于训练预测模型的大规模合成数据集。SGE 利用来自混合文库的带条形码的单个转录单元（TU）的高多重性整合，使不同的通路变体能够在每个细胞中随机组装。由于小型 DNA 构建体在合成、组装和递送方面呈指数级地更容易，这种设计将每个细胞变成一个独立的实验，并能够并行筛选数百万种组合。通过选择具有所需表型的细胞并对其相关条形码进行测序，SGE 绕过了顺序 DBT 循环的局限性，并大幅加快了发现时间线（图 1c）。

作为概念验证，我们应用 SGE 筛选数百万种通路组合，以在哺乳动物细胞中工程化两条必需氨基酸（EAA）生物合成通路。在先前的工作中，我们通过引入四个E. coli基因，使中国仓鼠卵巢（CHO）细胞能够在没有缬氨酸的情况下增殖，实现了 3.8 天的无缬氨酸倍增时间⁵。然而，尽管生物合成步骤存在重叠，我们仍无法赋予异亮氨酸独立性。在此，我们使用 SGE 在 CHO 细胞中重建并优化缬氨酸原养型，获得了倍增时间仅为 1.1 天、接近野生型的克隆。我们还发现了一种功能性六基因解决方案，使 CHO 细胞能够实现异亮氨酸原养型。在这两种情况下，我们都看到了强烈倾向于生物合成酶线粒体定位以优化通路功能的特征，这表明细胞器定位是哺乳动物代谢工程中一个尚未被充分探索的变量。扩展这一方法，我们将缬氨酸原养型引入人 T 淋巴细胞（Jurkat）细胞系——这是免疫细胞中代谢通路扩展的初步演示，在工程化抗应激细胞疗法中具有潜在应用。最后，我们表明 SGE 能够训练机器学习分类器，以识别可预测通路功能的遗传特征。总之，这些结果确立了 SGE 作为哺乳动物细胞中通路规模代谢工程的可扩展框架。

结果

SGE 的设计

SGE 的一个核心前提是，单个 TU 在合成、组装和递送方面比大型多基因构建体要容易指数倍。将单独条形码标记的 TU 混合文库以高感染复数（MOI）递送，使每个细胞接收通路组分的独特组合，并作为一个独立实验¹⁹。该策略避免了逐个构建和递送每个完整通路，从而能够对通路组成、化学计量、表达和亚细胞定位进行可扩展的探索。为了生成 TU 文库，带有条形码的 CDS 被合成，其两侧带有 IIS 型限制性酶切位点，并定向克隆到含有条形码启动子以及（在适当情况下）N 端细胞器定位信号（OLS）的混合载体中。改变启动子强度能够优化通路通量并减轻通路组分或中间产物带来的毒性^{20,21,22,23,24}，而细胞器靶向可以增加局部酶浓度或提供有利的生化环境^1,25。CDS、启动子和 OLS 通过 Golden Gate 克隆组装到慢病毒兼容的表达载体中，识别条形码彼此位于 3′ 非翻译区（UTR）内约 100 bp 范围内，以便通过单个 PCR 扩增子回收。

所得文库以高 MOI 递送，产生在 TU 组成、拷贝数和化学计量上各不相同的细胞。对细胞进行选择或功能筛选，并通过条形码测序鉴定富集的通路配置（图 1c）。尽管我们在此使用基于生长的选择，但 SGE 与生物传感器、荧光激活细胞分选和其他功能读出兼容。文库还可以纳入额外的 CDS、启动子–OLS 对、宿主因子扰动或替代递送机制（例如转座子或重组酶）。

我们最初选择慢病毒是因为滴定 MOI 简便且其宿主范围广泛，这使我们能够在多种细胞类型中部署相同的文库。然而，慢病毒递送也带来若干实际限制，包括需要在文库克隆后进行包装、对货物大小的限制以及可能导致条形码交换的潜在模板转换²⁶。因此，对于某些应用，尤其是在处理较大构建体或必须维持条形码连锁时，基于转座子的策略（例如 piggyBac）可能更合适。此外，如果使用慢病毒，纳入大规模并行报告基因检测中使用的双 5′ 加 3′ 条形码策略²⁷或直接长读长读出可能有助于缓解条形码交换问题。

SGE 缬氨酸/异亮氨酸文库的整理与构建

为了确认SGE作为代谢工程方法的可行性，我们聚焦于在哺乳动物细胞中重建EAA生物合成。这一挑战提供了一个严格的测试案例；EAA生物合成在5亿多年前的哺乳动物谱系中就已丢失，哺乳动物必须从环境中获取九种氨基酸。相比之下，许多微生物保留了合成这些化合物的能力^28,29。因此，鉴于其复杂性、生物技术价值和实验可操作性，将EAA途径工程化到哺乳动物细胞中，为探索代谢重编程的边界提供了一个强大的模型系统。关键的是，该系统提供了一种简单而有效的选择策略；细胞在缺乏目标氨基酸的条件下生长，只有具有功能性生物合成途径的细胞才能生长⁵。

我们最初聚焦于CHO细胞中缬氨酸和异亮氨酸的生物合成。在先前的工作中，我们运用理性设计原则，发现当以固定的1:1:1:1化学计量比表达时，E. coli的ilvN、ilvB、ilvC和ilvD能够赋予缬氨酸原养型⁵（图2a）。然而，尽管这组基因与缬氨酸生物合成在酶促步骤上存在重叠，但同一组基因却不足以赋予CHO细胞异亮氨酸生物合成能力（图2b）。我们推测，这可能是因为：(1) 异亮氨酸途径特异性底物2-氧代丁酸的可利用性不足；(2) 途径产物对途径早期步骤的反馈抑制；或 (3) 由催化亚基ilvB和调节亚基ilvN编码的乙酰羟酸合酶I（AHAS I）复合体的催化特性欠佳。

因此，我们构建了一个包含ilvN、ilvB、ilvC和ilvD以及五个额外途径组分的文库（图2b和表1）。具体而言，我们添加了源自E. coli的ilvA，以解决异亮氨酸特异性底物2-氧代丁酸可用性受限的假设。为解决反馈抑制的第二个挑战，我们纳入了两个ilvA突变变体，它们对异亮氨酸本身介导的负反馈抑制具有部分（ilvA^L481F）或完全（ilvA^L447F）不敏感性^30,31,32。在我们之前的工作中，我们观察到理性工程改造的缬氨酸原养型CHO细胞（pMTIV）在异亮氨酸浓度降低的培养基中生长更快，我们认为这反映了途径内强烈的反馈抑制⁵。纳入这些变体使我们能够测试在异源背景下调节反馈抑制是否能改善途径功能。最后，为了可能改善途径的催化特性，我们添加了ilvG和ilvM，它们分别是ilvB和ilvN的同工酶。ilvG和ilvM共同形成AHAS II复合物，类似于ilvB和ilvN形成的复合物（AHAS I），但具有不同的催化特性——最显著的是对异亮氨酸生物合成底物2-氧代丁酸的偏好（57–180倍），而非共享的异亮氨酸/缬氨酸生物合成底物丙酮酸^33,34。此外，AHAS II对异亮氨酸和缬氨酸的反馈抑制不敏感，不同于AHAS I，后者对这两种氨基酸的抑制均有响应。最后，包含一个绿色荧光蛋白（GFP）CDS作为乘客对照。

每个CDS与强EF1a或中等强度PGK启动子配对，有或无线粒体靶向序列（MTS）。包含线粒体定位是因为区室化可以改善途径功能，并且支链氨基酸（BCAA）生物合成发生在现存真核原养型生物的线粒体中^25,35（图2c）。总之，十个CDS、两个启动子和两种定位状态产生了40个成员的TU文库。每个启动子–OLS组合和CDS各由三个条形码代表，以区分独立的整合事件。

我们使用金门克隆组装文库（方法），并进行扩增子测序以监测病毒包装前TU的分布。文库内任意两个TU之间的读取计数差异不超过两倍，表明文库分布在组成部分方面均衡良好（扩展数据图1a）。文库被包装在慢病毒中，感染了242万个CHO细胞，感染后通过qPCR估计的MOI为8.8（扩展数据图1b）。

我们选择了十个CDS、两个启动子和两个OLS（共40个转录单元），这反映了生物学相关性与实际可采样性之间的平衡。由于我们知道四个基因足以建立缬氨酸原养型表型，我们至少希望采样所有可能的四转录单元组合，在40个成员的文库中有123,410种。然而，我们也有兴趣采样拷贝数的变异，这促使我们设定目标为采样所有五转录单元组合，共有1,086,008种。这个设计空间可以在单个10厘米培养皿中采样，我们认为这适合该方法的试点规模演示。即使对文库进行适度扩展，例如增加一个额外的启动子或OLS以产生60个转录单元的文库，也会将可能的五转录单元组合数量增加到约700万，使得全面采样在实践中变得更具挑战性。这个组合空间被探索的有效程度可以使用蒙特卡洛模拟来建模，如下所述。

感染后，我们进行了扩增子测序以监测转录单元分布，并观察到EF1a启动子在有无MTS的情况下均表现不足（扩展数据图1a）。由于此前尚未在混合慢病毒文库中对PGK和EF1a进行过系统性的头对头比较，这一效应的基础仍不清楚。在我们的混合设置中，几种机制可能导致EF1a驱动的转录单元表现不足。首先，由于EF1a是比PGK更强的启动子，其招募转录机器可能干扰包装过程中病毒基因组的转录，从而降低病毒滴度^36,37,38。其次，EF1a包含一个PGK中不存在的内含子，尽管该内含子在慢病毒包装过程中通常被保留，但内含子剪接位点与病毒剪接和包装机器之间的相互作用仍可能适度损害包装效率³⁹。最后，当同一细胞中存在多个途径组件时，EF1a驱动的高水平表达可能带来适应度成本，导致在感染靶细胞后EF1a相关转录单元的耗竭。我们还发现ilvA^L447F（读数的2.6%）和ilvG（读数的1.2%）相比其他CDS表现较少。

蒙特卡洛模拟以估计引入的多样性规模

TU 代表性的这种不平衡促使我们进行 Monte Carlo 模拟，以确定涉及代表性不足的 TU 的通路组合是否可能在我们的感染实验中被捕获。我们基于在 MOI 范围 2–14 下为 CHO-Div 群体测得的条形码分布，假设整合事件服从泊松分布，模拟了 TU 的随机选择。在 MOI 为 8 时，模拟表明我们将在 242 万个感染细胞中观察到 156 万种不同的通路组合，其中 12.4% 的细胞包含我们缬氨酸生物合成通路（ilvN、ilvB、ilvC 和 ilvD）的基因内容，而 2.0% 的细胞四个基因均定位于细胞质，1.6% 的细胞四个基因均定位于线粒体（Extended Data Fig. 1c）。我们推断，这很可能构成了足够多的真实通路解决方案，使我们能够在选择后发现功能性缬氨酸生物合成通路，并随后在缬氨酸限制培养基上进行选择。

更广泛地说，这些模拟说明了文库大小、感染细胞数量和 MOI 等关键实验参数如何共同决定单次筛选中可采样的组合设计空间的比例。因此，使用相同的模拟框架可以预测任何 SGE 实验可达到的覆盖度，为设计 SGE 筛选时选择文库复杂度、MOI 和培养规模提供了一个有用的起点。

