What does it cost to screen LNP formulations at the bench?
Before building a case for computational screening, it is worth being precise about what the alternative costs. Bench-based LNP formulation screening involves synthesis, characterization, and functional testing of each candidate formulation. The unit economics break down as follows at typical CDMO or academic facility rates.
Synthesis (microfluidic ethanol injection). A single LNP batch synthesized via microfluidic mixing using a benchtop NanoAssemblr or equivalent instrument requires approximately 1–3 hours of technician time, solvent and lipid material costs (ionizable lipid at $30–300/mg at research scale depending on custom synthesis requirements, helper lipids at $5–30/mg), and instrument time. A conservative all-in cost per synthesized batch is $200–600 for routine formulations with commercially available ionizable lipids; custom ionizable lipid batches can increase material cost substantially.
Characterization (DLS/ELS). Each batch requires size, PDI, and zeta potential measurement by dynamic light scattering and electrophoretic light scattering. Run time is approximately 15 minutes per sample on a Zetasizer Nano or equivalent. Total characterization cost (instrument amortization, technician time) is approximately $50–100 per sample.
Encapsulation efficiency (Ribogreen assay). Ribogreen fluorescence assay (Thermo Fisher Quant-iT RiboGreen) quantifies free versus encapsulated RNA. Each 96-well plate run requires approximately 2 hours setup and analysis time, reagent costs of $5–15/sample. Total cost with technician time: approximately $40–80 per sample.
In vitro functional testing (HepG2 or primary hepatocyte transfection, luciferase reporter). Cell culture, transfection, and luciferase readout take approximately 48–72 hours from LNP dose to result. Reagent costs (cell culture media, transfection reagents, One-Glo or Nano-Glo luciferase substrate) plus cell culture facility overhead: $100–300 per sample tested. If primary human hepatocytes are used instead of HepG2, reagent and procurement costs rise sharply (primary hepatocytes from commercial suppliers cost $500–2,000 per batch).
Summing across synthesis, characterization, EE measurement, and in vitro functional testing: the all-in cost per LNP formulation screened at bench scale is approximately $400–1,000 per candidate in a modest program, and $800–2,000+ per candidate if primary hepatocytes are used for functional validation.
The combinatorial scale problem
The LNP formulation space is defined by four primary variables: ionizable lipid identity, helper lipid identity and ratio, PEG-lipid identity and molar fraction, and cholesterol fraction. Each variable has meaningful diversity:
- Ionizable lipids: dozens of candidates in the published literature, plus proprietary variants. A conservative exploration set is 50–200 ionizable lipid candidates.
- Helper lipid combinations: DSPC, DOPE, DPPC, and their ratios relative to ionizable lipid (typically expressed as molar ratios with ionizable lipid at 40–50 mol%). Conservative exploration: 20–40 combinations.
- PEG-lipid variants: PEG2000-DMG, PEG2000-DSG, varying molar fractions (0.5–3 mol%). Approximately 10–20 variants with molar fraction variation.
- Cholesterol fraction: typically 30–40 mol%, but the precise fraction interacts with ionizable lipid head group geometry to modulate bilayer fluidity and pKa. 5–10 values to explore.
A combinatorial screening space of 100 ionizable lipids × 30 helper lipid combinations × 15 PEG-lipid variants × 8 cholesterol fractions = 360,000 candidate formulations. Even a restricted sub-library of 10,000 candidates (which is what the Gendelivr platform evaluates in silico) vastly exceeds what any bench program can physically screen in a development timeline.
At $500/candidate (mid-range bench cost), screening 10,000 formulations to bench scale would cost $5 million and require approximately 25 scientist-years of effort. At 10 formulations per week per scientist (a realistic throughput including characterization), it would take 19 years.
What computational screening actually does — and what it doesn't
The economic case for computational pre-screening depends on an honest accounting of what the simulation predicts accurately and where it has real uncertainty. Presenting computational screening as equivalent to bench data would be technically dishonest and scientifically counter-productive — bench validation is still required.
What QSPR and CGMD accurately predict (within training distribution):
- Ionizable lipid apparent pKa, based on headgroup molecular descriptors trained on measured pKa data from published and proprietary libraries. Prediction accuracy for known ionizable lipid chemistries: mean absolute error approximately 0.2–0.3 pKa units.
