Macrophages are immune cells that switch between two broad states: an inflammatory one (called M1) that clears debris and fights infection, and a healing one (M2) that rebuilds tissue. In chronic diabetic ulcers, spinal cord injury, and non-healing fractures, this switch stalls — cells stay locked in the inflammatory state and the tissue decays. A growing clinical literature argues the failure is fundamentally one of timing: the same inflammatory signals that jump-start repair become destructive when they persist, and today's macro-scale imaging only reveals the failure weeks after the cellular events have passed. The stated goal in that literature is a small implanted device that continuously measures local inflammatory markers and intervenes at exactly the right moment. The sensing technology already exists in the lab — plasmonic biosensors that detect inflammatory cytokines such as TNF-$\alpha$ down to $10^{-15}$ mol/L — but the clinical reviews stop at the concept. What they never specify is the chip that reads such a sensor inside a millimeter-scale, battery-powered implant: the analog front-end, the chain of amplifiers and filters that turns femtoamp-to-picoamp sensor currents into digital numbers, designed under nanowatt power budgets. Today that chip is drawn by hand over months by expert engineers, precisely the bottleneck that machine-learning-driven analog design automation was built to remove.
That automation has made real progress — transformer and diffusion generators now invent valid circuit topologies, and LLM agents size transistors to meet specifications — but every benchmark in the field lives in the telecom regime: gigahertz bandwidths, milliwatt power, transistors driven hard in strong inversion. Implant-grade biosensor front-ends invert each of those pressures simultaneously: signal currents near a picoamp or below, bandwidths set by biology (cytokine dynamics evolve over hours, so well under 1 Hz), nanowatt power budgets, and transistors operated in weak inversion, where a small geometry change moves the operating point exponentially. This note proposes to close that gap: build the translation layer from macrophage-modulation monitoring requirements to analog front-end specifications, a public benchmark of implant-grade front-end circuits in that inverted regime, an agent extension that can actually size them, and a closed-loop co-simulation that scores the whole system — did the intervention land inside the therapeutic window — rather than a single block's spec sheet.
Core Claim
The claim is specific and falsifiable. Automated analog sizing agents, as they exist, will fail on implant-grade biosensor front-end specifications, because those specifications are far outside the distribution the agents and their benchmarks were built for; and a modest extension — a biology-to-electronics specification compiler, weak-inversion design knowledge in the agent's toolset, and a parasitic-closure loop — will close most of that gap. The motivation comes from three directions in the literature. First, regime shift is a measured failure mode, not a hypothesis: AICircuit: A Multi-Level Dataset and Benchmark for AI-Driven Analog Integrated Circuit Design showed that models trained on isolated circuit blocks completely fail to predict the same blocks inside a full 28 GHz receiver, because inter-stage loading changes the problem; moving from telecom to sub-threshold biosensing specs is a comparable shift. Second, the weak-inversion regime is exactly where LLM agents are known to underuse domain knowledge: AutoCircuit: Agentic Pareto Front Exploration for Analog Circuit Designs found that giving an agent all available tools degraded performance relative to a curated subset, and that plain optimizers beat the agent on a small 12-parameter amplifier. Third, at picoamp signal levels, post-layout parasitic leakage and capacitance are the same order of magnitude as the signal itself, so single-pass sizing that ignores the physical layout is not merely suboptimal, it is invalid — OSIRIS: Bridging Analog Circuit Design and Machine Learning with Scalable Dataset Generation already showed that single-pass place-and-route tools fail against parasitic-aware closed loops on ordinary blocks, before the signal scale shrinks by six orders of magnitude.
One scoping decision follows from the literature: topology discovery is not where this problem lives. Generative engines such as AnalogGenie: A Generative Engine for Automatic Discovery of Analog Circuit Topologies, which produces simulatable netlists at 93.2% correctness and scales to 56–64 devices, have largely solved structure discovery, and biosensor front-end topologies (transimpedance amplifiers, potentiostats, chopper-stabilized sensor amplifiers) are textbook-stable anyway. The open hard problem is sizing in weak inversion plus physical parasitic closure under inverted specifications — which is where this proposal concentrates.