CHO 中缬氨酸原养型的重建与优化

将 CHO-Div 置于缬氨酸浓度降低（4.25 µM）的培养基中，以利于携带可共同构成缬氨酸生物合成途径的 TU 组合的细胞生长（图 3a）。同时，我们将一个对照群体（CHO-GFP）置于相同的选择方案下，该群体用仅携带 GFP 的对照病毒转导。在低缬氨酸培养基中培养 15 天后，我们在实验群体（CHO-Val）中发现了细胞的克隆性生长，而在对照 CHO-GFP 群体中则未发现（图 3b）。我们挑取了 16 个 CHO-Val 克隆，并分别在缬氨酸充足（170 µM）的培养基中传代，然后使用氨基酸缺失试验进行表征。作为细胞数量的替代指标，使用基于刃天青的试剂测量代谢活性⁴⁰，并通过测量无缬氨酸与缬氨酸充足条件下代谢活性的比值（Val 评分）来估计缬氨酸生物合成功能。如果克隆的 Val 评分至少是亲本对照的两倍，我们认为该克隆为缬氨酸原养型。所有选出的克隆（16/16）均达到该阈值（图 3c）。为了验证该功能性试验能准确反映缬氨酸原养型功能，我们进一步表征了两个克隆，CHO-Val-D1 和 CHO-Val-D3（命名规则详见方法）。我们在无缬氨酸培养基中培养这两个克隆 10 天，测量其绝对生长速率。对照细胞在无缬氨酸培养基中到第 7 天时死亡，而克隆 CHO-Val-D1 和 CHO-Val-D3 则在 10 天的无缬氨酸培养中增殖，倍增时间分别为 1.14 和 1.05 天（图 3d），相当于其在缬氨酸充足培养基中生长速率的 83% 和 80%（扩展数据图 2a）。与我们此前研究中通过理性设计改造的 pMTIV 细胞所表现出的 3.77 天倍增时间相比，这是无缬氨酸生长速率的显著改善⁵，而后者仅相当于其在缬氨酸充足条件下生长速率的 26%。

为了进一步确认细胞是内源性合成缬氨酸的，我们在无缬氨酸的“重”RPMI培养基中培养了CHO-Val-D1和CHO-Val-D3（扩展数据图2b和补充说明1）。两个克隆的¹³C_2–5缬氨酸标记率分别显著提高至96.5%和96.6%，而pMTIV细胞为32.2%，表明与理性设计的代谢途径相比，缬氨酸生物合成表型有强烈改善（图3e）。事实上，在克隆Val-D1和Val-D3中检测到的¹³C_2–5缬氨酸水平与¹³C_2–3丙氨酸水平相当，表明这些克隆中工程化的缬氨酸生物合成与哺乳动物细胞天然生物合成非必需氨基酸丙氨酸的能力相当，后者也是由丙酮酸生物合成的。

我们进行了扩增子测序，以鉴定缬氨酸原养型表型背后的转录单元（TU），并表征所有16个克隆之间的基因型差异。为了区分真正的整合事件与背景噪音，我们对TU读取计数应用了z分数阈值（扩展数据图3a）。在每个克隆内，TU条形码计数被分组为离散的丰度水平，与整合拷贝数的差异一致，使我们能够估计CHO-Val-D1和CHO-Val-D3分别含有九个和六个TU整合（扩展数据图3b）。这分别对应于约37和约23 kb的整合DNA，或约19和约11 kb的TU模块。在两个克隆中，我们都发现了对应于ilvN、ilvB、ilvC和ilvD的条形码，与理性设计的pMTIV途径的CDS内容一致，其中CHO-Val-D3还表现出ilvM（ilvN的同工酶）的条形码特征（扩展数据图3b）。为了验证这些归属，我们对CHO-Val-D1和CHO-Val-D3进行了长读长测序和RNA-seq。长读长测序确认了40个判定中的38个TU身份（扩展数据图3c,d），而RNA-seq表明条形码定义的基因型与转基因表达基本一致（扩展数据图3b,e）。

在验证了基于条形码的基因分型方法后，我们接下来通过鉴定所有原养型克隆（Val 得分 > 0.28）中存在的 CDS，定义了缬氨酸原养型背后的遗传配置。我们发现 ilvB、ilvC 和 ilvD 存在于每一个表现出原养型的克隆中，表明它们是赋予 CHO 细胞缬氨酸生物合成功能的必要条件（图 3f）；ilvN 在选择后出现在 94%（15/16）的克隆中，表明它强烈支持缬氨酸生物合成功能。这与先前的报道一致，即 ilvN 作为调节亚基增强 ilvB 的催化活性和底物特异性；单独的 ilvB 保留活性，但仅以 ilvBN 全酶约 15% 的水平运作⁴¹。在唯一缺少 ilvN 的克隆 CHO-Val-D8 中，细胞反而携带其同工酶 ilvM（补充图 1a），表明杂合 AHAS 全酶 ilvBM 可能发挥作用，此前已证明其具有功能，并与其经典的 AHAS I 和 AHAS II 对应物相比表现出不同的催化和反馈特性^41,42。

我们接下来试图通过检查强原养型克隆（Val 得分 > 0.75）中启动子–OLS 组合的 TU 条形码分布，来确定每个 CDS 的最佳启动子和定位背景。在这 13 个此类克隆中，我们使用每个 TU 的中位表示生成了复合条形码“指纹”。在该指纹中，由 PGK 启动子表达的线粒体定位的 ilvB、ilvC、ilvD 和 ilvN 尤为突出，定义了一条此前未通过理性设计确定的通路“解决方案”（图 3g）。鉴于每个 TU 有九个底层组合条形码（每个启动子–OLS 组合三个条形码，每个 CDS 三个条形码），我们能够检查通路解决方案中每个 TU 的底层条形码模式。我们发现，在强原养型缬氨酸克隆中，每个 TU 的底层条形码模式具有多样性，表明在低缬氨酸选择中受到青睐的不同克隆源自多个独立的整合事件（图 3h）。在扩增的克隆中，我们未观察到 ilvN、ilvB、ilvC 和 ilvD 共同位于细胞质的情况，而根据 MOI 为 8 时模拟随机 TU 组合的预期，所有四个基因的线粒体定位富集了 48 倍（图 3i）。

尽管线粒体定位的 PGK 驱动的 ilvA^L481F 也出现在复合条形码指纹中，但我们将其从通路解决方案中排除，因为它在起始 CHO-Div 群体中过度代表（扩展数据图 1a）。MOI = 8 时的模拟预测其存在于 59% 的克隆中（补充图 1b），这与它在所有缬氨酸原养型克隆中的存在很好地相关，表明在还原缬氨酸条件下选择后，它在我们的筛选中并未相对于预期过度代表。它在强原养型克隆 CHO-Val-D1 和 CHO-Val-D3 中的缺失进一步支持了这一结论（扩展数据图 3b）。

基于这些结果，我们尝试在单一DNA构建体上重建这一优化途径。该设计与我们之前的途径pMTIV相匹配，使用相同的调控背景和2A核糖体跳跃肽序列（图2a），但包含线粒体定位版本的ilvN、ilvB、ilvC和ilvD。该途径以约3-kb的片段合成，利用S. cerevisiae中的同源重组进行组装⁴³，并通过Flp-In重组递送至CHO细胞⁴⁴（图3j）。SGE指导的设计使CHO细胞能够在无缬氨酸的RPMI中以每倍增1.25天的速度生长，相当于其在缬氨酸充足条件下生长速率的74%，显著优于pMTIV细胞（图3k）。总之，这些结果表明，SGE能够在庞大的组合搜索空间内为复杂的代谢工程挑战发现功能性途径架构，并使其在单一DNA构建体上重建，从而赋予所需的细胞功能。

发现赋予CHO异亮氨酸原养型的代谢途径解决方案

为了鉴定赋予异亮氨酸原养型的TU组合——这是我们通过理性设计无法实现的结果——我们将CHO-Div细胞置于降低异亮氨酸的RPMI条件下（分别为1.5、3和6 µM）。同时，我们将对照CHO-GFP群体置于相同的选择方案下。在低异亮氨酸培养基中培养15天后，我们发现在所有条件下实验群体（CHO-Ile）中均出现大量克隆性生长，而对照CHO-GFP群体中则没有。在三种异亮氨酸条件下，我们挑选了46个克隆，并在异亮氨酸充足（60 µM）培养基中分别传代，然后进行功能表征。如前所述，我们通过测量每个克隆在无异亮氨酸培养基中培养的代谢活性，并相对于异亮氨酸充足培养基中的代谢活性进行归一化，来估计每个克隆的异亮氨酸生物合成能力（Ile评分）。总体而言，41%的克隆（19/46）表现出相对于对照改善的Ile评分（图4a）。我们进一步表征了两个克隆，CHO-Ile-10-H1和CHO-Ile-10-H2（命名规则详见方法），它们分别表现出强和中等程度的原养型。我们在无异亮氨酸培养基中培养这两个克隆10天，并在此期间测量绝对细胞数。对照细胞在第6天死亡，而克隆CHO-Ile-10-H1在无异亮氨酸培养基中增殖10天，倍增时间为1.69天，相当于其在异亮氨酸充足培养基中生长的48%，而CHO-Ile-10-H2最初在无异亮氨酸培养基中表现出相似的增长速率，但在培养第八天后逐渐减弱，在10天培养期间总体增殖速率为每倍增3.04天，相当于其在异亮氨酸充足培养基中生长速率的33%（图4b和扩展数据图4a）。

为了进一步确认细胞是内源性合成异亮氨酸，我们在重RPMI培养基中培养CHO-Ile-10-H1和CHO-Ile-10-H2（扩展数据图4b）。在两个克隆中，提取离子色谱、质谱（MS1和MS2）分析揭示了与M+2和M+6异亮氨酸一致的特征，如果异亮氨酸是内源性合成的，我们预期检测到这些同位素体（图4c,d、扩展数据图4c和补充说明2）。对于克隆CHO-Ile-10-H1和CHO-Ile-10-H2，分别有78.5%和39.5%的异亮氨酸为¹³C_2–6，后者在重复实验中表现出显著的标记变异性，与其较低效的原养型功能一致（图4e）。总之，这些克隆展示了哺乳动物细胞中的异亮氨酸原养型，这一代谢功能上次在哺乳动物谱系中出现是在5亿多年前。

与上述缬氨酸原养型克隆一样，我们对所有46个克隆（原养型和营养缺陷型）进行了扩增子测序，以鉴定异亮氨酸原养型背后的TU组合。应用z分数阈值以区分真实整合与背景（扩展数据图5a），我们在两个克隆中均检测到ilvB、ilvC、ilvD、ilvG和ilvM，其中CHO-Ile-10-H1还携带ilvA和ilvA^L481F，CHO-Ile-10-H2携带ilvN（扩展数据图5b）。基于相对条形码丰度水平，我们估计CHO-Ile-10-H1和CHO-Ile-10-H2分别含有13和10个TU整合，对应约52 kb和约41 kb的整合DNA，其中分别有26和21 kb直接由TU模块组成。

为验证我们基于条形码的基因分型方法，我们进行了长读长测序，发现其与扩增子测序在克隆CHO-Ile-10-H1和CHO-Ile-10-H2中分别对36/40和31/40个TU判定结果一致（扩展数据图5c,d）。转基因表达的RNA-seq测量结果在CHO-Ile-10-H1中与条形码定义的基因型基本一致，但在CHO-Ile-10-H2中一致性较低（扩展数据图5b,e）。这些差异与克隆异质性一致，因为长读长测序检测到的额外低丰度TU在扩增子测序中仅处于背景水平（扩展数据图5b–e），并可能导致在重复实验中观察到的生长和同位素标记变异性。关于CHO克隆中整合结构、克隆组成和全局转录组影响的全面表征见补充说明3。

我们接下来试图通过鉴定在所有强异亮氨酸原养型克隆（Ile分数 > 0.75）中普遍存在的CDS来确定异亮氨酸途径组成。引人注目的是，100%的强原养型克隆表现出对应于ilvB、ilvC、ilvD、ilvG、ilvM和ilvA^L481F的条形码（图4f）。为检查这些CDS的启动子和定位背景，我们使用强原养型克隆中TU表示的中位数生成了复合条形码指纹。与最佳缬氨酸途径配置一致，所有CDS主要以其线粒体定位配置存在，其中ilvA^L481F、ilvM、ilvB、ilvC和ilvD在PGK启动子下更受青睐，而ilvG在EF1a启动子下更受青睐（图4g）。值得注意的是，包含ilvA（催化异亮氨酸途径的第一步，即苏氨酸转化为2-氧代丁酸）对于最佳途径性能至关重要。特别是，部分反馈抗性变体ilvA^L481F在所有强原养型克隆中的富集表明，缓解反馈抑制是途径效率的关键决定因素。

We examined the nine underlying barcode possibilities for the favored pathway solution and found barcode diversity across the seven strongly prototrophic clones, suggesting that clones arose from independent integration events (Supplementary Fig. 2a). We further examined how often ilvC, ilvD, ilvB, ilvG, ilvM and ilvA^L481F were collectively observed to be cytoplasmically or mitochondrially localized in isolated clones, as well as how often they were observed irrespective of intracellular location and compared each to simulated expectation. We found that mitochondrial localization of these TUs was dramatically selected for at 217× expectation while cytoplasmic localization of all six CDSs was not observed at all across the 46 clones (Fig. 4h).

