Observation-Guided Computation: Sequential Reobservation for Robust Chromatographic Peak Detection
Year: 2026
Type: preprint
Abstract: Chromatographic peak detection is the bottom-most step of the data-processing chain, yet almost every classical detector defines its decision variables in units of signal amplitude, so those variables co-vary with baseline drift, injection amount, and detector gain, and a single-pass decision is made once against an unchanged trace. We reframe detection not as a better criterion but as a change in the computational organization of observation, and we do so in three ordered layers. Layer 1 — Observation-Guided Computation (OGC) is a sequential computational framework — an analytical data-interpretation principle rather than a new peak detector — built around the Observation Deviation (OD) measure established in our companion work13. OGC does not introduce another peak criterion or redefine OD; it embeds OD as an observation measure into an observe–detect–reconstruct–flatten–reobserve loop, so the observation condition is updated after each detected structure. Within a working window the trace is normalized to a unitless observation space, scored by OD against a single dimensionless threshold, and all final boundaries are re-anchored on the raw signal. OGC is deliberately independent of any particular candidate generator: it can be composed from any apex finder, and six heterogeneous engines all converge to a common performance plateau. Under one frozen configuration OGC reaches interval-coverage F1 = 1.000 on three of four synthetic benchmark sets with zero false positives, converts six heterogeneous apex-finder engines into region locators that all converge to F1 ≈ 1.000 (lifting bare derivative and moving-average detectors, unusable at F1 ≤ 0.090, by up to +0.967), is invariant to four orders of magnitude of dynamic range, and identifies high-confidence observable structures on two real GC–MS chromatograms where no global prominence fraction could serve both. Layer 2 — OGC-Net asks whether the representation of chemical observation can itself be learned: it algorithm-unrolls OGC's recursive loop into a differentiable K = 5-round network of 160,516 parameters. It runs a 150-point segment in 13.9 ms (batch 1) to 0.92 ms (batch 128) on CPU and, evaluated under the identical matching protocol, beats the frozen deterministic OGC by +0.116 peak-level F1 precisely on the harder, deployment-identical windowed distribution (0.6944 vs 0.5782). Layer 3 — OGC-Web delivers OGC-Net as a single offline, zero-backend HTML file whose forward pass is a pure-JavaScript port with base64-embedded weights, reproducing the peak-level agreement with the OGC reference client-side. Our central message is that chromatographic observation can be quantified and computationally organized as a sequential process, and that learning the framework — not fitting output labels — is the stronger objective when a deterministic gold-standard algorithm already exists.
Source: ChemRxiv
Authors: Zhang Zhi Xi, Chua Chun Kiang, Hua-Jun Zhang
Institution: Nanyang Technological University
Language: English
Subfield: Analytical Chemistry
Field: Chemistry
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7211941331