Bees & Insects

Phenotyping through the mess: Contamination-aware NIR models for prediction of amino acid composition in live black soldier fly larvae for selective breeding applications

---

Abstract

Commercial production of black soldier fly (Hermetia illucens, BSF) relies heavily on the nutritional quality of larvae, with amino acid (AA) composition being a major determinant of product value and downstream feed performance. Improving larval AA profiles through selective breeding is therefore of growing interest, yet advancement of breeding applications is constrained by the need for large volumes of accurate phenotypic data on live selection candidates. Conventional chemical analyses of nutritional composition are costly, time-consuming, and destructive, making them unsuitable for high-throughput phenotyping within breeding programs. Near-infrared (NIR) spectroscopy offers a rapid, non-destructive alternative for predicting biochemical traits, including AA composition. However, standard NIR calibration approaches assume that samples are clean and homogeneous, an assumption incompatible with industrial BSF production, where larvae are always in contact with heterogeneous rearing substrates whose residues strongly distort spectral signatures.Here we present a contamination-aware NIR modeling framework that integrates substrate spectra with clean and contaminated larval scans to predict AA composition of live larvae under realistic rearing environments. BSF were reared on 17 experimental diets, producing 204 group-level larval samples, each measured twice using point-based NIR spectroscopy, alongside paired wet- and dry-matter diet spectra. Clean larval groups were used exclusively for model training, with their spectra duplicated into a à¢â‚¬Å“dirtyà¢â‚¬Â predictor block to mimic substrate interference. Diet spectra were incorporated as structured predictor blocks. All contaminated larval samples, bearing real substrate residues, were reserved strictly for testing, providing a rigorous real-world performance benchmark.Using block-structured partial least squares regression, we modeled 20 free amino acids from chemically quantified composite larval samples. Under contaminated test conditions, prediction accuracy ranged from modest to strong. Top-performing traits included Glutamate (R² = 0.66), Proline (0.48), Glycine (0.45), Tryptophan (0.34), and Phenylalanine (0.30), while several essential amino acids (Ile, Met, Val, His) reached R² = 0.20à¢â‚¬â€œ0.26. These values show that meaningful biochemical signal can be recovered even when spectra are heavily confounded by substrate contamination, substantially outperforming traditional clean-only NIR pipelines.This study demonstrates that contamination-aware NIR models can successfully recover meaningful biochemical information from live, substrate-covered larvae, thereby enabling non-destructive, high-throughput, crate-level phenotyping. Such models unlock the possibility of large-scale genetic improvement programs targeting AA composition in BSF without removing larvae from production environments. This represents a major step toward integrating real-time nutritional phenotyping into industrial BSF breeding pipelines.

Keywords: 2026

How to Cite:

Gebreyesus, G., Jensen, K., Andersen, L., Noel, S., Buitenhuis, A. & Zaalberg, R., (2026) “Phenotyping through the mess: Contamination-aware NIR models for prediction of amino acid composition in live black soldier fly larvae for selective breeding applications”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2286852. doi: https://doi.org/10.31274/wcgalp.24144

Rights: 1