Lastly, we looked at CDS representation amongst isoleucine auxotrophic clones (Supplementary Fig. 2b). Intriguingly, we found presence of ilvB, ilvC and ilvD across 100% of auxotrophic clones indicating that there was selection for partial metabolic pathways. This is consistent with a model in which infected cells are resource sharing and metabolic intermediates are available in the medium allowing cells encoding only partial biosynthetic pathways a selective advantage. Such a mechanism would favor CDSs that encode later enzymatic steps in the biosynthetic pathway, which is what we observe with ilvD and ilvC being most overrepresented relative to simulated expectation followed by ilvM, ilvG and ilvB.

Together, these results demonstrate that SGE can uncover not only functional metabolic pathway architectures for previously intractable metabolic engineering problems but also provide insight into the mechanistic basis for their function.

Double isoleucine/valine prototrophy in CHO-Val and CHO-Ile clones

Having established that we could achieve valine and isoleucine prototrophy individually using SGE, we next asked whether these behaviors could coexist within the same cells given the substantial overlap between pathway components. To this end, we cultured all CHO-Val clones and CHO-Ile clones in a double isoleucine/valine dropout medium to assess whether dual prototrophy could arise without direct selection. Indeed, 37.5% (6/16) of CHO-Val clones and 23.9% (11/46) of CHO-Ile clones surpassed the threshold for double isoleucine/valine prototrophy (2× the metabolic activity of the parental control line; Ile/Val score > 0.28), although none met the strong double prototrophy threshold (Ile/Val score > 0.75) (Extended Data Fig. 6a). A comprehensive characterization of the barcode signatures underlying single versus double prototrophic clones is provided in Supplementary Note 4.

These results illustrate both the diversity of phenotypes accessible through SGE and the insights such diversity can provide for pathway optimization. Importantly, they underscore a key strength of the SGE framework: the optimal genetic solution for a single metabolic trait (for example, valine or isoleucine prototrophy) is not necessarily the optimal solution for combined traits (for example, valine and isoleucine prototrophy). Traditional DBT-based engineering approaches often struggle with such multiobjective tradeoffs but the scale and combinatorial depth of SGE allow for simultaneous selection across multiple traits, enabling discovery of genotypes that balance or even reconcile conflicting demands across pathways.

Engineering valine prototrophic human T cells

To examine the versatility of SGE across cell types, we wanted to ascertain whether it was possible to engineer Jurkat cells to become prototrophic for valine and isoleucine. Unlike CHO cells, Jurkat cells grow in suspension, allowing for seamless isolation, expansion and characterization of large numbers of individual clones after selection through either serial dilution or cell sorting. Using the same lentivirus library from previous experiments, we infected 3 million Jurkat cells (Jurkat-Div) and observed a similar barcode distribution to the distribution we previously observed in CHO-Div cells (Extended Data Fig. 7a). The MOI of the infection was estimated by qPCR to be 3.4 (Extended Data Fig. 7b).

Given the lower MOI relative to CHO cells, we performed Monte Carlo simulations at a range of MOIs (2–8) to determine the likelihood of finding known solutions given the measured barcode distribution (Extended Data Fig. 7c). At MOI = 4, our simulation indicated that we should see 5.9 × 10⁵ different pathways within the 3 million infected cells, of which 2.31% of cells should carry some combination of ilvN, ilvB, ilvC and ilvD, while 0.19% of cells should carry all four TUs localized to the mitochondria and 0.23% should carry all TUs localized to the cytoplasm. On the other hand, only 0.06% of Jurkat-Div cells were expected to carry any combination of our discovered isoleucine biosynthetic pathway consisting of ilvM, ilvG, ilvB, ilvC, ilvD and ilvA^L481F, with 0.00% of cells expected to carry entirely mitochondrially or entirely cytoplasmically localized versions of all six genes (Extended Data Fig. 7d). These estimates suggested that valine prototrophy would be readily accessible, whereas isoleucine prototrophy would be rare in the Jurkat-Div population.

Accordingly, Jurkat-Div cells were cultured on low-valine medium (0, 1.7 or 4.25 µM) alongside control cells infected only with GFP (Jurkat-GFP) (Extended Data Fig. 8a). At all valine concentrations, both cell populations saw a reduction in cell number over 14 days. By day 21, however, Jurkat-Div cells saw an increase in cell number across all conditions, whereas Jurkat-GFP cells did not recover. We serially diluted the resulting selected Jurkat-Div cell populations (Jurkat-Val) to isolate and expand 119 individual clones in valine-replete conditions before characterization using our amino acid dropout assay (Fig. 5a). As before, we considered a clone prototrophic if its Val score was at least twofold that of a parental control (Val score > 0.23), and 71% (85/119) of clones met this threshold (Fig. 5b). We further characterized two clones exhibiting valine prototrophy, Jurkat-Val-31 and Jurkat-Val-32, culturing both clones in valine-free medium over 11 days and measuring their absolute growth rate. While control cells exhibited stagnant cell numbers over 11 days of culture, clones Jurkat-Val-31 and Jurkat-Val-32 proliferated at 2.02 and 1.87 days per doubling, respectively, corresponding to 50% and 52% of their growth rates in valine-replete medium (Fig. 5c and Extended Data Fig. 8b).

We next asked whether isoleucine prototrophy could similarly be achieved in Jurkat cells. Jurkat-Div cells were cultured in low-isoleucine medium (0, 6 or 12 µM) alongside Jurkat-GFP controls. Across all conditions and 32 days of culture, Jurkat-Div cells failed to outcompete control cells, indicating that isoleucine prototrophic phenotypes were not present in the population (Extended Data Fig. 8c).

Together, these results highlight that simulated pathway diversity, determined using TU distributions and MOI, can be used to predict and guide pathway discovery across cell types.

Characterization of valine prototrophic human T cell clones

To identify the TU combinations underlying valine prototrophy in Jurkat cells, we performed amplicon-seq on all isolated clones, applying a z-score threshold to distinguish signal from background (Extended Data Fig. 9a). Clones Jurkat-Val-31 and Jurkat-Val-32 exhibited similar TU profiles, comprising mitochondrially localized ilvM, ilvB, ilvN, ilvC and ilvD expressed from PGK promoters, along with cytoplasmically localized ilvB in the same promoter context (Extended Data Fig. 9b,c). This configuration closely mirrors the optimal mitochondrially localized pathway architecture we identified in CHO cells.

On the basis of relative barcode abundance, we estimated that each clone harbored nine TU integrations corresponding to ~32 kb of integrated DNA, of which ~15 kb was directly made up by TU modules. We additionally performed long-read sequencing for clone Jurkat-Val-31 and discovered concordance between long-read TU calls and amplicon-seq TU calls for 40/40 TUs (Extended Data Fig. 9d,e). Barcodes mapped to their assigned promoter–OLS in 95.5% of reads and to their assigned CDS in 100% of reads (Extended Data Fig. 9f and Supplementary Table 1). Transcripts from SGE transgenes measured by RNA-seq were in good agreement with the TU barcodes detected in amplicon-seq (Extended Data Fig. 9c,g).

Next, to define the pathway underlying valine prototrophy in Jurkat cells, we identified CDSs that were universally present across all prototrophic clones. ilvB, ilvC, ilvD and ilvN were present in all Jurkat valine prototrophic clones, matching the CDS content of the solution previously identified for the CHO-Div population (Fig. 5d). To determine the optimal promoter and localization context for each CDS, we generated a composite barcode fingerprint on the basis of the median representation of each TU across clones. This analysis revealed that ilvC, ilvD and ilvN were favored in their mitochondrially localized context under control of the PGK promoter while ilvB appeared to be favored under control of a PGK promoter in both the cytoplasm and the mitochondria (Extended Data Fig. 10a).

However, when looking at the three promoter barcodes and three CDS barcodes underlying these five TUs, we noticed that many clones exhibited barcode signatures shared with other clones suggesting that multiple sequenced samples had arisen from the same founder clone (Extended Data Fig. 10b). We expect to observe such clonal ‘siblings’ if particular pathway solutions confer a strong selective advantage and undergo expansion during selection.

To quantify this effect, we examined global similarity in TU profiles across clones. We identified 69 clones with at least one sibling and 50 unique ‘singleton’ clones (Extended Data Fig. 10c). Using graph-based clustering, we found ten separate sibling groups varying in size (2–24 clones) and sibling group size did not correlate to average Val score (Extended Data Fig. 10d). Inspection of the largest sibling groups (2, 4, 9 and 5, consisting of 24, 12, ten and seven clones, respectively) revealed a consistent pattern; ilvN, ilvC and ilvD were uniformly localized to mitochondria, while ilvB was observed localized to the mitochondria, cytoplasm or both compartments (Extended Data Fig. 10e). To avoid bias from clonal expansion, we removed duplicate siblings and analyzed CDS usage across the population consisting only of unique clones, which yielded a largely unchanged CDS usage profile (Extended Data Fig. 10f; compare to Fig. 5d). We also recomputed a composite barcode fingerprint, which yielded a pathway solution consisting only of mitochondrially localized ilvN, ilvB, ilvC and ilvD (Fig. 5e). Consistent with this, the collective presence of ilvN, ilvB, ilvC and ilvD was enriched 31-fold relative to expectation, while their colocalization to mitochondria was enriched 283-fold (Fig. 5f).

We also examined CDS representation amongst 24 valine auxotrophic Jurkat clones carrying at least one integration (Supplementary Fig. 3). Unlike the auxotrophic CHO clones selected on low-isoleucine medium, which all carried a partial isoleucine biosynthetic pathway, selection for partial biosynthetic pathways in Jurkat clones selected on low-valine medium was much less pronounced with slight CDS overrepresentation of ilvD, ilvC and ilvB relative to expectation. This may reflect the lower MOI used to infect Jurkat cells or that clonal resource sharing is less pronounced in suspension cell culture.

These results highlight the generalizability of SGE across cell types including rodent and human cell lines and its ability to uncover pathway solutions that converge on shared design principles despite differences in cellular context.

Machine learning to predict prototrophy phenotypes from SGE-generated data

Lastly, we asked whether machine learning could be used to predict prototrophy phenotypes directly from the TU composition of each clone. We decided to apply a random forest model for classification because of its strength in interpreting nonlinear relationships between features, its robustness to overfitting and its ability to provide feature importance scores (Fig. 5g). As input features, we used the proportion of reads assigned to each TU within a clone (analogous to the clone-level TU abundance heat maps). Each clone’s phenotypic performance was binarized as prototrophic or auxotrophic using the Val score threshold of 0.23. The samples were partitioned into four mutually exclusive class-balanced folds to perform stratified fourfold cross-validation (Supplementary Fig. 4). For each fold, the model was trained on the other three folds and evaluated on the held-out fold for four separate 75:25 train–test splits. When looking at metrics across folds, the model successfully classified samples into prototrophic or auxotrophic with a mean accuracy of 0.942 ± 0.005, a mean area under the receiver operating characteristic (ROC) curve of 0.993 ± 0.006 and a mean area under the precision–recall curve of 0.997 ± 0.002 across folds, indicating strong discriminative ability and stability across different train–test sample compositions (Fig. 5h).

