MAKO: Multi-Domain Research Innovations

• MAKO in submillimeter imaging is a 350 μm camera prototype that utilizes advanced LEKID arrays for photon-noise-limited imaging with high multiplexing efficiency.
• MAKO in control systems is a meta-adaptive Koopman operator framework that leverages deep lifting networks for robust, real-time prediction and stability in nonlinear systems.
• MAKO in digital pathology is a unified benchmarking framework using attention-based models to predict breast cancer recurrence risk with competitive performance against transcriptomic assays.

MAKO is an acronym applied to several distinct research concepts spanning instrumentation for submillimeter imaging, meta-adaptive operator theory for control of nonlinear systems, and digital pathology benchmarking frameworks for cancer risk prediction. Each context reflects a technically rigorous approach to scaling, interpretability, or adaptivity within its discipline, as detailed in the source literature.

1. MAKO for Submillimeter Astronomical Imaging Arrays

MAKO, in the instrumentation context, denotes a pathfinder 350 μm imaging camera designed for on-sky demonstration of low-cost, high-density detector arrays targeting large-format submillimeter telescopes such as CCAT (Swenson et al., 2012). The primary scientific motivation is to enable photon-noise-limited imaging with over $10^6$ pixels, surpassing prior state-of-the-art ($\sim$10^4$$in SCUBA-2), via three enabling advances: (i) extreme multiplexing density (≫$$10^3$$pixels per feedline), (ii) single-layer fabrication, and (iii) low per-pixel cost.</p> <p>MAKO deploys lumped-element <a href="https://www.emergentmind.com/topics/kinetic-inductance-detectors-mkid" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">kinetic inductance detectors</a> (LEKIDs) based on titanium nitride (TiN) films patterned onto high-resistivity silicon. Each pixel comprises an inductive meander, which directly absorbs 350 μm radiation, and an interdigitated capacitor to tune the resonance frequency. The resonance constraint is$$f_0 = 1/(2\pi \sqrt{LC})$, with the kinetic inductance fraction$\alpha_k = L_k/L \approx 0.5$$–0.8 for optimized responsivity and absorber filling.</p> <p>Multiplexing is achieved through frequency-domain readout, where each resonator has a distinct$$f_0$$and all couple to a single 50 Ω feedline. Pixel pitch in frequency (channel spacing$$\delta f$) is set to at least five times the linewidth,$\delta f \gtrsim 5(f_0/Q_r)$. At$Q_r \approx 5\times 10^4$and$\sim$0 in the 50–250 MHz range, $\sim$1 is 10–25 kHz, enabling $\sim$2 pixels per 200 MHz readout band.

Fabrication employs a single-mask optical lithography process for both the absorber and capacitor, yielding $\sim$$3\% array yield, with performance strongly linked to TiN deposition. MAKO is designed as a drop-in replacement for SHARC-II at the <a href="https://www.emergentmind.com/topics/chicken-swarm-optimization-cso" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">CSO</a> Nasmyth focus, with shared cold-optical chain, back-illumination through the substrate, and multi-stage pulse-tube and helium sorption cryogenics (detectors at$$\sim$4 mK).

Laboratory testing has demonstrated loaded $\sim$5 in 100-pixel arrays, bifurcation (nonlinear) behavior in line with kinetic-inductance theory, and frequency noise consistent with photon and generation-recombination processes ($\sim$6 W/$\sim$7). On-sky mapping speed and beam quality are expected to match or exceed existing kilo-pixel arrays.

Scalability studies indicate that extending MAKO's approach to CCAT ($\sim$8 pixels) would require $\sim$9100 readout lines and stringent resonance-frequency uniformity ($10^4$0), with ongoing work on wafer-scale TiN and custom ASIC/gPU readouts (Swenson et al., 2012).

Table 1: Typical LEKID Pixel Parameters

Parameter	Value / Range
Resonance frequency ($10^4$1)	50–250 MHz
Total inductance ($10^4$2)	~50–100 nH
Capacitance ($10^4$3)	~20–50 pF
Internal Q ($10^4$4)	$10^4$5
Coupling Q ($10^4$6)	$10^4$7
Loaded Q ($10^4$8)	$10^4$9
Resonator BW ($10^3$0)	2–5 kHz
Channel spacing ($10^3$1)	≥10–25 kHz
Max. multiplex/band	$10^3$28,000–20,000

2. MAKO as Meta-Adaptive Koopman Operator Framework

In nonlinear control, MAKO (Meta-Adaptive Koopman Operator) refers to a meta-learning-based extension of Koopman operator theory for learning-based model predictive control (MPC) under parametric uncertainty (Han et al., 10 Oct 2025). The central model describes discrete-time nonlinear systems with unknown, fixed parameters, $10^3$3 with $10^3$4.

MAKO meta-learns a deep-lifting network $10^3$5, sharing observables across a multi-modal set of system parameters, accompanied by task-specific linear Koopman triplets $10^3$6. Meta-training minimizes multi-step prediction error over parameter-induced task distributions, while online adaptation employs closed-form gradient updates on $10^3$7 to fit new, previously unseen $10^3$8.