- Particle size and polydispersity as a function of lipid composition and N/P ratio, using CGMD models that capture bilayer packing geometry.
- Encapsulation efficiency as a function of ionizable lipid surface charge density at synthesis pH — a parameter that correlates strongly with RNA electrostatic binding during microfluidic assembly.
- ApoE adsorption probability as a function of surface PEG density, particle charge, and particle size — using protein adsorption models trained on corona proteomics data from LNPs of known composition.
What computational screening does not predict:
- In vivo editing efficiency — which depends on Cas9 protein activity, guide RNA on-target efficiency, chromatin state at the target locus, and post-endosomal Cas9 nuclear localization. These are not LNP formulation parameters.
- Off-target editing rates — which depend entirely on guide RNA design, not formulation.
- In vivo biodistribution at tissue-level granularity beyond ApoE-LDLR mediated hepatic uptake.
- Long-term in vivo safety signals.
The computational layer is a pre-filter, not a replacement for in vitro or in vivo validation. The economic argument is about the cost of what you don't make: if computational screening eliminates 9,800 of 10,000 candidates before bench synthesis, the bench team synthesizes and tests 200 candidates instead of 10,000. The savings are in the formulations not made, not in the elimination of validation.
The hit rate comparison
What constitutes a "hit" in LNP screening? The typical working definition in hepatic delivery programs is a formulation achieving:
- Particle size between 80–160 nm (compatible with hepatic sinusoidal fenestrae of 100–180 nm diameter)
- PDI < 0.15 (acceptable homogeneity)
- EE > 85% (efficient RNA encapsulation)
- In vitro transfection efficiency in HepG2 cells > 50% of positive control performance (using a benchmark LNP of known composition)
In fully random screening programs (selecting candidates without computational guidance), hit rates for satisfying all four criteria simultaneously typically fall in the range of 2–8%. In a random screen of 1,000 candidates, 20–80 hits are expected.
In computationally pre-filtered screens — where the in silico ranking has already filtered for predicted pKa 6.2–6.8, predicted EE > 80%, and predicted size 80–160 nm — the bench hit rate for satisfying all four criteria rises substantially. In Gendelivr's experience, computationally pre-filtered candidate sets achieve bench hit rates of approximately 35–55%, meaning that a bench campaign of 200 pre-filtered candidates yields 70–110 formulations meeting all criteria.
To achieve the same 70–110 hits from random screening would require synthesizing and testing approximately 1,400–3,500 candidates. The cost difference between testing 200 pre-filtered candidates versus 1,400–3,500 random candidates represents the direct economic value of the computational pre-filter.
The economic comparison in numbers
| Metric | Unguided bench screen | Computationally guided screen |
|---|---|---|
| Candidates evaluated in silico | 0 | 10,000 |
| Candidates sent to bench | 2,000 | 200 |
| Expected hits meeting all criteria | 100 (5% hit rate) | 90 (~45% hit rate) |
| Bench cost per candidate (all-in) | $600 | $600 |
| Total bench cost | $1,200,000 | $120,000 |
| Timeline to equivalent hit count | 18–24 months | 4–6 months |
The numbers above use conservative assumptions. The hit rate improvement from computational pre-filtering is likely conservative; the bench cost per candidate could be higher if primary hepatocytes are used for functional validation; and the timeline compression is often more dramatic in programs that were previously constrained by available scientist headcount rather than formulation quality.
Where the platform investment goes
The Gendelivr platform's core investment is in the training data and model architecture for pKa prediction, encapsulation efficiency estimation, and protein corona composition modeling. These models are trained on proprietary and published LNP characterization datasets spanning thousands of formulation–outcome pairs — which is what allows the in silico screening to operate in the relevant region of formulation space rather than treating the full 360,000-candidate space as uniformly uncertain.
The platform does not generate simulated bench data to present to customers as real experimental results. The output of the in silico screen is a ranked list of candidates with predicted property ranges and confidence intervals, which the bench team uses to prioritize synthesis. The bench data is always generated by the customer's team or a CRO — not imputed from the model.
The economic argument for computational pre-screening is not primarily about replacing bench science. It is about letting bench scientists spend their time on the formulations most likely to work, rather than on random sampling of a space too large to exhaustively explore by hand.