Research Program
The program has four connected pieces, and the application side is treated as the source of specifications and the evaluation setting, not as a claim about curing wounds.
The first piece is a specification compiler that turns biological monitoring requirements into electrical ones, so the circuit work is anchored to real numbers rather than invented ones. The logic reduces to one inequality the compiler must satisfy per block:
$$i_{n,\mathrm{rms}} \;=\; \sqrt{S_i \, B} \;<\; R_{\mathrm{sens}}\, C_{\min}$$
where $S_i$ is the front-end's input-referred current-noise spectral density in A²/Hz, $B$ is the signal bandwidth (fixed by the hours-scale kinetics of the M1-to-M2 transition, hence far below 1 Hz), $R_{\mathrm{sens}}$ is the sensor's responsivity in amperes per mol/L, and $C_{\min}$ is the concentration resolution the biology demands. The plasmonic detection limit of $10^{-15}$ mol/L reported in Bioaugmented osteosynthesis: precise monitoring and intervention of the bone healing microenvironment sets the sensing ambition; the biomarker panel comes from Adipose-Derived Mesenchymal Stem Cells Accelerate Diabetic Foot Ulcer Healing by Promoting Macrophage M2 Polarization Through Downregulation of EREG and CSTA, which identified the EREG and CSTA hub genes from human diabetic-ulcer expression data with an externally validated model (area under the ROC curve of 0.944 on cohort GSE143735). The compiler also carries the power budget, derived from the implant scale argued for in the feasibility study of A feasibility study on AI-controlled closed-loop electrical stimulation implants.
The second piece is a benchmark, which we call Bio-AFE-Bench, built on the SkyWater 130nm open PDK so it composes directly with the field's existing infrastructure. It contains roughly eight front-end blocks: a transimpedance amplifier for plasmonic photodiode readout, a potentiostat front-end for an electrochemical cytokine immunosensor, a chopper-stabilized amplifier for pH sensing, an impedance-measurement front-end, weak-inversion bias and reference circuits, and a low-pass filter — included deliberately, since OSIRIS documented that low-pass filters are where existing tools crashed while its parasitic-aware agent succeeded. Blocks are split into Easy, Medium, and Hard tiers following the AMS-SizingBench convention, and every circuit is scored twice: pre-layout spec attainment in Ngspice, and post-layout attainment after parasitic extraction, with design-rule-check and layout-versus-schematic validity, mirroring OSIRIS's evaluation.
The third piece is the agent extension. We start from the AutoSizer architecture — a manager LLM that classifies devices into critical and secondary groups, orchestrates numerical optimizers, and reflects its way out of local minima — and add two ingredients with direct precedent: a curated weak-inversion toolset built around the $g_m/I_d$ methodology, curated because AutoCircuit demonstrated that an over-full toolbox distracts the agent; and an OSIRIS-style two-level parasitic-closure loop in which sizing decisions are re-evaluated against extracted post-layout parasitics rather than idealized schematics.
The fourth piece is the closed-loop co-simulation that provides the system-level score. A sensor model produces noisy biomarker currents; the automatically sized front-end (with extracted parasitics) digitizes them; an on-implant state estimator — a Random Forest, chosen because the feasibility study measured Random Forest at a mean absolute error of $0.03 \pm 0.01$ within-subject while running at roughly 45% of a perceptron's runtime — infers the inflammatory state; an actuation model shifts the polarization dynamics, calibrated to the in vitro shift reported by the EREG/CSTA paper (a 14% reduction in M1 polarization alongside a 181% increase in M2 polarization under co-culture) and to the near-complete wound closure by Day 21 in $db/db$ diabetic mice from Optogenetic engineered umbilical cord MSC-derived exosomes for remodeling of the immune microenvironment in diabetic wounds and the promotion of tissue repair. The macrophage population model enforces the biphasic window documented in the osteosynthesis review: early inflammatory signaling is required for repair to start, sustained signaling causes tissue destruction. We are deliberately honest about the status of this loop: the biological dynamics are calibrated from published endpoints, not fitted time-series data, so the loop is a stress test of circuit sensitivity and control timing, not a therapeutic prediction.
