How ICA-Var Detects COVID Variants Early from Wastewater Using Machine Learning

Background and Motivation

Since the COVID‑19 pandemic began, the rapid evolution of SARS‑CoV‑2 has produced multiple variants with differing transmissibility and immune‑escape capabilities, challenging traditional clinical surveillance that relies on individual testing and extensive laboratory resources.

Wastewater‑based epidemiology (WBE) offers a community‑level, unbiased monitoring approach by detecting viral RNA shed in sewage, providing early warning of emerging variants without requiring individual participation.

Limitations of Existing Wastewater Methods

Current tools such as Freyja and COJAC depend on predefined variant barcode libraries derived from databases like GISAID. When novel variants lacking known mutation signatures appear, these methods often fail to identify them promptly, reducing the effectiveness of WBE.

ICA‑Var Method Overview

The University of Nevada, Las Vegas team introduced ICA‑Var (Independent Component Analysis of Variants), an unsupervised machine‑learning pipeline that applies independent component analysis (ICA) to wastewater mutation‑frequency matrices, extracting independent co‑variant mutation patterns that correspond to distinct viral lineages.

After ICA, a dual‑regression step projects the independent components back onto the original samples, quantifying each variant’s relative abundance over time and space.

Sample Collection and Laboratory Processing

From August 2021 to November 2023, 3,659 wastewater samples were collected across urban and rural sites in southern Nevada. Samples were kept on ice and processed within 36 hours.

Nucleic acids were extracted using Promega’s Wizard Enviro Total Nucleic Kit (A2991) with a modified protocol that employed proteinase digestion and Macherey‑Nagel NucleoMag Beads (744970). RNA >10 ng was reverse‑transcribed with NEB LunaScript RT SuperMix.

Sequencing libraries were prepared with Paragon Genomics CleanPlex SARS‑CoV‑2 FLEX Panel and sequenced on Illumina NextSeq 500/1000 (300‑cycle flow cell).

Bioinformatic Workflow

Raw reads were trimmed with cutadapt 4.2, aligned to the reference genome NC_045512.2 using bwa mem 0.7.17‑r1188, and primer sequences were removed with fgbio TrimPrimers 2.1.0 (hard‑clip mode). Variant calling employed iVar v1.4.1, and coverage/depth metrics were calculated with samtools v1.16.1.

After removing duplicates and controls, 2,684 samples passed initial QC. A stringent filter retained only samples with ≥50× depth and ≥80 % genome coverage, yielding 1,385 high‑quality samples representing 59,422 distinct mutations.

Validation and Comparative Performance

ICA‑Var’s results were benchmarked against clinical sequences from GISAID (8,810 high‑coverage Nevada genomes, Sep 2021–Nov 2023) and against the gold‑standard tool Freyja.

ICA‑Var consistently detected Delta, Omicron, and recombinant XBB variants earlier than Freyja, with an average lead time of 1–4 weeks for emerging lineages such as EG.5, HV.1, and BA.2.86.

Urban vs. Rural Variant Dynamics

Weekly analyses revealed that 16 of the 18 monitored variants were first detected in urban wastewater before appearing in rural samples, confirming a typical city‑to‑rural spread pattern.

Temporal Mutation Evolution Analysis

The team identified 177 mutation sites with significant temporal contributions and compared them to hallmark mutations of Delta (B.1.617.2), Omicron BA.1, and recombinant XBB.1. Distinct temporal patterns aligned with known epidemiological waves, illustrating the method’s ability to track mutation dynamics.

Potential Future Variant Signatures

Hierarchical clustering of 113 candidate novel mutations (derived from 15 known variants) produced six feature clusters. Four clusters overlapped with mutations observed in late‑2023 variants, while clusters 1 and 6 contained mutations absent from current clinical data, suggesting possible precursors of future lineages.

Conclusions

ICA‑Var demonstrates that integrating unsupervised machine learning with dual‑regression analysis can overcome the reliance on predefined variant barcodes, delivering earlier and more reliable detection of SARS‑CoV‑2 variants from wastewater. The approach also elucidates spatial transmission patterns and highlights mutation signatures that may herald new variants, offering a cost‑effective, high‑resolution tool for public‑health surveillance.