DNA Assembly & Plasmid Manufacturing

The workflow

Plasmid manufacturing is a multi-stage experimental pipeline. Each stage produces an intermediate that feeds the next: purified DNA goes into assembly reactions, assembly products are transformed into cells, colonies are screened, and verified clones are scaled up. On a modern biofoundry platform, most of this runs on automation hardware. The liquid handling, thermal cycling, colony picking, and plate routing happen with minimal hands-on time. But the decisions about what to assemble, under which conditions, and which clones to advance are still made manually between stages.

In a typical Golden Gate (MoClo) pipeline at a biofoundry, source plasmids are extracted via automated alkaline lysis and bead-based purification into 96-well plates. An acoustic liquid handler dispenses DNA parts at nanoliter precision, followed by thermal cycling with digest/ligation cycles scaled to part count. A single plate holds 94 assemblies plus two controls. Assembly products are transformed into competent cells and plated on selective media. Colonies are picked, verified by colony PCR, and positive clones are cherry-picked for culture scaleup and DNA quantification. Sequence verification runs either on-site via nanopore or through external sequencing services.

The automation typically runs on a three-layer stack. At the bottom, firmware controls individual instruments: liquid handlers, acoustic dispensers, thermal cyclers, plate readers. A scheduling layer coordinates plate routing, device reservations, and timing across instruments. At the top, an orchestration layer manages experiments, protocols, samples, and worklists in a database. A single campaign generates over a million logged variable values, tens of thousands of device operations, and hundreds of thousands of system messages.

In practice, the heavy manual workload sits between stages. When an assembly fails, someone has to dig through logs and plate data to understand why, then decide whether to adjust part ratios, redesign overhangs, change the thermal cycling protocol, or retry with a fresh enzyme lot. When yields are low, someone scripts a new worklist with adjusted parameters. When results are ambiguous, teams often accept suboptimal outcomes rather than spend another cycle troubleshooting. None of this analysis carries over to the next campaign in a structured way.

ProviGen closes this gap. The platform sits at the orchestration layer and reads every data source the automation stack produces: dispense volumes, device logs, colony counts, PCR results, sequencing outcomes, and execution metadata. After each stage, it updates a model of what works, what fails, and why. Before the next batch, it recommends which reagent conditions, part ratios, and cycling parameters to use for a given set of constructs. Those recommendations are written back as structured worklists that the automation executes directly.

Decision points at each stage

At every stage of this cycle, teams make decisions that affect downstream success. Currently those decisions are based on fixed defaults, SOPs, or operator intuition. None of them systematically learn from previous campaigns.

1. Part molar ratios. The backbone-to-insert ratio is typically fixed at 1:2, but the optimal ratio depends on part count, overhang set, and fragment lengths. The model recommends ratios per construct and learns from colony counts, PCR positive rates, and sequencing results.
2. Cycle count. Common defaults scale cycle count by part number (e.g. 30 cycles for simple assemblies, up to 60 for complex ones). The model refines this based on overhang fidelity data and observed assembly efficiency at the specific platform.
3. Process parameters. Reaction volume, enzyme concentration, ligation temperature, and inactivation time all affect assembly yield, quality, and run-to-run variance. Small changes compound across stages. The model learns which parameter combinations produce reliable outcomes for a given assembly type and construct.
4. Colony screening. After picking, the model flags constructs where the positive rate is unexpectedly low compared to similar assemblies. This helps teams catch assembly or transformation problems early rather than discovering them at sequence verification.
5. Failure response. When an assembly fails, the question is whether to retry with modified conditions or redesign the parts. The model classifies failure modes: no colonies points to a transformation problem, all empty vector to an assembly problem, mixed bands to partial assembly.
6. Cherry-pick decisions. Which clones to advance based on partial QC data. The model draws on colony PCR band quality, growth rate, and prior nanopore results for similar constructs.

Beyond the assembly reaction itself, even standard steps have tunable parameters that the model can learn to understand and predict: electroporation voltage (1500 to 2500 V, strain-dependent), pulse duration, recovery volume and time, recovery temperature. And there are batch-level variables that drift silently: competent cell lot and time-since-thaw, enzyme lot activity, SOC media batch, antibiotic plate freshness.