In total, 112/119 clones were correctly classified and the top four features driving classification were mitochondrially localized ilvN, ilvC, ilvD and ilvB under control of a PGK promoter, matching the pathway solutions that we discovered by manual analysis and interpretation. This classification accuracy also held true when removing duplicate sibling clones from the dataset before model training; the same top four features were discovered and correct classifications were produced for 53/59 clones (Supplementary Fig. 5a,b).

Five auxotrophic clones were misclassified as prototrophic; two (Jurkat-Val-55 and Jurkat-Val-104) carried all four CDSs that form our discovered pathway solution and likely reflect examples of gene silencing or barcode swapping and two clones (Jurkat-Val-79 and Jurkat-Val-113) lacked ilvN only, potentially reflecting insufficient AHAS activity for prototrophic function, while the final clone (Jurkat-Val-110) lacked both ilvB and ilvD (Supplementary Fig. 6a). On the other hand, two prototrophic clones were misclassified as auxotrophic. Clone Jurkat-Val-60 (Val score: 0.31) appears to be clonally impure, carrying very lowly abundant ilvC and ilvD, which the model may not consider sufficient to generate a prototrophic phenotype. Clone Jurkat-Val-68 (Val score: 1.05) carried ilvN, ilvC and ilvD under control of PGK-MTS, as found in our discovered optimal pathway solution, while ilvB was under the control of EF1a-MTS, which is a much less represented possibility in our starting SGE library and may, therefore, not have been classified as contributing to the valine prototrophic phenotype (Supplementary Fig. 6b).

These results demonstrate that SGE-generated data are well suited for machine-learning-based modeling and such models can effectively extract biologically meaningful features, enabling pathway-scale analysis for solutions to complex metabolic engineering problems.

Discussion

Here, we introduce SGE, a method for efficiently exploring vast combinatorial design spaces comprising multiple interacting TUs in mammalian cells. By leveraging single cells as independent experiments and using random high-copy integration of individual TUs per cell, SGE enables high-throughput assembly and screening of millions of synthetic metabolic pathways. This approach bypasses the inefficiencies associated with building and delivering large DNA constructs, substantially accelerating pathway discovery and optimization.

Using this approach, we generated and screened millions of synthetic metabolic pathways to engineer EAA prototrophy in mammalian cells. We achieved (1) near-wild-type growth in valine-free medium and (2) isoleucine prototrophy in mammalian cells. The functional solutions involved integration of 23–52 kb of synthetic DNA—construct sizes impractical to screen using conventional methods. We also provide initial evidence that leveraging organellar compartmentalization can enhance engineered pathway functionality in mammalian cells. Lastly, we demonstrate the versatility of SGE across mammalian cell types (species, in suspension or adhesion) and present a pilot integration with machine learning to predict functional pathway combinations.

A consistent theme across the metabolic engineering efforts described here was the preference for mitochondrial localization of BCAA pathway enzymes. This was unexpected given that the genes in our library were sourced from E. coli, a prokaryote lacking subcellular compartments. However, mitochondria are thought to be of bacterial origin and share notable protein composition with bacteria, which may help explain this preference⁴⁵. Equivalent BCAA biosynthesis genes in extant eukaryotic BCAA prototrophic microbes such as fungi are also localized to mitochondria, providing a clue that there may be biochemical advantages associated with mitochondrial localization, such as increased pyruvate concentrations relative to cytosol^46,47 or greater availability of relevant pathway cofactors (for example, iron–sulfur clusters and thiamine diphosphate)^25,48,49. Similarly, in plants, these BCAA biosynthetic enzymes are targeted to chloroplasts, which are also an abundant source of pyruvate⁵⁰. Another possibility is that BCAA biosynthetic enzymes simply benefit from being much more concentrated in a limited organellar physical space, thereby boosting pathway flux and function, as previously demonstrated for other pathways localized to the mitochondria in S. cerevisiae³⁵. This finding highlights the utility of incorporating organellar diversity in the SGE screen and encourages incorporation of yet more organellar targeting signals in future engineering efforts in mammalian cells.

Use of SGE comes with several limitations, most notably the requirement for a selectable phenotype. Essential nutrients provide a built-in live–dead screen that effectively identifies cells carrying optimal solutions. However, when the desired phenotype is not linked to cell survival, alternative selection strategies such as coupling pathway activity to fluorescent biosensors, surface marker expression or engineered genetic circuits that connect pathway output to a selectable reporter may be applied^51,52. Another limitation of SGE as currently implemented is the use of lentivirus for random integration. We observed substantial biases in TU representation introduced during viral packaging. Specifically, EF1a-driven constructs were markedly underrepresented compared to those with PGK promoters and similar biases were noted for certain CDSs in ways not attributable solely to size. Transitioning to transposon-based delivery systems could alleviate this issue. Lastly, while amplicon-seq can inform what TUs have been integrated in clones of interest, it provides little information about which genes are expressed. In cell lines prone to silencing, this may confound interpretation. Pairing SGE with either bulk or single-cell RNA-seq would provide expression-level information, enabling more accurate assessment of functional pathway configurations.

While SGE excels at rapidly exploring vast design spaces through random integration, successful multi-TU solutions identified in this manner may need to be reconstituted into compact, single-vector formats for practical deployment. However, TUs may behave differently when assembled into condensed architectures⁵³. A hybrid strategy may be effective in this case, using SGE for high-throughput discovery, followed by CLASSIC¹² or related methods for validation and optimization in compact constructs in lower throughput.

Despite its current limitations, SGE represents a powerful strategy for the scalable engineering of increasingly complex biosynthetic traits in mammalian cells. In this study, we demonstrated the discovery of a functional six-gene isoleucine biosynthesis pathway from a 40-member TU library. We anticipate that SGE can be extended to larger libraries by proportionally increasing the MOI during integration, enabling access to even more complex phenotypes. These may include prototrophy for additional EAAs, growth factors and vitamins, enhanced tolerance to environmental stressors such as hypoxia, shear stress and metabolic byproducts or rewiring of metabolic programs toward alternative cell states or identities. The inclusion of elements that modulate host factors (CRISPRa/i machinery) would further expand the design space explored by this approach.

Ultimately, the ability to engineer mammalian cells with complex pathways has broad implications across basic research, biotechnology and medicine. For instance, a mammalian cell engineered to be prototrophic for all amino acids would require a reduced number of nitrogen sources, opening the door to using stable isotope labeling to chart presently unknown corners of the mammalian metabolic map⁵⁴. Meanwhile, mammalian cells engineered to become independent of specific nutrients could lower media costs and simplify formulations for biomanufacturing applications such as cultivated meat and viral vector production^55,8. Lastly, cell therapies may be metabolically augmented to better withstand the hostile tumor microenvironment, which is marked by nutrient deprivation, acidosis, oxidative stress and inflammation, with potential for improving the performance of cell-based therapeutics used in the treatment of solid tumors^56,57,58,59.

Methods

Cell lines, media, cell counting and images

CHO Flp-In cells (Thermo Fisher, R75807) and Jurkat cells (M. Pacold lab, New York University) were used in all experiments. All cell lines tested negative for Mycoplasma. For growth assays involving amino acid dropout formulations, medium was prepared from an amino-acid-free RPMI 1640 powder base (US Biological, R8999-20) and custom combinations of amino acids were added back in as needed to match the standard amino acid concentrations for RPMI 1640 except for isoleucine, which was considered 100% supplemented at 0.06 mM. Additionally, RPMI medium was supplemented with 2 mM sodium pyruvate as previously described⁵. Custom amino acid dropout medium was adjusted to a pH of 7.3 with HCl, sterile-filtered and supplemented with 10% dialyzed FBS (Fisher Scientific, 26400044) and penicillin–streptomycin (100 U per ml) before use. For metabolomics experiments, medium was prepared from an amino-acid-free, glucose-free RPMI 1640 powder base (US Biological, R9010-01) and custom combinations of amino acids were added in to match the standard concentrations for RPMI 1640. In the place of glucose, isotopically heavy ¹³C_U glucose (Cambridge Isotope Labs, CLM-1396-PK) and ¹³C₃ sodium pyruvate (Millipore Sigma, 490717) were added at 11.1 mM and 2 mM, respectively. Cell counts were performed using a Countess 3 FL. Images of clones were captured on an EVOS M5000 using a ×10 objective and scale bars were added using Fiji.

SGE library construction

Promoters and OLSs were PCR-amplified and allotted unique barcodes using PCR primers encompassing said barcodes and BsmBI restriction sites. Each barcoded promoter–OLS combination was cloned into an assembly vector with lentiviral long terminal repeats using Gibson assembly. CDSs of interest were codon-optimized and synthesized commercially. Each CDS was allotted three unique barcodes by PCR amplification with primers encompassing said barcodes along with BsmBI restriction sites to facilitate downstream Golden Gate cloning. Each PCR-amplified, barcoded CDS was cloned into a pCR-Blunt-II-TOPO vector. Concentrations of each TOPO-cloned CDS were measured using a Qubit fluorometer before pooling of all CDSs of interest at equimolar concentration. Concentrations of each promoter–OLS plasmid were similarly measured using a Qubit fluorometer before pooling of all promoter–OLS plasmids at equimolar concentration. The pooled CDS mixture was spiked into the Golden Gate reaction at 5× the molar concentration of the pooled assembly vector mix, which was used at 100 ng. We added 1.5 μl of 10× NEB T4 ligase buffer, 0.15 μl of BSA, 1.1 μl of BsmBI-v2 (10,000 U per ml), 1 μl of T4 ligase (20,000 U per ml) and then supplemented with H₂O to a total volume of 15 μl. Reactions were run for 90 cycles of 42 °C for 3 min and 16 °C for 4 min before one cycle of 50 °C for 5 min and 80 °C for 5 min. Resulting libraries were transformed into ElectroMAX Stbl4 E. coli (Invitrogen, 11635-018) for amplification before library extraction and amplicon-seq to measure library distributions. Promoter–OLS and CDS sequences are shown in Supplementary Table 8.

Lentiviral packaging and infection

Lentivirus was packaged by plating 4 × 10⁶ HEK293T cells on 10-cm plates and incubating cells overnight at 37 °C. Cells were transfected with a plasmid mix consisting of 3.5 µg of the SGE library, 6.0 µg of psPAX2 (Addgene, 12260) and 3.0 µg of pMD2.G (Addgene, 12259) using Lipofectamine 2000 (Thermo Fisher Scientific, 11668019) in accordance with the manufacturer’s instructions. Transfected HEK293T cells were incubated for 48 h, before medium was collected and replenished. Cells were then incubated for an additional 24 h before medium was collected again. Collected medium was centrifuged at 200g for 5 min and the resulting supernatant was filtered using a 0.45-µm filter. Virus was concentrated using Millipore Amicon Ultra-15 centrifugal filter units (Millipore Sigma, UFC903024) for 30 min at 4,000g before storage at −80 °C until use. For CHO infections, cells were plated at 1.5 M cells per 10 cm and incubated overnight before infection. At the time of infection, the medium was replaced with fresh medium containing polybrene at 8 µg ml⁻¹ and concentrated virus. Cells were incubated with the virus for 24 h before replenishing with fresh medium. For Jurkat infections, 3 million cells were resuspended in 6 ml of medium containing polybrene at 8 μg ml⁻¹ and concentrated virus before centrifugation at 931g for 2 h at 30 °C. An additional 6 ml of medium (not containing polybrene) was subsequently added to cells dropwise before 24 h of incubation at 37 °C, after which medium/virus was replenished with fresh medium.

MOI estimation by qPCR

gDNA was extracted from cells using the QIAamp DNA mini kit (Qiagen, 51304), in accordance with the manufacturer’s protocol. qPCR reactions were performed as 10-μl reactions with 0.5 µl of 5 µM primer mix using SYBR green master I on a QuantStudio 7 Pro. Primers were designed to amplify amplicons 60–120 bp in size. MOI was estimated by measuring amplification of a sequence universally found in each TU found in the SGE library and relativizing to amplification of CHO targets, Dhfr and Hprt (expected to be single copy), and Jurkat targets, RPP30 and ALB (expected to be found in two copies), using the ∆C_t method. All reactions were performed in triplicate.

