Defense Allocation Costs Explain Antifungal Synergy-Score Discordance

When two antibiotics are combined, clinicians decide whether the pair is "synergistic" using one of two mutually inconsistent mathematical yardsticks. The first, the Fractional Inhibitory Concentration Index (FICI), asks whether each drug behaves as a diluted copy of the other; the second, Bliss response surface analysis, asks whether their effects multiply as if acting independently. On identical checkerboard assays these two yardsticks routinely disagree: for isavuconazole paired with the flavonoid isoquercitrin, FICI called 100% of 30 Candida glabrata isolates synergistic while Bliss analysis called only 60% (Synergistic In Vitro Interaction of Isavuconazole and Isoquercitrin against Candida glabrata), and for amphotericin B plus colistin the disagreement flips direction by species — FICI found synergy in 33% of Candida auris strains where Bliss found 73% (In Vitro Activity of Amphotericin B in Combination with Colistin against Fungi Responsible for Invasive Infections). The field treats this as measurement noise to be standardized away.

I propose it is not noise. It is a signal about how fungal populations pay for drug defense. The core claim: when a fungal population is exposed to a drug combination, individual cells reallocate metabolic machinery from growth to defense — efflux pumps and stress-response enzymes — and this reallocation has a measurable cost in lag time and growth rate. Populations that bet-hedge, splitting into a fast-growing defense-poor majority and a slow defense-rich minority, produce exactly the non-linear, time-dependent dose-response surfaces that make the Loewe and Bliss null hypotheses diverge. If this is right, then the FICI-versus-Bliss disagreement is not an artifact but a free diagnostic of defense allocation in the population, and a mechanistically grounded kinetic scoring protocol can replace both ad hoc metrics, predict which isolates will resist combination therapy through genetic compensation, and predict which combinations will hold up evolutionarily.

Core Claim

The proposal rests on one mechanism and two testable consequences. The mechanism extends the resource-allocation theory developed for yeast nutrient shifts by Phenotypic heterogeneity in lag reflects an evolutionarily stable bet-hedging strategy, which showed that clonal yeast facing a glucose-to-maltose switch split into "recoverers" and "arresters," and that fast pre-shift growth imposes prolonged post-shift lag through a strict power-law trade-off in how enzymes are allocated. That model assumes two non-toxic carbon sources and static enzyme pools; it has never been applied to a toxic stressor. I extend it so that a defense allocation share $\phi$ of the cell's enzyme budget buys drug protection at the price of growth:

$$\mu(\phi, c) = \mu_{\max}\,(1-\phi)\,\exp\!\bigl(-c/K(\phi)\bigr), \qquad K(\phi) = K_0\, e^{\,a\phi}$$

Here $\mu$ is the growth rate, $c$ the drug concentration, $\mu_{\max}$ the drug-free maximum, and $K(\phi)$ the protection scale that grows exponentially with the defense share $\phi$. Because raising $\phi$ after drug exposure takes time, populations enter a lag — the trailing growth that static 24- or 48-hour endpoint assays truncate, and that the Droplet 48 kinetic platform has already shown produces severe diagnostic errors in single-drug susceptibility testing (Rapid automated antifungal susceptibility testing system for yeasts based on growth characteristics). Two consequences follow. First, a mixture of cells with different pre-set $\phi$ values yields a population dose-response surface that violates Loewe additivity and Bliss independence in different directions, so FICI-Bliss discordance should be predictable from kinetic signatures of allocation heterogeneity. Second, because defense machinery in fungal pathogens is built from redundant, paralogous parts — efflux transporters and stress pathways that compensate for one another — a combination therapy only becomes durably lethal when it makes compensation unaffordable, that is, when the energy budget cannot support the $\phi$ the drug pressure demands.

Research Program

Three linked components. First, a standardized kinetic interaction-mapping protocol: checkerboard assays read by continuous resazurin fluorescence (the Droplet 48 measurement principle), with pre-registered endpoint extraction at multiple time points so the same raw wells yield FICI, Bliss, and time-resolved growth curves simultaneously. This makes scoring heterogeneity itself an experimental variable rather than a nuisance, directly addressing the open problem that researchers switch between Loewe and Bliss without justification (Synergistic In Vitro Interaction of Isavuconazole and Isoquercitrin against Candida glabrata). Second, a hierarchical fit of the allocation model above to the kinetic traces, recovering per-isolate parameters: the distribution of $\phi$ across the population (the bet-hedging rheostat), the switching time $\tau$, and the growth tax on defense. Third, a redundancy layer: using the Saccharomyces cerevisiae synthetic genetic array network — roughly 4,000 interactions among ~1,000 genes (Global Mapping of the Yeast Genetic Interaction Network) — transferred by orthology to C. glabrata, to prioritize which paralogous defense pairs (efflux transporters, ERG-pathway regulators) most plausibly buffer single-target adjuvants. This is needed because prior redundancy-mapping work has only ever been run under nutrient-rich, drug-free conditions on ~2,700 genes (Exploring weighting schemes for the discovery of informative generalized between pathway models to uncover pathways in genetic interaction networks); whether compensation holds under chemical stress is precisely the unmapped question.

