Home blood collection for decentralized clinical studies is becoming practical: nurses, couriers, or robots visit patients, draw blood, and carry it to a lab. The blood is increasingly used as a stand-in for organs that cannot be sampled, such as the brain. In that setting the molecular readout decays in transit. Circulating cell-free RNA (cfRNA, RNA fragments floating in plasma) is broken down quickly by enzymes, and time-dependent cell damage in whole blood distorts the mix of cell types that downstream analysis tries to recover. The same fleets are going electric, and electric vehicles must stop to charge. Every minute a vehicle spends at a charger with tubes on board is a minute of decay.
This note proposes treating the amount of charge taken at each stop as a decision variable tied to the molecular value of the cargo. The objective is not only "arrive before a deadline." It is "arrive with the most usable assay signal." The two-sentence version is this. Routing work on partial charging already shows that taking only the energy you need saves time. We add that when and how much to charge should depend on how fragile the specimens currently on board are, with fragility measured by the downstream assay itself. A plausible consequence is that vehicles should fill their batteries while empty and take only minimal top-ups while carrying samples. We then test whether this beats the obvious fix of a per-sample hold-time deadline.
Core Claim
The core claim is conditional and falsifiable. If the loss of assay usefulness grows steadily with holding time, and grows at different rates for different assays, then a single maximum hold time is either too loose for fragile panels or wastefully tight for robust ones. A routing objective weighted by assay decay should then dominate hold-time constraints on the trade-off between fleet cost and specimen quality. The same logic extends to vehicles with refrigerated compartments. There, cold storage draws battery power in proportion to time spent loaded, so charging and specimen age become coupled a second time.
If decay turns out to behave like a cliff at a fixed time, the hold-time baseline will match our method. We will report that outcome as the main finding, not bury it.
What Exists and What It Leaves Open
The partial-charging side supplies the method. Hybrid Memetic Search for Electric Vehicle Routing With Time Windows, Simultaneous Pickup-Delivery, and Partial Recharges solves routing with time windows, simultaneous pickup and delivery, and partial recharges. It combines large-neighborhood search with a population-based memetic search, a station-insertion operator (PSSI), and cluster decomposition for 1,000-customer instances. Its pickups, however, are inert commodities with no perishability. Deep Reinforcement Learning for the Electric Vehicle Routing Problem With Time Windows trains an attention-based policy with feasibility masking. It assumes full recharges and names both that assumption and its short-sighted charging behavior as sources of detours. Fleet studies such as A Multi-Class Dispatching and Charging Scheme for Autonomous Electric Mobility On-Demand motivate our design. They show that matching charge level to the upcoming trip class cuts response times by over 20%. That work matches charge to the next job; we match it to the cargo already on board.
The transcriptomics side supplies the utility function and motivates the gap. Integrative single-cell and cell-free plasma RNA transcriptomics identifies biomarkers for early non-invasive AD screening finds a 34-gene plasma cfRNA panel for Alzheimer's disease that is enriched for microglia and oligodendrocyte markers. Its Random Forest reaches an area under the ROC curve (AUC) of 0.89. The authors name cfRNA instability and transit delay as primary barriers to clinical use. Cell-type deconvolution of bulk-blood RNA-seq reveals biological insights into neuropsychiatric disorders uses CIBERSORTx and bMIND to recover cell-type-specific signal from low-depth whole-blood RNA-seq. Its estimates agree with single-cell references at R² between 0.61 and 0.84, and it finds 1,211 eGenes (genes whose expression is linked to a genetic variant) that bulk analysis misses. That signal depends on accurate cell proportions, which is exactly what ex vivo lysis corrupts.
The logistics-adjacent work frames the need without modeling vehicles or energy. Intelligent blood collection robot: key technologies, clinical applications, and future challenges and Digitizing clinical trials describe the collection setting. Surrogate DC Microgrid Models for Optimization of Charging Electric Vehicles under Partial Observability shares only the word "surrogate": it optimizes grid-side charging schedules, not routes or cargo. The research file found no paper that joins these threads.
Research Program
The work has three parts.
The first part measures decay curves, because nothing in the file gives them. We will run an ex vivo holding study. Paired blood draws are held for a grid of durations at ambient and 4°C, then processed for plasma cfRNA-seq and whole-blood RNA-seq. For each hold condition we compute three assay-level retention scores:
- the rank correlation of the 34-gene panel score with the zero-hold sample;
- the drift of CIBERSORTx cell proportions from zero-hold;
- the retention of cell-type-specific expression estimates from bMIND.
Fitting these gives a retention function $U_a(h, \tau)$. It is the fraction of assay $a$'s usable signal remaining after hold time $h$ at temperature $\tau$, with value 1 at zero hold. If the wet-lab study is delayed, we will run the optimization side under a family of parametric curves (exponential, sigmoidal, step) as a sensitivity analysis. We will state plainly that those results are conditional on the assumed curve.
The second part is the optimizer. Let $x_{ij}$ indicate that arc $(i,j)$ is traveled at cost $c_{ij}$. Let $h_s$ be the time from draw to lab for specimen $s$, and let $w_s$ weight that specimen's scientific value. The objective is
$$\min \; \sum_{i,j} c_{ij} x_{ij} \;+\; \lambda \sum_{s} w_s \big(1 - U_{a(s)}(h_s, \tau_s)\big),$$
where $\lambda$ sets the price of lost signal and $a(s)$ is the assay assigned to specimen $s$. Battery use on an arc is traction energy plus refrigeration power $p$ times travel-and-dwell time whenever the vehicle carries specimens. Charge added at a station takes time given by a charging curve that flattens near full.