Monte Carlo simulation of TU integrations across variable MOIs

To model the distribution of TU combinations integrated into a cell population, we performed Monte Carlo simulations assuming Poisson-distributed MOIs. For CHO cells, we simulated 2,420,000 cells at MOI values of 2, 4, 6, 8, 10, 12 and 14, spanning a range around the estimated MOI of 8.8. For Jurkat cells, 3,000,000 cells were simulated at MOIs of 2, 4, 6 and 8, centered around the estimated MOI of 3.4. Integration events were simulated by first drawing the number of integrations per cell from a Poisson distribution. For cells with one or more integrations, TUs were sampled with replacement according to empirically defined distributions. The resulting sets of integrated TUs were sorted and recorded as combinations, irrespective of integration order. For each MOI, the frequency of unique TU combinations was tallied and used to benchmark the probability of observing specific TU sets. These simulations were used not only to quantify diversity in our study but also to provide a general framework for designing SGE experiments. By varying the library size, MOI and number of simulated cells, one can estimate the achievable coverage of TU combinations under their specific experimental parameters. The analysis code used to run these simulations is available in the GitHub repository associated with this paper.

Clonal isolation and amino acid prototrophy screens

After selection on the indicated amino acid dropout medium, CHO clones were manually isolated using a P200 pipette tip, transferred to a 96-well plate and expanded before phenotypic screening. For each selection condition (CHO-Val in 4.25 µM valine RPMI; CHO-Ile in 1.5, 3 and 6 µM isoleucine RPMI), clones were distributed across two separate 10-cm dishes to allow sampling of all infected cells but minimize risk of clones merging. Clones from each dish were assigned to a separate row of the 96-well plate and named according to plate position (for example, C1–C8 and D1–D8). CHO-Ile clones were further labeled with numeric identifiers corresponding to the isoleucine concentration used for selection (2.5%, 5% and 10% of 0.06 mM). For instance, CHO-Ile-10-XX refers to a CHO clone subject to 10% Ile RPMI or RPMI with 6 µM Ile. For Jurkat cells, clones were isolated by serial dilution and numbered sequentially.

We screened for amino acid prototrophic clones using PrestoBlue HS cell viability reagent (Thermo Fisher, P50200), a resazurin-based viability reagent that fluoresces in response to cellular reducing activity. Fluorescence intensity is proportional to the number of metabolically active (viable) cells in the medium. For each amino acid dropout condition, we relativized triplicate wells for each clone in the amino acid dropout condition to triplicate wells in an amino-acid-replete condition to control for differences in overall growth rate and seeding density. Amino-acid-replete medium was made by making amino acid dropout medium and adding in the missing amino acid. Fluorescence measurements were conducted on an Agilent Biotek Synergy Neo2.

Amplicon-seq library preparation

gDNA from clones of interest was prepared using the Zymo Quick-DNA 96 Plus kit (Zymo, D4070) in accordance with the manufacturer’s protocol. PCR1 (UMI addition) was performed using primers annealing to a universal primer sequence found in the 3′ UTR across all SGE TU vectors. PCR1 used 250 ng of gDNA template, 2.5 µl of each 10 µM primer and Q5 polymerase in a total reaction volume of 50 µl. PCR1 cycling conditions were as follows: 98 °C for 5 min, followed by four cycles of 98 °C for 20 s, 65 °C for 20 s and 72 °C for 30 s, with a final extension at 72 °C for 60 s. PCR cleanup was performed using AmpureXP beads (0.8×) and eluted in 15 µl of 0.1× TE. For PCR2 (sample indexing and sequencing adaptor addition), 10 µl of eluate from the previous reaction was used in another 50-µl reaction using Q5 polymerase and 5 µl of a 5 µM indexing primer mix. SYBR green dye was also added to the reaction mix to allow us to monitor the reaction progression and stop reactions before saturation (typically 16–19 cycles). PCR2 cycling conditions were as follows: 98 °C for 5 min, followed by 16 cycles of 98 °C for 10 s, 65 °C for 10 s and 72 °C for 20 s. PCR cleanup was again performed using AmpureXP beads (0.8×) and eluted in 15 µl of 0.1× TE. Sequencing was performed on a NextSeq500 75 cycle kit to yield paired-end 36-bp reads.

Amplicon-seq analysis

FASTQ files from paired ends were merged using pear (version 0.9.11). Stitched FASTQ files were processed to recover known barcode combinations identifying promoter–OLS combinations and CDSs to quantify TU integration events. Reads were parsed to identify two universal priming regions flanking the barcode segments. For reads containing both universal primers, two 8-bp barcodes were extracted and concatenated to form a unique identifier for each TU.

Each extracted barcode combination was matched against a known barcode reference library for exact matches. However, we noticed that a subset of clones exhibited a large proportion of reads that did not match any in the reference library but exhibited a 1-bp mismatch relative to reads that were found in the reference library. As no two barcodes in the library exhibited a Hamming distance < 2 and 99.9% of possible barcode pairings exhibited a Hamming distance ≥ 3, we opted to allow for 1-bp mismatches if possible to do so unambiguously. Briefly, if no exact barcode match was found, sequences were compared to a precomputed lookup table containing all possible 1-bp mismatch variants of the reference barcodes. Reads with a unique (unambiguous) 1-bp mismatch were assigned to the corresponding reference barcode, while ambiguous 1-bp mismatches (matching multiple barcodes) were excluded from downstream quantification. Samples with fewer than 50,000 matching reads were further excluded from downstream analysis.

For analysis of CDS presence within each clone, read counts for each TU were used to calculate a z score. Given three barcodes per promoter–OLS combination and three barcodes per CDS, each TU had nine underlying barcodes. Read counts for each TU were aggregated and log-transformed before calculating a z score for each TU on a per sample basis to separate true signal from background. For generation of per-clone heat maps, the percentage of reads assigned to a given TU within a sample was calculated, normalized to total reads per sample. For generation of heat maps to identify optimal pathways, the median percentage representation of each TU across all indicated clones was displayed as a composite heat map. Raw read counts are shown in Supplementary Table 9 and TU aggregated read counts are shown in Supplementary Table 10.

Metabolomics

Cells were cultured in RPMI medium containing ¹³Cglucose and ¹³Csodium pyruvate with or without valine/isoleucine (as indicated) before cell harvest. Cell pellets were generated by trypsinization, followed by low-speed centrifugation, and the pellet was frozen at −80 °C until further processing. A metabolite extraction was carried out on each sample with an extraction ratio of 1 × 10⁶ cells per ml (80% methanol containing internal standards, 500 nM). The liquid chromatography (LC) column was a Millipore ZIC-pHILIC (2.1 × 150 mm, 5 μm) coupled to a Dionex Ultimate 3000 system and the column oven temperature was set to 25 °C for the gradient elution. A flow rate of 100 μl min⁻¹ was used with the following buffers: (A) 10 mM ammonium carbonate in water (pH 9.0) and (B) neat acetonitrile. The gradient profile was as follows: 80–20% B (0–30 min), 20–80% B (30–31 min) and 80–80% B (31–42 min). The injection volume was set to 1 μl for all analyses (42 min of total run time per injection). MS analyses were carried out by coupling the LC system to a Thermo Q Exactive HF MS instrument operating in heated electrospray ionization mode. MS was performed for 30 min with a polarity switching data-dependent top 3 method for both positive and negative modes, with targeted MS2 scans for the monoisotopic, ¹³C_U and ¹³C_U/¹⁵N_U valine m/z values. The spray voltage for both positive and negative modes was 3.5 kV and the capillary temperature was set to 320 °C with a sheath gas rate of 35, auxiliary gas rate of 10 and maximum spray current of 100 μA. The full MS scan for both polarities were conducted at 120,000 resolution with an automatic gain control (AGC) target of 3 × 10⁶ and a maximum injection time (IT) of 100 ms; the scan range was 67–1,000 m/z. Tandem MS spectra for both positive and negative mode used a resolution of 15,000, AGC target of 1 × 10⁵, maximum IT of 50 ms, isolation window of 0.4 m/z, isolation offset of 0.1 m/z, fixed first mass of 50 m/z and three-way multiplexed normalized collision energies of 10, 35 and 80. The minimum AGC target was 1 × 10⁴ with an intensity threshold of 2 × 10⁵. All data were acquired in profile mode. All valine/isoleucine data were processed using Thermo XCalibur Qualbrowser for manual inspection and annotation of the resulting spectra and peak heights referring to authentic valine/isoleucine standards and labeled internal standards as described.

Identification and analysis of ‘sibling’ clones

To detect samples with globally similar barcode profiles, all pairwise comparisons between samples were computed using a Pearson correlation coefficient. Pairs were retained if each sample had at least two nonzero barcode features (to avoid spurious correlations from sparse data) and the Pearson correlation coefficient between the sample profiles was ≥0.9. To organize putative sibling clones identified through barcode profile correlation into groups, we applied a graph-based clustering approach. The set of high-correlation sample pairs was used to construct an undirected graph in which each node represented a sample and each edge represented a strong global match using the NetworkX Python package, with edges connecting sample pairs that met the correlation threshold. Connected components within the graph were then identified, each representing a cluster of samples inferred to be derived from the same original clone or barcoded integration event. Samples not belonging to any component were defined as singletons.

Random forest classification

To identify TU features predictive of valine prototrophy, we trained a random forest classifier using relative barcode read counts as input features. Relative barcode read counts were calculated as the absolute read count for each TU per clone normalized by the total barcode read count for that clone. In total, 119 samples represented by 40 features (that is, TUs) were used to train the model. A binary phenotype label was defined by thresholding the Val score at 0.23, classifying samples as either prototrophic (Val score > 0.23) or nonprototrophic. For each clone, the random forest classifier output an estimated probability of being prototrophic or nonprototrophic, representing the averaged class likelihood produced across all decision trees. Discrete class predictions were obtained by assigning clones to the class with the higher predicted probability. We used stratified fourfold cross-validation to ensure balanced class distributions across splits. In each fold, a random forest classifier was trained on 75% of the data and evaluated on the remaining 25%. Model performance was assessed using confusion matrices, classification reports, area under ROC curves, area under precision–recall curves and average precision. Confusion matrices were plotted per fold and mean performance metrics were reported across folds with s.d. Feature importances were extracted from each trained model to identify the most predictive TUs. The top ten features were reported per fold and mean feature importances were calculated to identify features consistently ranked as important across folds.

RNA-seq

RNA-seq was performed commercially (Plasmidsaurus), yielding at least 10 million single-end reads per sample following transcript purification with poly(A) mRNA capture and cDNA synthesis with oligo-dT primers. Given that SGE transgenes were not polyadenylated (to facilitate use of lentivirus as a delivery method), detection of SGE transgenes using this methodology is underpowered and expected to yield lower reads per kilobase per million mapped reads (RPKM) than endogenous polyadenylated transcripts. Each sample type was submitted in duplicate. Resulting raw fastq files were subjected to adaptor trimming and quality control using trim_galore (version 0.6.10) and cutadapt (version 4.4) with default parameters. For transgene quantification, trimmed reads from CHO samples were aligned to a custom reference genome based on the Cricetulus griseus assembly CriGri-PICRH-1.0 (GCF_003668045.3) to which SGE transgene sequences (ilvA, ilvB, ilvC, ilvD, ilvG, ilvM, ilvN and EGFP) were appended as separate contigs. Trimmed reads from Jurkat samples were aligned to a custom human genome reference based on GRCh38/hg38, to which SGE transgene sequences were appended as separate contigs. Reads were aligned to custom genomes using STAR (version 2.7.11b) using default STAR parameters. For each sample, the number of reads mapped to each contig was extracted from the BAM files using SAMtools (version 1.21) idxstats. This produced per-contig counts along with contig lengths. To account for both sequencing depth and differences in transgene length, expression levels were quantified using RPKM, providing a length-normalized measure of transgene expression comparable across samples and transgenes.