Learning over time

Each assembly run generates data at every stage: dispense volumes, device logs, colony counts, PCR results, and eventually full sequence verification. These observations arrive at different times. Dispense confirmations are immediate. Colony counts follow after transformation. PCR results come after screening. Sequencing arrives last. The model updates as each signal comes in, so a colony count already informs the next batch design before sequencing confirms which clones are correct.

Published overhang fidelity data^[1] covering all 256 possible 4-nucleotide pairs provides a strong prior on assembly success before any wet lab work begins. As real results accumulate, the model refines these priors for the specific platform, part libraries, and conditions in use.

Batch-level drift is another signal the model tracks. Competent cell quality, enzyme lot activity, and plate reader calibration change gradually. This information typically disappears into troubleshooting notes or operator memory. Connected to the model, it becomes reusable process intelligence that carries over across campaigns.

Multi-objective reality

Plasmid manufacturing involves multiple objectives that trade off against each other. Yield, purity (supercoiled fraction, absence of genomic DNA), plasmid quality, impurity profile (endotoxin, host-cell protein, RNA), robustness across operators and batches, turnaround time, and cost per batch all matter. Higher yield conditions may compromise supercoiling fraction. Faster turnaround may mean less stringent QC. Optimizing one metric while ignoring the rest leads to processes that pass one test but fail in production.

The model handles these tradeoffs explicitly by mapping the Pareto frontier (a set of conditions where no single objective can be improved without compromising another). Teams define what matters and how much, and the model proposes experiments that move toward that frontier rather than chasing a single metric.

MoClo at scale

Modular Cloning (MoClo) uses standardized Type IIS restriction enzymes (BsaI, BsmBI) to assemble multiple DNA parts in a single reaction. Parts are designed with compatible 4-nucleotide overhangs that dictate assembly order. Related standards follow the same modular logic: Golden Braid extends MoClo with alternating enzyme pairs for unlimited hierarchical levels; Gibson Assembly joins overlapping fragments without restriction enzymes; BASIC Assembly uses linkers for combinatorial pathway construction. The decision loop applies to all of these. The parameters differ but the optimization structure is the same.

The design space grows combinatorially. Even a moderate parts library creates thousands of relevant combinations, and each new construct may behave differently during assembly depending on its specific overhangs, fragment lengths, and part interactions. It is impossible to manually test each beforehand, which is where a predictive model becomes highly valuable: it generalizes from past assemblies to predict conditions for new constructs that have never been built before. Multi-level assembly (Level 0 parts into Level 1 transcription units into Level M multigene constructs) compounds this further across tiers.

Fragment count is the primary driver of difficulty. Assemblies with 2 to 3 parts typically succeed at high rates. Above 6 parts, success drops significantly. Assemblies with up to 24 fragments are at the hard frontier. The main challenge is overhang fidelity: some 4-nucleotide overhang pairs have near-zero ligation efficiency, and with more fragments the probability of hitting a problematic pair increases multiplicatively. Additional failure modes include dam-methylation sensitivity (E. coli-propagated parts with dam sites at the BsaI recognition sequence silently fail) and compounding failure probability across multi-level assemblies.

Standard pain points in this space are clear targets for the decision loop. There is typically no systematic record of which parameter combinations work for which part families. Colony PCR gives a binary result (correct band or not) that does not distinguish partial assemblies from contamination. Batch-level drift in competent cell quality, enzyme lots, and media goes untracked. Platform-level anomalies (volume drift from stale acoustic surveys, equipment reservation conflicts, retry loops) are caught manually rather than flagged from execution metadata.

---

The precedent for model-guided experimental planning is growing. Merck KGaA developed BayBE,^[2] a Bayesian experimental planning framework that powered dozens of internal R&D use cases before being open-sourced. Ginkgo Bioworks integrated Zymergen's machine learning and automation stack specifically to explore genetic design space more efficiently across their foundry operations.^[3]

ProviGen helps teams make better decisions with fewer experimental runs. Every batch produces results, updates the model, and improves the next batch. Adaptive parameters per construct replace fixed defaults. Partial results from early stages inform the next round before full verification is complete. Failed assemblies teach the model what to avoid. Batch drift is caught from execution metadata rather than discovered through unexplained failures.

If you are interested in exploring what a decision loop could look like for your assembly workflow, reach out to contact@provigen.ai.