The drug pairs anchor the program in published benchmarks. Amphotericin B plus lansoprazole is the strongest mechanistic probe: lansoprazole blocks mitochondrial Complex III, drains bioenergetics, and restores amphotericin B killing with ΣFICI of 0.25–0.50 across 90% of 20 C. auris isolates from four clades (Lansoprazole interferes with fungal respiration and acts synergistically with amphotericin B against multidrug-resistant Candida auris). My model says this works not merely through reactive oxygen species but because energy depletion makes the defense allocation $\phi$ unaffordable — compensation costs ATP the cell no longer has. Itraconazole plus aprepitant is the counter-case: aprepitant perturbs metal-ion homeostasis, yet the fungus responds with compensatory upregulation of ergosterol biosynthesis genes (Aprepitant, an antiemetic agent, interferes with metal ion homeostasis of Candida auris and displays potent synergistic interactions with azole drugs) — exactly the dynamic, transcription-triggered buffering that the redundancy literature suspects but rarely captures, and which chromatin-level regulation can mediate under azole challenge (Set1-mediated histone H3K4 methylation is required for azole induction of the ergosterol biosynthesis genes and antifungal drug resistance in Candida glabrata).

Proposed Experiments

Experiment 1 tests the discordance-prediction claim on the two published combination panels: the 30-isolate C. glabrata set from the isavuconazole/isoquercitrin study and the 20-isolate, four-clade C. auris set from the lansoprazole study, plus the amphotericin B/colistin pairing from the 86-isolate invasive-fungus panel (In Vitro Activity of Amphotericin B in Combination with Colistin against Fungi Responsible for Invasive Infections). Each checkerboard is read kinetically; the allocation model is fitted per isolate; and the pre-registered prediction is that isolates where FICI and Bliss disagree are those whose kinetic traces show high recoverer fractions and extended trailing — the bet-hedging signature. Success is per-isolate prediction of discordance at 80% accuracy or better, a threshold justified by the size of the disagreement signal already on record (100% versus 60% in C. glabrata; 33% versus 73% in C. auris).

Experiment 2 tests the redundancy claim in C. glabrata. Single paralog deletions in efflux and stress pathways are predicted to show little phenotype in drug-free conditions — the classic masking that formal lesion-analysis work showed requires multi-element perturbation to unmask (Who Does What? A Novel Algorithm to Determine Function Localization) — and to remain masked under single-drug azole pressure, but to be revealed when the energy drain (lansoprazole-style Complex III inhibition) is added, because compensation becomes metabolically unpayable. This directly interrogates the limitation of single-target silencing protocols such as knocking down one efflux transporter while its paralogs compensate (siRNA-Mediated Silencing of Antifungal Resistance Genes: A Research Protocol). The measurement is the gap between single- and double-perturbation drug response, quantified across serial drug concentrations with and without the energy stressor.

Experiment 3 tests evolutionary durability. Populations are serially passaged under single adjuvant, single antifungal, and the combination, with resistance emergence rate as the primary endpoint and the fitted allocation cost (lag extension per unit drug) as the predictor. The prediction is that combinations imposing the highest defense-allocation cost select resistance most slowly, and that the resistance that does emerge follows the reversible-aneuploidy route documented for Candida albicans under aureobasidin A — chromosome gains that upregulate multiple resistance loci at once and revert when the drug is withdrawn (Chromosome 1 trisomy confers resistance to aureobasidin A in Candida albicans). Karyotyping and growth-cost measurement of resistant clones close the loop: resistant clones should carry a measurable growth-defense tax that the allocation model predicted in advance.

Baselines

The naive reference point is the field's current default: a static endpoint FICI checkerboard under EUCAST conditions, exactly as run in the isavuconazole/isoquercitrin study, which reported 100% of 30 C. glabrata isolates as synergistic by FICI against 60% by Bliss response surface analysis (Synergistic In Vitro Interaction of Isavuconazole and Isoquercitrin against Candida glabrata). This baseline keeps the discordance unexplained and unexplained discordance blocks clinical translation, since a pair can look synergistic or indifferent depending on the arbitrary choice of null model. Beating it means delivering a scoring procedure that reproduces both metrics where they agree, explains per-isolate disagreements mechanistically, and is pre-registered so no post hoc metric choice is possible.