For differential expression analysis, trimmed reads from CHO samples were quasi-mapped against the C. griseus transcriptome derived from assembly CriGri-PICRH-1.0 (GCF_003668045.3) to which SGE transgenes were appended. Trimmed reads from Jurkat samples were quasi-mapped against a Gencode v49 human transcriptome reference to which SGE transgenes were appended. Quasi-mapping was conducted using Salmon (version 1.10.1) with a k-mer size of 25. Transcript abundances were imported to R using tximport and differential expression analysis was performed using DESeq2. Genes with adjusted P < 0.05 and |log₂FC (fold change)| > 1 were considered significantly differentially expressed.

For Gene Ontology (GO) enrichment analysis in CHO samples, functional enrichment was performed using mouse orthologs as a proxy. Gene symbols for genes with adjusted P < 0.05 and |log₂FC| > 1 were mapped to mouse Entrez identifiers using clusterProfiler against the org.Mm.eg.db database. GO categories were considered significantly enriched at adjusted P < 0.05. For GO enrichment analysis in Jurkat samples, functional enrichment was performed for genes with adjusted P < 0.05 and |log₂FC| > 1 using clusterProfiler against the org.Hs.eg.db database.

Long-read sequencing

gDNA was isolated from 3–6 million cells using either the QIAamp DNA mini kit (Qiagen) or the Monarch genomic purification kit (New England Biolabs), incorporating an RNase treatment phase in all instances. To achieve target fragment lengths, 2 µg of gDNA was diluted in 150 µl of nuclease-free water and sheared using g-TUBE (Covaris) at 3,000g. Following shearing, the DNA was purified using the Zymo DNA clean and concentrator kit and analyzed on an Agilent TapeStation 4150 (Genomic DNA ScreenTape) to verify the distribution.

Nanopore libraries were constructed using the Oxford Nanopore native barcoding kit 24 V14 (SQK-NBD114.24). Sample purity and concentration were validated by NanoDrop spectrophotometry and Qubit fluorometry, respectively. Sequencing was performed by loading 80–100 fmol of the library onto R10 MinION flow cells or 100–120 fmol onto PromethION flow cells. To optimize throughput, flow cells were washed and reloaded once pore activity reached approximately 30%.

Raw Oxford Nanopore POD5 files were processed using the Dorado analysis engine (https://github.com/nanoporetech/dorado/releases/tag/v1.0.0). The POD5 files were basecalled again using the super model (dna_r10.4.1_e8.2_400bps_sup@v5.2.0). The resulting BAM files were demultiplexed into a separate BAM file per barcode using Dorado demux. Each BAM file was aligned against its corresponding reference FASTA files using the Dorado align command. Per-base coverage depth tracks were generated using BEDTools (version 2.30.0)⁶².

BAM files were analyzed to assign TU identities on a per-read basis using contig alignments. Reads were filtered to retain alignments with MAPQ ≥ 50 and length ≥ 200 bp. Primary and supplementary alignments were included, whereas secondary alignments were excluded.

For transgene quantification, PGK-MTS (508-bp promoter + 108-bp MTS) and EF1a-MTS (1,187-bp promoter + 108-bp MTS) sequences were used as promoter contigs. Reads were classified as no-MTS or MTS promoter variants according to whether alignments to these contigs extended beyond positions 550 and 1230 for PGK and EF1a, respectively. For ilvA, ilvA^L447F and ilvA^L481F, the native ilvA sequence was used as the alignment contig and variant identity was resolved by scanning each read for 15-bp motif windows corresponding to native or mutant sequence states at contig positions 1341 and 1443, respectively (ilvA: native/native; ilvA^L447F: mutant/native; ilvA^L481F: native/mutant), allowing a Hamming distance of 1. After promoter and CDS assignment, only reads with exactly one promoter call and exactly one CDS call were retained for TU quantification.

For estimation of barcode swapping rates, we additionally filtered reads to retain those with high-confidence alignments to both DNA sequences flanking the 127-bp SGE amplicon allowing a maximal distance of 150 bp between the two alignments. Alignments were used to reconstruct full-length read sequences on the basis of hard clipping information contained within the CIGAR string. Reconstructed reads were parsed to identify the two universal priming regions flanking the 8-bp SGE barcodes, allowing a Hamming distance of 5, and the corresponding 8-bp barcodes were matched to their assigned promoter–OLS or CDS, allowing a Hamming distance of 1.

For identification of lentiviral integration sites, we looked for split-read junctions between integrated sequences (that is, TU modules and lentiviral sequences) and the host genome. Candidate junctions were retrieved from reads in which both alignments had MAPQ ≥ 50 and length ≥ 100 bp; one genomic breakpoint was assigned per read. Breakpoints were clustered by genomic position within a 10-kb window to define candidate integration sites. A candidate integration site was retained if its breakpoint cluster was supported by at least two reads. Each site was mapped to its relevant reference genome annotation (CriGri-PICRH-1.0 for CHO, GRCh38/hg38 for Jurkat) using ChIPseeker and classified by genomic feature (promoter, exon, intron or intergenic).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability

Amplicon, long-read sequencing data were deposited to the Sequence Read Archive under BioProject PRJNA1282762 (ref. ⁶³). Metabolomics data were deposited to Zenodo (https://doi.org/10.5281/zenodo.21874813)⁶⁴.

Code availability

Computational pipeline for long-read sequencing can be accessed from GitHub (https://github.com/FenyoLab/ont_analysis_pipeline). All other scripts written to analyze data and generate figures can also be accessed from GitHub (https://github.com/julietrolle/shotgun-genetic-engineering).

References

1. Srinivasan, P. & Smolke, C. D. Biosynthesis of medicinal tropane alkaloids in yeast. Nature 585, 614–619 (2020).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

2. Luo, S. et al. Construction and modular implementation of the THETA cycle for synthetic CO₂ fixation. Nat. Catal. 6, 1228–1240 (2023).

   Article  CAS  Google Scholar 

3. Zhang, J. et al. A microbial supply chain for production of the anti-cancer drug vinblastine. Nature 609, 341–347 (2022).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

4. Stout, A. J., Mirliani, A. B., Soule-Albridge, E. L., Cohen, J. M. & Kaplan, D. L. Engineering carotenoid production in mammalian cells for nutritionally enhanced cell-cultured foods. Metab. Eng. 62, 126–137 (2020).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

5. Trolle, J. et al. Resurrecting essential amino acid biosynthesis in mammalian cells. eLife 11, e72847 (2022).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

6. Zhao, Y. et al. IL-10-expressing CAR T cells resist dysfunction and mediate durable clearance of solid tumors and metastases. Nat. Biotechnol. 42, 1693–1704 (2024).

   Article  CAS  PubMed  Google Scholar 

7. Lim, W. A. The emerging era of cell engineering: Harnessing the modularity of cells to program complex biological function. Science 378, 848–852 (2022).

   Article  CAS  PubMed  Google Scholar 

8. Collins, L. T., Ponnazhagan, S. & Curiel, D. T. Synthetic biology design as a paradigm shift toward manufacturing affordable adeno-associated virus gene therapies. ACS Synth. Biol. 12, 17–26 (2023).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

9. Petzold, C. J., Chan, L. J. G., Nhan, M. & Adams, P. D. Analytics for metabolic engineering. Front. Bioeng. Biotechnol. 3, 135 (2015).

   Article  PubMed  PubMed Central  Google Scholar 

10. Freemont, P. S. Synthetic biology industry: data-driven design is creating new opportunities in biotechnology. Emerg. Top. Life Sci. 3, 651–657 (2019).

    Article  PubMed  PubMed Central  Google Scholar 

11. Pinglay, S. et al. Synthetic regulatory reconstitution reveals principles of mammalian Hox cluster regulation. Science 377, eabk2820 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

12. Rai, K. et al. Ultra-high-throughput mapping of genetic design space. Nature 650, 1035–1044 (2026).

    Article  CAS  PubMed  Google Scholar 

13. Liberante, F. G. & Ellis, T. From kilobases to megabases: design and delivery of large DNA constructs into mammalian genomes. Curr. Opin. Syst. Biol. 25, 1–10 (2021).

    Article  CAS  Google Scholar 

14. Park, S.-Y. et al. Driving towards digital biomanufacturing by CHO genome-scale models. Trends Biotechnol. 42, 1192–1203 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

15. Kotidis, P. et al. CHOGlycoNET: comprehensive glycosylation reaction network for CHO cells. Metab. Eng. 76, 87–96 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

16. Hefzi, H. et al. Multiplex genome editing eliminates lactate production without impacting growth rate in mammalian cells. Nat. Metab. 7, 212–227 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

17. Gutierrez, J. M. et al. Genome-scale reconstructions of the mammalian secretory pathway predict metabolic costs and limitations of protein secretion. Nat. Commun. 11, 68 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

18. Rai, K., Wang, Y., O’Connell, R. W., Patel, A. B. & Bashor, C. J. Using machine learning to enhance and accelerate synthetic biology. Curr. Opin. Biomed. Eng. 31, 100553 (2024).

    Article  CAS  Google Scholar 

19. Gam, J. J., DiAndreth, B., Jones, R. D., Huh, J. & Weiss, R. A ‘poly-transfection’ method for rapid, one-pot characterization and optimization of genetic systems. Nucleic Acids Res. 47, e106 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

20. Smanski, M. J. et al. Functional optimization of gene clusters by combinatorial design and assembly. Nat. Biotechnol. 32, 1241–1249 (2014).

    Article  CAS  PubMed  Google Scholar 

21. Shin, J., South, E. J. & Dunlop, M. J. Transcriptional tuning of mevalonate pathway enzymes to identify the impact on limonene production in Escherichia coli. ACS Omega 7, 18331–18338 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

22. Paddon, C. J. et al. High-level semi-synthetic production of the potent antimalarial artemisinin. Nature 496, 528–532 (2013).

    Article  CAS  PubMed  Google Scholar 

23. Cautereels, C. et al. Combinatorial optimization of gene expression through recombinase-mediated promoter and terminator shuffling in yeast. Nat. Commun. 15, 1112 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

24. Patel, Y. D. et al. Control of multigene expression stoichiometry in mammalian cells using synthetic promoters. ACS Synth. Biol. 10, 1155–1165 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

25. Avalos, J. L., Fink, G. R. & Stephanopoulos, G. Compartmentalization of metabolic pathways in yeast mitochondria improves the production of branched-chain alcohols. Nat. Biotechnol. 31, 335–341 (2013).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

26. Hill, A. J. et al. On the design of CRISPR-based single-cell molecular screens. Nat. Methods 15, 271–274 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

27. Zheng, Y. & VanDusen, N. J. Massively parallel reporter assays for high-throughput in vivo analysis of cis-regulatory elements. J. Cardiovasc. Dev. Dis. 10, 144 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

28. Payne, S. H. & Loomis, W. F. Retention and loss of amino acid biosynthetic pathways based on analysis of whole-genome sequences. Eukaryot. Cell 5, 272–276 (2006).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

29. Guedes, R. L. M. et al. Amino acids biosynthesis and nitrogen assimilation pathways: a great genomic deletion during eukaryotes evolution. BMC Genomics 12, S2 (2011).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

30. Eisenstein, E. et al. An expanded two-state model accounts for homotropic cooperativity in biosynthetic threonine deaminase from Escherichia coli. Biochemistry 34, 9403–9412 (1995).

    Article  CAS  PubMed  Google Scholar 

31. Slater, S. et al. Metabolic engineering of Arabidopsis and Brassica for poly(3-hydroxybutyrate-co-3-hydroxyvalerate) copolymer production. Nat. Biotechnol. 17, 1011–1016 (1999).