Empirically, MAKO outperforms strong baselines such as deep stochastic Koopman (DeSKO) in modeling (mean squared error $10^3$9 over 16-step prediction) and closed-loop control across benchmarks: cartpole (variable pole length and mass), synthetic GRN, and chemical reactor+separator. Real-time feasibility is confirmed ($f_0 = 1/(2\pi \sqrt{LC})$00.02s/step). Theoretical contributions include Lyapunov-based proofs of online adaptation convergence and closed-loop stability under mild regularity, as well as robust variants addressing model residuals.

Limitations include dependence on finite-dimensional invariant subspace assumptions (i.e., lifted linear dynamics may imperfectly capture strong nonlinearity) and the open question of persistent-excitation sufficiency in meta-training and online adaptation. Potential extensions include uncertainty quantification, hybrid/system time-varying extensions, and experimental validation on real hardware (Han et al., 10 Oct 2025).

3. MAKO Benchmarking in Digital Pathology

MAKO ("Mammary Analysis for Knowledge of Outcomes") denotes a unified benchmarking framework for interpretable prediction of breast cancer recurrence risk (PAM50-based ROR-P score) from routine H&E whole-slide images (WSIs) (Kaczmarzyk et al., 16 Aug 2025). MAKO systematically compares 12 pathology foundation models (including CONCH, UNI, H-optimus-0, Virchow2) and two non-pathology baselines (ResNet50, ViT-DINOv2) across three tasks: binary classification (low/medium vs. high risk), regression (continuous ROR-P), and survival stratification (10-year recurrence).

The methodology involves patch-based feature extraction (128×128 μm² at ~0.50 μm/px), embedding via a pretrained encoder, and aggregation by attention-based multiple instance learning (ABMIL) with gated attention. Model outputs undergo downstream classification, regression, or Cox survival modeling.

Training and validation leverage the Carolina Breast Cancer Study (CBCS, $f_0 = 1/(2\pi \sqrt{LC})$1 WSIs, ROR-P labels) and external benchmarking on TCGA BRCA ($f_0 = 1/(2\pi \sqrt{LC})$2 WSIs). Performance metrics include ROC AUC, Pearson $f_0 = 1/(2\pi \sqrt{LC})$3, and Cox concordance (C-index). Multiple pathology foundation models outperform baselines; for classification: CONCH (AUC=0.809 CBCS/0.852 TCGA), UNI (0.808/0.825), and Phikon (0.799/0.824), versus ResNet50 (0.745/0.772). For regression, H-optimus-0, Virchow2, and Provi-GigaPath yield the highest $f_0 = 1/(2\pi \sqrt{LC})$4; all models match or surpass ResNet50 and ViT-DINOv2. Survival stratification (C-indices 0.55–0.62) is non-inferior to transcriptomic assays.

Interpretability is addressed via attention heatmaps and HIPPO perturbation analysis. Excluding tumor regions (necessity) drops high-risk prediction, while restricting to tumor (sufficiency) generally preserves or raises predicted risk. MAKO enables discovery of candidate tissue biomarkers: 120 patches featuring nuclear pleomorphism, high mitotic figures, necrosis, and paucity of tumor-infiltrating lymphocytes, which systematically elevate ROR-P predictions when inserted into low-risk WSIs across all pathologist-trained models.

MAKO highlights the effective, scalable, and interpretable use of pathology foundation models for recurrence risk prediction in ER+/HER2- breast cancer. It demonstrates parity with transcriptomics while offering workflow advantages and candidate biomarker discovery capabilities (Kaczmarzyk et al., 16 Aug 2025).

Table 2: Performance of Top Pathology Foundation Models

Task/Metric	Model	CBCS	TCGA
Classification (AUC)	CONCH	0.809	0.852
Classification (AUC)	UNI	0.808	0.825
Regression ($f_0 = 1/(2\pi \sqrt{LC})$5)	H-optimus-0	0.638	0.443
Regression ($f_0 = 1/(2\pi \sqrt{LC})$6)	Virchow2	0.627	0.587
Survival (C-index)	All ABMIL models	0.55–0.62	—

4. Common Technical Themes and Innovations

Despite disciplinary differences, MAKO is consistently characterized by:

• Emphasis on scalable, high-density or multi-task architectures (LEKID arrays, Koopman lifting, pathology encoders).
• Utilization of multiplexing, either in frequency (imaging arrays), meta-learned feature spaces (systems control), or dense attention-based representations (digital pathology).
• Integration with existing workflow and infrastructure (instrument drop-in, online adaptation, digital slide processing).
• Strong focus on interpretability (closed-form learning rules, attention heatmaps, perturbation diagnostics).

5. Future Directions and Open Challenges

Instrumental MAKO development points to further scaling via ASIC/GPU readouts and wafer-scale lithography (Swenson et al., 2012). Meta-adaptive Koopman frameworks highlight the need for real-world experimental validation, deeper understanding of persistent excitation for meta-learning, and robust uncertainty quantification (Han et al., 10 Oct 2025). Pathology benchmarking frameworks motivate prospective clinical trials, integration with additional data modalities, and analytical advances in interpretability and biomarker validation (Kaczmarzyk et al., 16 Aug 2025). These reflect the ongoing synthesis of multiplexed measurement, adaptive learning, and interpretable prediction across domains.