Baselines
The naive reference point is random search over the sizing space, which is the same naive reference OSIRIS used when it reported cutting the time to find high-quality layouts by nearly 50% relative to random search; on top of that we add a "do almost nothing" baseline in which a textbook front-end design is ported to SkyWater 130nm and evaluated post-layout, to check whether trivial designs already clear the spec — if they do, the benchmark is too easy and must be tightened before any agent comparison is meaningful.
The strongest existing competitor is AutoSizer, run unmodified: the LLM multi-agent sizing system with its reflective manager loop, from AutoSizer: Automatic Sizing of Analog and Mixed-Signal Circuits via Large Language Model (LLM) Agents. It is the strongest because it holds the field's best published number on the closest benchmark — a 100% success rate on the Hard tier of AMS-SizingBench, against 50% for standard Bayesian optimization — and our benchmark is deliberately built to be a sibling of AMS-SizingBench. Fairness means identical circuits, identical Ngspice simulator and SkyWater 130nm PDK, an identical simulation-count budget per circuit, and the same hyperparameter tuning effort the original paper used; nobody gets extra compute. Beating it means a higher post-layout success rate on Bio-AFE-Bench at equal simulation count, or the same success rate with materially fewer simulations. Classical references are batch Bayesian optimization via Batch Bayesian Optimization via Multi-objective Acquisition Ensemble for Automated Analog Circuit Design under the same budget, since plain Bayesian optimization is the reference AutoSizer itself reported at 50% on Hard circuits. No reported result exists on Lacuna for any of these tools on biosensing specifications — that absence is precisely the gap this proposal measures.
The ablation that isolates the key ingredient removes, one at a time, the three additions: the specification compiler (the agent receives only a raw electrical spec sheet with no biological context), the curated weak-inversion toolset, and the parasitic-closure loop. The precedent for expecting a single ingredient to dominate is AutoSizer's own finding that removing its self-reflection loop caused an 80% drop in success rate; our ablations are designed to find out which ingredient, if any, carries that weight here. On the system side, the closed-loop controller is compared against a static one-shot bolus, a hand-timed fixed schedule, and a simple threshold controller that actuates whenever any single biomarker crosses a fixed line — the last of these is the strongest system-level baseline, because if a dumb threshold on a well-sensed signal navigates the biphasic window as well as a state-aware controller, the sophisticated estimator is not earning its power budget.
Why not just…?
Why not just run the existing tools on the new specs unchanged? This is not a strawman — it is Experiment 1 and the strongest baseline, and it might work. The evidence for shortfall is indirect but consistent: AICircuit measured catastrophic accuracy loss under a milder regime shift than this one, and AutoCircuit showed agents underusing exactly the domain heuristics that weak-inversion design requires. But if AutoSizer unmodified clears, say, 90% post-layout success at equal simulation budget, the "regime gap" claim is dead, and we will say so; in that case the deliverable reframes to the benchmark and the closed-loop system score, which would remain unaddressed by any prior work. The experiment settles it either way.
Why not just hand-design the front-end, as every implantable-electronics group does today? For one chip, hand design is genuinely fine, and we will include one expert-designed reference circuit per block as a golden model regardless. But hand design is exactly the "dark art" bottleneck the automation literature documents, and its cost structure bites hardest where specifications churn — different cytokine panels, sensor modalities, and power budgets per indication or even per patient. More fundamentally, a hand design answers no scientific question about whether automation generalizes across specification regimes, which is the question this proposal exists to test.
Why not just do the control loop in software on an off-the-shelf microcontroller? The feasibility study did exactly that, and well: its Random Forest surrogate hit a within-subject error of $0.03 \pm 0.01$ predicting muscle twitch force. But its inputs were electrophysiological signals many orders of magnitude larger than a plasmonic sensor's picoamp output, and its measured cross-subject error of $0.20 \pm 0.04$ is itself a warning about how little margin a noisy biological signal leaves a controller. Resolving picomolar cytokine transients requires the custom low-noise analog chain first; a general-purpose microcontroller's analog noise floor and power draw cannot deliver it at implant scale. We absorb that paper's surrogate finding as a component of our loop rather than competing with it.
