Rethinking massive multiplexing in whispering gallery mode biosensing

Accurate, label-free quantification of multiple analytes in complex biological media remains a major challenge due to limited multiplexing, signal cross-correlations, and inconsistency across sensor samples and measurement runs. We introduce a multiplexed whispering-gallery-mode (WGM) biosensing framework that overcomes these barriers by jointly advancing photonic integration and data analytics. Our glass-chip platform enables massive, parallelized and flexible multiplexing of >10000 microresonators organized into up to 100 sensing channels, with universal and modular chip design and detection hardware, while maintaining loaded Q-factors of 10^6. Our novel hybrid deep-learning framework BioCCF that integrates domain adaptation with cross-channel fusion enables harmonization of responses across sensing chips and extraction of nonlinear correlations in complex mixtures. Using a highly heterogeneous dataset comprising over 200 hours of sensing data acquired from nine chips with different channel configurations, biological replicates, and repeated regeneration cycles, we demonstrate recalibration-free identification of solution (99.3% accuracy) and quantification of immunoglobulin G components with relative prediction error of 10^-4 under 5 min. The affordability and modularity of the platform enable distributed data acquisition and aggregation into shared repositories, providing a pathway toward continuously improving model generalization, cross-validation and a scalable, community-driven paradigm for biosensing.

💡 Research Summary

The paper presents a comprehensive solution to two long‑standing bottlenecks in whispering‑gallery‑mode (WGM) biosensing: (1) the inability to scale multiplexed detection to the number of analytes present in complex biological samples, and (2) the lack of robust data‑analysis pipelines that can handle the inevitable variability between sensor chips, measurement runs, and experimental conditions.

Hardware platform – The authors fabricate a glass‑chip sensor that integrates up to 10 000 high‑Q (≈10⁶) microresonators distributed across as many as 100 logical channels. Microparticles (106–125 µm silica spheres) are positioned using 3D‑printed masks and permanently immobilized with a low‑refractive‑index polymer (MY‑133MC). A spin‑coated “distancing” layer of ≈430 nm separates the resonators from the substrate, dramatically reducing coupling losses while preserving strong evanescent field overlap. By varying the number of channels (4, 9, 16, up to 100) the authors demonstrate that the median resonator count per channel drops from ~235 to ~100, yet inter‑chip variance also declines, establishing a practical trade‑off between density and uniformity. Optical characterization shows that, in aqueous environments, the loaded Q‑factor reaches the mid‑10⁵ range for >60 % of resonators, with a maximum exceeding 3 × 10⁶.

Domain adaptation (DA) – Even with identical design rules, each chip exhibits a unique spectral fingerprint because of microscopic size variations and fabrication tolerances. To map heterogeneous measurements onto a common feature space, the authors convert wavelength‑resolved spectra (≈4800 points per sweep) into intensity‑based representations, then apply several linear DA techniques: Principal Component Analysis (PCA), Transfer Component Analysis (TCA), Maximum Independence Domain Adaptation (MIDA), and Joint Probability Domain Adaptation (JPDA). Performance is quantified by the root‑mean‑square error (RMSE) between the first adapted feature (AF1) and a generalized spectral shift derived from all resonators in a channel. With 100 resonators per channel, TCA achieves the lowest median RMSE (0.037) and the smallest outlier spread, indicating that it can reliably align source and target domains across chips without sacrificing sensitivity. The calibrated AF1 difference between regeneration (PBS‑REG) and buffer (PBS) phases serves as a universal reference, enabling chip‑to‑chip consistency without explicit recalibration.

BioCCF deep‑learning framework – Building on the DA‑harmonized data, the authors introduce Biological Cross‑Channel Fusion (BioCCF), a hybrid neural architecture that merges a CNN‑biLSTM classifier with a Transformer‑Encoder regressor through a cross‑attention fusion block. Each logical channel (including dedicated control channels) contributes a time‑series of AF1 values; the CNN extracts local patterns, the biLSTM captures temporal dynamics, and the Transformer learns long‑range dependencies across channels. Cross‑attention explicitly models inter‑channel correlations, allowing the network to remain robust when some resonators are missing or exhibit irregular behavior. The model is trained on a heterogeneous dataset comprising >200 h of sensing data from nine chips, spanning multiple channel configurations, biological replicates, and repeated regeneration cycles.

Experimental validation – In a series of immunoassay experiments, BioCCF identifies the presence of a target solution with 99.3 % accuracy and quantifies individual IgG components in complex mixtures with a relative prediction error of 10⁻⁴ (≈0.01 % deviation) within 5 minutes of measurement. Crucially, these results are obtained without any chip‑specific calibration, demonstrating that the combination of DA and cross‑channel fusion effectively eliminates the need for per‑chip recalibration.

Implications and outlook – The platform’s affordability (glass‑chip, modular optics) and scalability (up to 10⁴ resonators) make it suitable for point‑of‑care and field deployments. Moreover, the authors envision a distributed data‑acquisition ecosystem where multiple laboratories upload raw sensorgrams to a shared repository; the aggregated dataset continuously refines the BioCCF model, improving generalization and cross‑validation in a community‑driven manner. This paradigm shift could accelerate the adoption of high‑performance optical biosensors for personalized medicine, environmental monitoring, and beyond.