    Article  CAS  PubMed  Google Scholar 

32. Taillon, B. E., Little, R. & Lawther, R. P. Analysis of the functional domains of biosynthetic threonine deaminase by comparison of the amino acid sequences of three wild-type alleles to the amino acid sequence of biodegradative threonine deaminase. Gene 63, 245–252 (1988).

    Article  CAS  PubMed  Google Scholar 

33. Barak, Z., Chipman, D. M. & Gollop, N. Physiological implications of the specificity of acetohydroxy acid synthase isozymes of enteric bacteria. J. Bacteriol. 169, 3750–3756 (1987).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

34. Vinogradov, V. et al. Acetohydroxyacid synthase isozyme I from Escherichia coli has unique catalytic and regulatory properties. Biochim. Biophys. Acta 1760, 356–363 (2006).

    Article  CAS  PubMed  Google Scholar 

35. Dong, C. et al. Cloning and characterization of a panel of mitochondrial targeting sequences for compartmentalization engineering in Saccharomyces cerevisiae. Biotechnol. Bioeng. 118, 4269–4277 (2021).

    Article  CAS  PubMed  Google Scholar 

36. Curtin, J. A., Dane, A. P., Swanson, A., Alexander, I. E. & Ginn, S. L. Bidirectional promoter interference between two widely used internal heterologous promoters in a late-generation lentiviral construct. Gene Ther. 15, 384–390 (2008).

    Article  CAS  PubMed  Google Scholar 

37. Palmer, A. C., Egan, J. B. & Shearwin, K. E. Transcriptional interference by RNA polymerase pausing and dislodgement of transcription factors. Transcription 2, 9–14 (2011).

    Article  PubMed  PubMed Central  Google Scholar 

38. Uchida, N., Hanawa, H., Yamamoto, M. & Shimada, T. The chicken hypersensitivity site 4 core insulator blocks promoter interference in lentiviral vectors. Hum. Gene Ther. Methods 24, 117–124 (2013).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

39. Cooper, A. R., Lill, G. R., Gschweng, E. H. & Kohn, D. B. Rescue of splicing-mediated intron loss maximizes expression in lentiviral vectors containing the human ubiquitin C promoter. Nucleic Acids Res. 43, 682–690 (2015).

    Article  CAS  PubMed  Google Scholar 

40. O’Brien, J., Wilson, I., Orton, T. & Pognan, F. Investigation of the Alamar blue (resazurin) fluorescent dye for the assessment of mammalian cell cytotoxicity: resazurin as a cytotoxicity assay. Eur. J. Biochem. 267, 5421–5426 (2000).

    Article  PubMed  Google Scholar 

41. Vyazmensky, M. et al. Interactions between large and small subunits of different acetohydroxyacid synthase isozymes of Escherichia coli. Biochemistry 48, 8731–8737 (2009).

    Article  CAS  PubMed  Google Scholar 

42. Shi, C., Huo, X., You, R., Tao, Y. & Lin, B. High yield production of L-isoleucine through readjusting the ratio of two direct precursors in Escherichia coli. Bioresour. Technol. 418, 131889 (2025).

    Article  CAS  PubMed  Google Scholar 

43. Pinglay, S. et al. Mammalian genome writing: unlocking new length scales for genome engineering. Cell 189, 356–374 (2026).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

44. O’Gorman, S., Fox, D. T. & Wahl, G. M. Recombinase-mediated gene activation and site-specific integration in mammalian cells. Science 251, 1351–1355 (1991).

    Article  PubMed  Google Scholar 

45. Roger, A. J., Muñoz-Gómez, S. A. & Kamikawa, R. The origin and diversification of mitochondria. Curr. Biol. 27, R1177–R1192 (2017).

    Article  CAS  PubMed  Google Scholar 

46. Pan, D., Lindau, C., Lagies, S., Wiedemann, N. & Kammerer, B. Metabolic profiling of isolated mitochondria and cytoplasm reveals compartment-specific metabolic responses. Metabolomics 14, 59 (2018).

    Article  PubMed  PubMed Central  Google Scholar 

47. Matuszczyk, J.-C., Teleki, A., Pfizenmaier, J. & Takors, R. Compartment-specific metabolomics for CHO reveals that ATP pools in mitochondria are much lower than in cytosol. Biotechnol. J. 10, 1639–1650 (2015).

    Article  CAS  PubMed  Google Scholar 

48. Braymer, J. J. & Lill, R. Iron–sulfur cluster biogenesis and trafficking in mitochondria. J. Biol. Chem. 292, 12754–12763 (2017).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

49. Subramanian, V. S., Nabokina, S. M., Lin-Moshier, Y., Marchant, J. S. & Said, H. M. Mitochondrial uptake of thiamin pyrophosphate: physiological and cell biological aspects. PLoS ONE 8, e73503 (2013).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

50. Schulze-Siebert, D., Heineke, D., Scharf, H. & Schultz, G. Pyruvate-derived amino acids in spinach chloroplasts: synthesis and regulation during photosynthetic carbon metabolism. Plant Physiol. 76, 465–471 (1984).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

51. Zhang, J., Jensen, M. K. & Keasling, J. D. Development of biosensors and their application in metabolic engineering. Curr. Opin. Chem. Biol. 28, 1–8 (2015).

    Article  PubMed  Google Scholar 

52. Liu, S. et al. Iterative transcription factor screening enables rapid generation of microglia-like cells from human iPSC. Nat. Commun. 16, 5136 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

53. Johnstone, C. P. et al. Gene syntax defines supercoiling-mediated transcriptional feedback. Science 392, eadw1925 (2026).

    Article  CAS  PubMed  Google Scholar 

54. Mueller, D. & Heinzle, E. Stable isotope-assisted metabolomics to detect metabolic flux changes in mammalian cell cultures. Curr. Opin. Biotechnol. 24, 54–59 (2013).

    Article  CAS  PubMed  Google Scholar 

55. Riquelme-Guzmán, C., Stout, A. J., Kaplan, D. L. & Flack, J. E. Unlocking the potential of cultivated meat through cell line engineering. iScience 27, 110877 (2024).

    Article  PubMed  PubMed Central  Google Scholar 

56. Bader, J. E., Voss, K. & Rathmell, J. C. Targeting metabolism to improve the tumor microenvironment for cancer immunotherapy. Mol. Cell 78, 1019–1033 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

57. Arner, E. N. & Rathmell, J. C. Metabolic programming and immune suppression in the tumor microenvironment. Cancer Cell 41, 421–433 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

58. Aboelella, N. S., Brandle, C., Kim, T., Ding, Z.-C. & Zhou, G. Oxidative stress in the tumor microenvironment and its relevance to cancer immunotherapy. Cancers (Basel) 13, 986 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

59. Denk, D. & Greten, F. R. Inflammation: the incubator of the tumor microenvironment. Trends Cancer 8, 901–914 (2022).

    Article  CAS  PubMed  Google Scholar 

60. Umbarger, H. E. & Brown, B. Isoleucine and valine metabolism in Escherichia coli. J. Biol. Chem. 233, 415–420 (1958).

    Article  CAS  PubMed  Google Scholar 

61. Eisenstein, E. Cloning, expression, purification, and characterization of biosynthetic threonine deaminase from Escherichia coli. J. Biol. Chem. 266, 5801–5807 (1991).

    Article  CAS  PubMed  Google Scholar 

62. Quinlan, A. R. & Hall, I. M. BEDTools: a flexible suite of utilities for comparing genomic features. Bioinformatics 26, 841–842 (2010).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

63. Trolle, J. et al. Sequence Read Archive https://www.ncbi.nlm.nih.gov/bioproject/PRJNA1282762 (2026).

64. Trolle, J. et al. A shotgun approach for highly multiplexed mammalian metabolic engineering. Zenodo https://doi.org/10.5281/zenodo.21874813 (2026).

Download references

Acknowledgements

We thank M. Pacold for sharing Jurkat cells and members of the J.D.B. lab for helpful discussions. Figure schematics were made using BioRender.com and ChatGPT provided assistance in generating code for data analysis. We thank M. Maurano and his team for help with DNA sequencing.

Funding

This study was supported in part by the National Human Genome Research Institute, National Institutes of Health (NIH) grant RM1-HG009491 to J.D.B., NIH grant DP5OD036167 to S.P. and Chan Zuckerberg Initiative grant 2024-349901 to S.P. J.T. is supported by a SeaBridge Fellowship funded by the Washington Research Foundation, Brotman Baty Institute and the Chan Zuckerberg Initiative.

Author information

Authors and Affiliations

1. Institute for Systems Genetics, Department of Biochemistry and Molecular Pharmacology, NYU Langone Health, New York, NY, USA

   Julie Trolle, Sofia Sessa, Aleksandra Wudzinska, Mark Grivainis, David Fenyö, Sudarshan Pinglay & Jef D. Boeke

2. Department of Genome Sciences, University of Washington, Seattle, WA, USA

   Julie Trolle & Sudarshan Pinglay

3. Seattle Hub for Synthetic Biology, Seattle, WA, USA

   Julie Trolle & Sudarshan Pinglay

4. Graduate Program in Molecular Engineering, University of Washington, Seattle, WA, USA

   Katerina Rincones

5. Department of Biochemistry and Molecular Pharmacology, NYU Langone Health, New York, NY, USA

   Tori Rodrick & Drew R. Jones

6. Brotman Baty Institute for Precision Medicine, Seattle, WA, USA

   Sudarshan Pinglay

7. Department of Biomedical Engineering, NYU Tandon School of Engineering, Brooklyn, NY, USA

   Jef D. Boeke

Authors

1. Julie Trolle
2. Sofia Sessa
3. Aleksandra Wudzinska
4. Mark Grivainis
5. Katerina Rincones
6. Tori Rodrick
7. Drew R. Jones
8. David Fenyö
9. Sudarshan Pinglay
10. Jef D. Boeke

Contributions

J.T. and S.P. conceptualized the project. J.T., S.S., A.W., K.R. and T.R. performed the experiments. J.T. and M.G. built the computational pipelines and performed the data analysis. D.R.J. supervised the metabolomics experiments. D.F. supervised mathematical modeling. S.P. and J.D.B. supervised all other work. J.T., S.P. and J.D.B. wrote the manuscript with input from all authors. J.D.B. secured funding for the project.

Corresponding authors

Correspondence to Sudarshan Pinglay or Jef D. Boeke.

Ethics declarations

Competing interests

Two patent applications related to this work have been filed with J.T., S.P. and J.D.B. listed as inventors on one or both applications (US Patent application nos. 63/533,483 and 63/828,744). J.D.B. is a founder and director of CDI Labs, a founder of and consultant to Opentrons LabWorks/Neochromosome and a founder of JATech and serves or served on the scientific advisory boards of CZ Biohub New York, Logomix, Rome Therapeutics, SeaHub, Tessera Therapeutics and the Wyss Institute.

Peer review

Peer review information

Nature Biotechnology thanks Caleb Bashor and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data

Extended Data Fig. 1 Transcription unit distribution in CHO pre-infection (DNA Library) and post-infection (CHO-Div) of SGE library.

(a) Transcription unit distribution prior to lentiviral packaging compared to transcription unit distribution after infection of CHO cells. Y-axis indicates promoter/OLS context while the X-axis indicates CDS. Numbers are percentages; that is, all TU numbers add up to 100%. (b) qPCR to determine copy number of TU integrants relative to two single copy genomic targets (Dhfr, Hprt) for estimation of MOI. (c) Probability of finding a known valine biosynthesis pathway combination in CHO-Div based on a Monte Carlo simulated infection of 2.42 M cells with the library distribution measured for CHO-Div cells in (A). Depicted is the probability of finding ilvN, ilvB, ilvC and ilvD irrespective of subcellular compartmentalization (black), all 4 genes localized to the cytoplasm (red), or all 4 genes localized to mitochondria (green) at a range of MOIs. Panel a created in BioRender; Trolle, J. https://biorender.com/43izsew (2026).