We embed this objective in the hybrid memetic search. For a fixed route and station sequence, choosing charge amounts reduces to a small continuous problem that is convex whenever $U$ is concave in $h$. We solve it exactly inside route evaluation. The insertion operators are unchanged apart from their scoring.
The third part tests the load-aware charging hypothesis directly. We inspect optimized routes to see whether large charges concentrate on empty legs, and we compare against a hand-coded rule that does exactly that.
Baselines
The naive reference point is the current practice for a courier: full recharge at every charging stop, routes built for distance only, and specimen quality reported after the fact. Every other method has to beat this to justify its complexity.
The strongest existing competitor is the hybrid memetic search from Hybrid Memetic Search for Electric Vehicle Routing With Time Windows, Simultaneous Pickup-Delivery, and Partial Recharges, with partial recharges and an added per-specimen maximum hold time. It already handles partial charging and pickups at scale. With a deadline it is exactly what a practitioner would deploy. We will not fix the deadline at one value. We will sweep it and trace its full curve of fleet cost against expected retention. Both methods use the same instances, the same charging curve, the same refrigeration load, and the same wall-clock budget per instance. Beating it means our curve lies strictly above its curve by a margin larger than seed-to-seed variation, on at least the jd instances. The file has no reported cost figures for this method on akb or jd ("no reported result found on Lacuna" for exact distances). We will therefore rerun it ourselves.
The learned full-recharge policy from Deep Reinforcement Learning for the Electric Vehicle Routing Problem With Time Windows is a fast alternative. It reports an optimality gap of 8.5% to 11.8% against CPLEX on small Schneider instances and solves a 100-customer instance in about 6.7 minutes. We include it, retrained on our instances with the same deadline sweep, to measure what full recharging costs once cargo decays.
Two ablations isolate the key ingredients. One uses our objective but forces full charges, which isolates the value of continuous charge amounts. The other uses our objective but a single average retention curve for all specimens, which isolates the value of assay-specific decay. A third comparison is the "charge big when empty, top up when loaded" rule applied on top of the deadline baseline.
Why not just…?
Why not just impose a maximum hold time in an existing solver? This is the real threat. If measured retention is flat and then drops sharply, a well-chosen deadline captures nearly all the value. The deadline sweep in the baselines settles it: we report the gap between the two curves at matched fleet cost. The file offers only indirect support for expecting a gap. Fleet studies such as Optimal Vehicle Dimensioning for Multi-Class Autonomous Electric Mobility On-Demand Systems find that rigid service requirements inflate the fleet size needed. We treat that as a hypothesis, not evidence.
Why not just use a load-aware charging rule of thumb? It may recover most of the gain, which is why it is a named comparison. If it matches our method, the practical recommendation is the rule, and the optimizer's role shrinks to showing when the rule fails. Candidates include mixed assays, tight time windows, and refrigerated vehicles where idle time also costs energy.
Why not just add stabilizing collection tubes and stop worrying about transit? Chemical stabilization reduces decay but may not remove it. The holding study will include stabilized and unstabilized arms, so we can measure how much routing still matters after stabilization. If it matters little, the proposal narrows to unstabilized assays, such as the cfRNA pipeline in Integrative single-cell and cell-free plasma RNA transcriptomics identifies biomarkers for early non-invasive AD screening, and to settings where stabilizers are impractical.
Proposed Experiments and Decisive Checks
We will extend the akb and real-geography jd benchmark sets (up to 1,000 customers and 100 stations) and the Schneider instances. Each customer becomes a patient whose specimen carries an assay label and a retention curve from part one. Vehicles gain a refrigeration draw.
The primary metrics are fleet cost and value-weighted expected retention. We also carry retention through to the downstream assay. Measured degradation levels are mapped onto held-out portions of the 337-sample cfRNA cohort and the 1,730-person whole-blood cohort, and we ask how much the panel's AUC (0.89 when reported) and the deconvolution R² (0.61–0.84) fall under each routing policy's distribution of hold times.
The decisive check has three outcomes. The proposal succeeds if, at equal fleet cost, the assay-weighted optimizer keeps measurably more deconvolution accuracy or classifier AUC than the best deadline in the sweep. It narrows to a practical rule if the load-aware rule of thumb closes most of that gap. It fails honestly if the measured decay curves are step-like and the deadline curve coincides with ours.
References
- Hybrid Memetic Search for Electric Vehicle Routing With Time Windows, Simultaneous Pickup-Delivery, and Partial Recharges
- Deep Reinforcement Learning for the Electric Vehicle Routing Problem With Time Windows
- A Multi-Class Dispatching and Charging Scheme for Autonomous Electric Mobility On-Demand
- Integrative single-cell and cell-free plasma RNA transcriptomics identifies biomarkers for early non-invasive AD screening
- Cell-type deconvolution of bulk-blood RNA-seq reveals biological insights into neuropsychiatric disorders
- Intelligent blood collection robot: key technologies, clinical applications, and future challenges
- Digitizing clinical trials
- Surrogate DC Microgrid Models for Optimization of Charging Electric Vehicles under Partial Observability
- Optimal Vehicle Dimensioning for Multi-Class Autonomous Electric Mobility On-Demand Systems