The strongest existing competitor is the Droplet 48 kinetic susceptibility platform, which achieved over 90% essential agreement with the Sensititre YeastOne gold standard for 8 of 9 antifungals (posaconazole at 86.62%), categorical agreement above 90% for four agents, and 100% reproducibility within ±2 two-fold dilutions across 144 clinical isolates (Rapid automated antifungal susceptibility testing system for yeasts based on growth characteristics). It will be run on the same isolate panels, the same fluorescence hardware, and the same wells as my protocol, extended to two-drug plates with an equal tuning budget for its native kinetic features (threshold, slope, acceleration). Beating it means showing that its raw kinetic features cannot predict FICI-Bliss discordance or resistance-emergence rates, while the allocation-model parameters fitted to the identical traces can. This is the fairest and most dangerous comparison, and it is decided entirely within Experiment 1 and Experiment 3.

The key ablation isolates the proposal's central ingredient: the allocation model versus a generic growth-curve description. A logistic or Gompertz fit is applied to the same kinetic traces with the same number of free parameters, and the two parameterizations are compared on per-isolate discordance prediction and on predicting the serial-passage outcomes of Experiment 3. If the generic fit performs as well, the defense-allocation story adds narrative but not predictive power, and the proposal's contribution would collapse to the kinetic protocol itself. For the record, no reported result found on Lacuna for any kinetic predictor of FICI-Bliss discordance or of combination-durability outcomes; the closest prior quantitative anchor is the lansoprazole study's 90% growth inhibition and fungicidal restoration within 6 hours (Lansoprazole interferes with fungal respiration and acts synergistically with amphotericin B against multidrug-resistant Candida auris), which is a single-metric FICI result with no kinetic or redundancy analysis.

Why not just…?

Why not just run a bigger FICI-versus-Bliss concordance study? Because the two published studies already quantify the discordance — 100% versus 60% in C. glabrata (Synergistic In Vitro Interaction of Isavuconazole and Isoquercitrin against Candida glabrata) and 33% versus 73% in C. auris (In Vitro Activity of Amphotericin B in Combination with Colistin against Fungi Responsible for Invasive Infections) — and more isolates would only refine the correlation, not establish a cause. The direction of the disagreement flips between species, which no concordance study explains but a heterogeneity mechanism must. Experiment 1 is designed to answer exactly the question a concordance study cannot: which isolates disagree, and why.

Why not just standardize on one metric, say Bliss, and move on? Choosing a single null model hides the phenomenon. Since FICI overcalls synergy in C. glabrata and undercalls it in C. auris relative to Bliss, neither metric is uniformly correct, and the discrepancies are species- and isolate-dependent. My claim is that the disagreement itself carries mechanistic information; discarding one metric discards that information. The decisive check is whether the disagreement is systematically predictable from kinetic allocation signatures — if it were random noise, no per-isolate predictor would reach the 80% accuracy target.

Why not just extend the Droplet 48 platform to combinations and report better endpoint agreement? This is the most serious alternative, and I treat it as a genuine threat rather than arguing around it. A purely descriptive kinetic extension would be simpler and might stabilize endpoint choice at early read times before trailing emerges. That is why the Droplet 48 baseline runs on identical wells with an equal tuning budget, and why the ablation pits the allocation model against generic curve fits: if raw kinetic features suffice for discordance prediction, the honest conclusion is that the model contributes interpretation, not prediction, and I would report that. But raw features are not designed to forecast across conditions — which combination will be undermined by paralog compensation (Experiment 2) or how fast resistance emerges under passage (Experiment 3) — and those cross-condition predictions are where the allocation model must earn its place or fail visibly.

Why not just screen more repurposed adjuvants with FICI, as the lansoprazole study did? That path is already well trodden: 727 FDA-approved drugs were screened and lansoprazole identified by exactly this procedure (Lansoprazole interferes with fungal respiration and acts synergistically with amphotericin B against multidrug-resistant Candida auris). But the aprepitant study shows why screening alone stalls: the fungus answers with compensatory upregulation of the very pathways the adjuvant bypasses (Aprepitant, an antiemetic agent, interferes with metal ion homeostasis of Candida auris and displays potent synergistic interactions with azole drugs), and a static FICI screen cannot see that response coming. Predicting compensation and its metabolic cost is the point of this proposal, not a side benefit.

Decisive Checks

The proposal stands or falls on three numbers. First, per-isolate prediction of FICI-Bluss discordance from allocation-model parameters at 80% or better, with the Droplet 48 raw-feature baseline and the generic curve-fit ablation both falling short on the same data. Second, a measurable unmasking effect in Experiment 2: single paralog perturbations that are phenotypically silent under azole alone become growth-inhibitory only when energy stress is added, quantified as a significant interaction between perturbation order and stressor presence. Third, a negative correlation across combinations between fitted allocation cost and resistance-emergence rate over serial passage, with emerging resistant clones showing reversible aneuploidy and a growth tax consistent with the model's prediction. Any one of these failing would falsify the specific mechanism while still leaving the standardized kinetic interaction protocol — the methodological deliverable — intact.