Extended Data Fig. 2 Characterization of clones selected on low valine medium, including ¹³C tracing to confirm endogenous biosynthesis of valine in Clone CHO-Val-D1.

(a) Absolute cell count of clones CHO-Val-D1 and CHO-Val-D3 cultured in valine-replete RPMI compared to the parental cell line, previously engineered pMTIV cells, and SGE-Informed Design. Error bars represent standard deviation across triplicate wells. (b) Schematic detailing ¹³C labeling strategy to detect biosynthesized valine. Prototrophic cells were cultured on valine-free ¹³C labeled medium while control cells were cultured in a ¹³C valine-replete equivalent. (c) Extracted ion chromatography shows detection of ¹³C Valine in Clone CHO-Val-D1, which matches the retention time of spiked-in internal standard ¹³C₅/¹⁵N₁ Valine. Detection of ¹³C Valine in parental control cells is minimal by comparison. (d) MS1 peak corresponding to ¹³C₅ Valine is seen for Clone CHO-Val-D1 and not for the parental control cells. (e) MS2 confirms that fragmentation patterns for putative ¹²C₅ valine, ¹³C₅ valine and the ¹³C₅/¹⁵N₁ valine internal standard match expectation.

Extended Data Fig. 3 Genotypic characterization of clones CHO-Val-D1 and CHO-Val-D3.

(a) Log-transformed read counts were used to calculate a Z-score for each TU across all valine prototrophic clones. A Z-score threshold of 1.15 was used to define TU presence vs. TU absence across the 16 clones. (b) Percentage of amplicon-seq reads corresponding to each TU within clones CHO-Val-D1 and CHO-Val-D3. (c) Percentage of long reads corresponding to each TU in clones CHO-Val-D1 and CHO-Val-D3. (d) Concordance of TUs called by amplicon-seq and by long-read sequencing in clones CHO-Val-D1 and CHO-Val-D3. Amplicon-seq presence/absence calls utilize a Z-score threshold of 1.15 to separate signal from background, while no threshold was utilized for long-read sequencing. (e) Transgene quantification of SGE-introduced CDSs by RNA-seq in clones CHO-Val-D1 and CHO-Val-D3. Given near-identical CDS for ilvA, ilvA-L447F and ilvA-L481F, these could not be differentiated from each other in this analysis and reads mapping to any of these are presented collectively.

Extended Data Fig. 4 Characterization of clones selected on low isoleucine medium including ¹³C tracing to confirm endogenous biosynthesis of isoleucine in Clone CHO-Ile-10-H1.

(a) Absolute cell counts of clones CHO-Ile-10-H1 and CHO-Ile-10-H2 cultured on isoleucine-replete medium compared to the parental cell line. Error bars represent standard deviation across triplicate wells. (b) Schematic detailing ¹³C labeling strategy to detect biosynthesized isoleucine. Prototrophic cells were cultured on isoleucine-free ¹³C labeled medium while control cells were cultured in a ¹³C isoleucine-replete equivalent. (c) MS2 confirms that fragmentation patterns for putative ¹²C isoleucine, ¹³C₂ isoleucine, ¹³C₆ isoleucine and the ¹³C₆/¹⁵N₁ isoleucine internal standard match expectation.

Extended Data Fig. 5 Genotypic characterization of clones CHO-Ile-10-H1 and CHO-Ile-10-H2.

(a) Log-transformed read counts were used to calculate a Z-score for each TU across all isoleucine prototrophic clones. A Z-score threshold of 1.15 was used to define TU presence vs. TU absence across the 46 clones. (b) Percentage of amplicon-seq reads corresponding to each TU within clones CHO-Ile-10-H1 and CHO-Ile-10-H2. (c) Percentage of long reads corresponding to each TU in clones CHO-Ile-10-H1 and CHO-Ile-10-H2. (d) Concordance of TUs called by amplicon-seq and by long-read sequencing in clones CHO-Ile-10-H1 and CHO-Ile-10-H2. Amplicon-seq presence/absence calls utilize a Z-score threshold of 1.15 to separate signal from background, while no threshold was utilized for long-read sequencing. (e) Transgene quantification of SGE-introduced CDSs by RNA-seq in clones CHO-Ile-10-H1 and CHO-Ile-10-H2. Given near-identical CDS for ilvA, ilvA-L447F and ilvA-L481F, these could not be differentiated from each other in this analysis and reads mapping to any of these are presented collectively.

Extended Data Fig. 6 Sampling CHO-Val and CHO-Ile clones for double isoleucine/valine prototrophy.

(a) PrestoBlue amino acid dropout assay of 16 CHO-Val clones and 46 CHO-Ile clones cultured on isoleucine-free, valine-free RPMI. Metabolic activity was measured in triplicate following 3 days of growth on double isoleucine-free, valine-free medium. Error bars represent standard deviation across triplicate wells. (b) Percentage of double isoleucine/valine prototrophic clones that contain the indicated CDS. (c) CDS presence amongst double isoleucine/valine prototrophic CHO-Val clones and double isoleucine/valine prototrophic CHO-Ile clones. (d) Phenotype distribution amongst 16 CHO-Val clones and 46 CHO-Ile clones. (e) CDS presence by phenotype in CHO-Val (left) and CHO-Ile clones (right) as well as the expected distribution according to simulated data. Notable differences across phenotypes among CHO-Ile clones are highlighted in red boxes for ilvA, ilvA-L481F, and ilvG.

Extended Data Fig. 7 Transcription unit distribution in Jurkat pre-infection (DNA Library) and post-infection (Jurkat-Div) of SGE library.

(a) Transcription unit distribution prior to lentiviral packaging compared to transcription unit distribution after infection of Jurkat cells (b) qPCR of TU integrants relative to two double copy genomic targets (RPP30, ALB) for estimation of MOI. (c) Probability of finding a presumed valine biosynthesis pathway combination post-infection based on a simulated infection of 3 M Jurkat cells with the library distribution measured for Jurkat-Div cells in (A). Depicted is the probability of finding ilvN, ilvB, ilvC and ilvD irrespective of subcellular compartmentalization (black), all 4 genes localized to the cytoplasm (red), or all 4 genes localized to mitochondria (green) at a range of MOIs. (d) Probability of finding a presumed isoleucine biosynthesis pathway combination post-infection based on a simulated infection of 3 M Jurkat cells with the library distribution measured for Jurkat-Div cells in (A). Depicted is the probability of finding ilvM, ilvG, ilvB, ilvC, ilvD and ilvA-L481F irrespective of subcellular compartmentalization (black), all 6 genes localized to the cytoplasm (red), or all 6 genes localized to mitochondria (green) at a range of MOIs. Panel a created in BioRender; Trolle, J. https://biorender.com/43izsew (2026).

Extended Data Fig. 8 Long-term selection of Jurkat-Div cells led to identification of valine prototrophic Jurkat clones.

(a) Cell counts of Jurkat-Div and Jurkat-GFP control cells over 21 days of selection on a range of low valine RPMI conditions. (b) Growth curve of clones Jurkat-Val-31 and Jurkat-Val-32 cultured on valine-replete medium. Error bars represent standard deviation across triplicate wells. (c) Cell counts of Jurkat-Div and Jurkat-GFP control cells over 32+ days of selection on a range of low isoleucine RPMI conditions.

Extended Data Fig. 9 Characterization of Jurkat-Div clones selected on low valine medium.

(a) Log-transformed read counts were used to calculate a Z-score for each TU across all valine prototrophic clones. A Z-score threshold of 1.15 was used to define TU presence vs. TU absence across the 119 clones. (b) Percentage of amplicon-seq reads corresponding to each TU within clone Jurkat-Val-32. (c) Percentage of amplicon-seq reads corresponding to each TU within clone Jurkat-Val-31. (d) Percentage of long reads corresponding to each TU in clone Jurkat-Val-31. (e) Concordance of TUs called by amplicon-seq and by long-read sequencing in clone Jurkat-Val-31. Amplicon-seq presence/absence calls utilize a Z-score threshold of 1.15 to separate signal from background, while no threshold was utilized for long-read sequencing. (f) Frequency of barcode swapping in clone Jurkat-Val-31 as detected by long-read sequencing. (g) Transgene quantification of SGE-introduced CDSs by RNA-seq in clone Jurkat-Val-31. Given near-identical CDS for ilvA, ilvA-L447F and ilvA-L481F, these could not be differentiated from each other in this analysis and reads mapping to any of these are presented collectively.

Extended Data Fig. 10 Many clones arose from the same founder clones within the Jurkat-Val population.

(a) Composite barcode ‘fingerprint’ across strongly valine prototrophic clones. For each TU within a sample, the percentage of barcodes assigned to that TU was calculated. The median percentage representation of each TU across all clones is displayed as a composite heatmap. (b) Relative abundance of TUs, which comprise the valine prototrophic pathway ‘solution’ and their 9 underlying barcode combinations (3 per promoter/OLS, 3 per CDS) across all strongly valine prototrophic clones (Val Score > 0.75). (c) Distribution of clones that are part of sibling clonal groups and clones that are singletons. (d) Graph-based clustering identified 10 separate sibling groups in Jurkat clones selected on low valine medium ranging in size from 2–24 clones each. Simple linear regression indicates that sibling group size does not significantly correlate to an increase in average Val score (R² = 0.15; 95% CI for slope: −0.007 to 0.021; DFn = 1; DFd = 8; P-value = 0.2736). (e) Representative heatmaps for each of the 4 largest sibling groups: Groups 2, 4, 9, and 5. (f) Percentage of all valine prototrophic clones (Val Score > 0.23) and all strongly valine prototrophic clones (Val Score > 0.75) that contain each depicted CDS as measured by amplicon-seq after removal of duplicate sibling clones.

Supplementary information

Reporting Summary (download PDF )

Peer Review file (download PDF )

Supplementary Tables 1–10 (download XLSX )

Supplementary Table 1: Long reads mapped to SGE transgenes in Jurkat-Val-31. Supplementary Table 2: Long reads mapped to SGE transgenes in CHO-Val-D1, CHO-Val-D3, CHO-Ile-10-H1 and CHO-Ile-10-H2. Supplementary Table 3: SGE transgene integration site mapping in CHO-Val-D1, CHO-Val-D3, CHO-Ile-10-H1 and CHO-Ile-10-H2. Supplementary Table 4: Differentially expressed genes across all pairwise comparisons between CHO parental cells (valine-replete medium), CHO-Val-D1 (valine-replete medium) and CHO-Val-D1 (valine-free medium), as well as between CHO parental cells (valine-replete medium) and CHO-Val-D3 (valine-free medium). Supplementary Table 5: Differentially expressed genes across all pairwise comparisons between CHO parental cells (isoleucine-replete medium), CHO-Ile-10-H1 (isoleucine-replete medium) and CHO-Ile-10-H1 (isoleucine-free medium), as well as between CHO parental cells (isoleucine-replete medium) and CHO-Ile-10-H1 (isoleucine-free medium). Supplementary Table 6: SGE transgene integration site mapping in Jurkat-Val-31. Supplementary Table 7: Differentially expressed genes across all pairwise comparisons between Jurkat parental cells (valine-replete medium), Jurkat-Val-31 (valine-replete medium) and Jurkat-Val-31 (valine-free medium). Supplementary Table 8: Sequences for promoters, OLSs and CDSs used in the SGE library. Supplementary Table 9: Raw barcode read counts from amplicon-seq across all sequenced clones and controls. Supplementary Table 10: TU aggregated barcode counts from amplicon-seq across all sequenced clones and controls.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Cite this article

Trolle, J., Sessa, S., Wudzinska, A. et al. Highly multiplexed mammalian metabolic engineering with a shotgun approach. Nat Biotechnol (2026). https://doi.org/10.1038/s41587-026-03318-7

Download citation

• Received: 08 July 2025

• Accepted: 20 August 2026

• Published: 06 October 2026

• Version of record: 06 October 2026

• DOI: https://doi.org/10.1038/s41587-026-03318-7