Why not just improve the biology — better exosomes, gene targets, metabolic reprogramming — and skip the circuits? Those lines of work, from embryonic-stem-cell conditioned media that clears the lipid overload locking macrophages in the M1 state (Rescuing macrophage normal function in spinal cord injury with embryonic stem cell conditioned media) to silk nanoparticles that quench the reactive-oxygen-species feedback sustaining inflammation (Silk-engineered bioactive nanoparticles for targeted alleviation of acute inflammatory disease via macrophage reprogramming), are therapies with static delivery kinetics — and the clinical literature's central complaint is that static kinetics cannot navigate a biphasic window in which the same signal is first required and then destructive. Our work is complementary: it builds the sense-and-time substrate those therapies lack, and its actuation model is calibrated from their published in vivo endpoints.
Proposed Experiments and Decisive Checks
Experiment 1 measures the regime gap: AutoSizer unmodified, batch Bayesian optimization, and random search run on Bio-AFE-Bench under identical simulation budgets, scored on pre-layout and post-layout success rate, simulations-to-convergence, and OSIRIS-style parasitic degradation. The reference points are AutoSizer's published 100% Hard-tier success and Bayesian optimization's 50% on AMS-SizingBench; the hypothesis is a substantial drop on the bio specifications. Experiment 2 runs the ingredient ablations, reporting success-rate deltas large enough to compare against the 80% single-ingredient drop AutoSizer's own ablation produced. Experiment 3 scores the whole loop: window-hit rate across simulated healing trajectories for the closed-loop controller versus the static bolus, hand-timed schedule, and threshold controller, with the state estimator's discrimination validated against the external GSE143735 cohort (target: comparable to the 0.944 area under the curve reported for the EREG/CSTA model), and wound-closure trajectories anchored to the Day-21 endpoint from the $db/db$ mouse study. Experiment 4 is the kill criteria, stated in advance: if AutoSizer unmodified reaches 90% post-layout success at equal budget, the regime-gap claim is falsified; if closed-loop timing shows no outcome advantage over a hand-timed schedule under realistic sensor noise, the motivation collapses to a power-and-area argument; and if the macrophage dynamics model proves too underdetermined for any of these distinctions, we report the circuit-level results as the contribution and downgrade the closed-loop section to explicit scenario analysis. In vivo validation in a $db/db$ wound model is the stated long-term destination, not a claim of this work; the proposal's testable output is silicon-ready, post-layout-verified front-end designs and the first benchmark that binds analog design automation to a real biological specification.
References
- AICircuit: A Multi-Level Dataset and Benchmark for AI-Driven Analog Integrated Circuit Design
- AutoCircuit: Agentic Pareto Front Exploration for Analog Circuit Designs
- OSIRIS: Bridging Analog Circuit Design and Machine Learning with Scalable Dataset Generation
- AnalogGenie: A Generative Engine for Automatic Discovery of Analog Circuit Topologies
- Bioaugmented osteosynthesis: precise monitoring and intervention of the bone healing microenvironment
- Adipose-Derived Mesenchymal Stem Cells Accelerate Diabetic Foot Ulcer Healing by Promoting Macrophage M2 Polarization Through Downregulation of EREG and CSTA
- A feasibility study on AI-controlled closed-loop electrical stimulation implants
- Optogenetic engineered umbilical cord MSC-derived exosomes for remodeling of the immune microenvironment in diabetic wounds and the promotion of tissue repair
- AutoSizer: Automatic Sizing of Analog and Mixed-Signal Circuits via Large Language Model (LLM) Agents
- Batch Bayesian Optimization via Multi-objective Acquisition Ensemble for Automated Analog Circuit Design
- Rescuing macrophage normal function in spinal cord injury with embryonic stem cell conditioned media
- Silk-engineered bioactive nanoparticles for targeted alleviation of acute inflammatory disease via macrophage reprogramming