Larvae rearing, dietary treatments, and collection procedure

The BSF larvae were collected at Enorm Biofactory (Flemming, Denmark) at the age of five to six days. Up until then, the larvae were fed with the standard feed used at Enorm Biofactory. After arrival at the research facility, the larvae were randomly assigned to different diet groups. The base ingredient of all diets was chicken feed (Paco Start, Danish Agricultural Grocery Company, Fredericia, Denmark), which was the standard diet used at the research facility. The composition of the dry matter chicken feed was 18.5% protein, 3.6% fat, 67.9% digestible carbohydrates, 4.7% fibre, 5.3% crude ash, and included 0.99 g lysine/kg, 0.45 g methionine/kg, 0.34 g cystine/kg, 0.98 g calcium/kg, 0.72 g phosphorus/kg and 0.15 g sodium/kg. Casein or sucrose was added at different proportions, which resulted in a total of seventeen diets (Table 1). The seventeen diets were developed to vary in protein content (2.3–89.8%) and digestible carbohydrate content (8.5–96.0%), which was done to maximize the variation in body composition between our larval samples. The diets were prepared the day before the larvae arrived. For each diet, four transparent plastic cups (70:95 mm H:Ø) were prepared with 60 g dry feed, 140 ml water, and 1 g agar. Afterwards, approx. 300 six-day old larvae were added to each cup. The cups were closed off with a lid with a 65 mm Ø hole covered with a fine mesh for ventilation. The cups were placed in a climate chamber, where the larvae were reared under controlled conditions (28 °C, 70% RH) until harvesting. The larvae were harvested on four consecutive days: day twelve, thirteen, fourteen and fifteen after egg hatching. The larvae were harvested at different ages to maximize the variation in body composition between our larval samples. On each harvesting day, for each of the seventeen dietary treatments, one of the four cups was taken out of the climate chamber and prepared for NIR spectral acquisition, followed by the preparation of the larval sample for chemical analysis of AAs.

NIR spectral acquisition

All samples of live larvae were scanned using a FOSS DS2500 NIR spectrometer. A cup with larvae was taken out of the climate chamber, and the content of the cup was put in a fine sieve and then rinsed with water. Once all residues were removed, the larvae were dried with paper towels. A randomly selected sample of the cleaned and dried live larvae was put in a small quartz cup for the NIR-scanning. The small quartz cup was filled with as many larvae as possible to minimize the surface area without any larvae (pin hole gaps). The absence of pin hole gaps is essential for successful NIR-scanning. The quartz cup with larvae was gently closed off with a metal lid and placed in the scanner. The sample was scanned twice. After the scanning was completed, the live larvae were moved to a paper towel and placed back in the cup again for a second scanning (repacking), using the same scanning procedure as for the first scanning. The repacking procedure was done to reduce the random error component of the NIR-spectral data. Upon scanning, the quartz cup contained between 111 and 337 larvae, depending on the dietary treatment. Data recording was done every 0.5 nm from 400 to 2500 nm.

Each sample was packed twice, and each packed sample was scanned in duplicate, thus, for each sample we acquired four NIR-spectral scannings. The resulting spectra of these four scannings were averaged. The average NIR-spectra were preprocessed to enhance signal quality and optimize model performance. Multiple spectral preprocessing techniques were applied sequentially to minimize noise and scattering effects using the “prospectr” package (Stevens and Ramirez-Lopez, 2025) in R version (R Core Team, 2024). First, Savitzky-Golay smoothing (first derivative, polynomial order 2, window size 21) was used to reduce noise while preserving spectral features (Savitzky and Golay, 1964). Standard normal variate transformation was then applied to correct baseline shifts and scattering variations (Barnes et al., 1989). Both methods operate independently on each sample by using absorbance values across wavelengths within that sample, do not estimate parameters across samples or use response information, and therefore do not introduce information leakage. Finally, multiplicative scatter correction was implemented to normalize spectral differences due to physical variations in sample presentation. Different combinations of spectral preprocessing were tested for each component, and the best performing models were selected.

Freeze-drying and chemical extraction of amino acids

After the acquisition of the NIR spectra, each sample of live larvae was weighed. The number of larvae in the sample was determined upon visual inspection. Each sample was placed in a small plastic zip-lock bag and moved to a freezer at a temperature of –20 °C. The samples were freeze-dried, and after freeze-drying, each sample was weighted again and then homogenized using an Emerio MC126196 blender for 1–2 min, followed by manual grinding with a mortar and pestle. Approximately 100 mg of freeze-dried material was transferred into 1.5 ml screw-cap tubes, followed by the addition of 1 ml of 60% methanol. Two metal beads were added to each tube, and the samples were homogenized in a Retsch MM 400 mixer mill at 20 Hz for 10 min. The homogenates were centrifuged at $17000\mathrm{×}g$ for 10 min, and the supernatant was collected. This supernatant was diluted 5-fold in deionized water and filtered through 0.22 μm pore-size filter plates to remove particulates.

LC-MS/MS analysis of amino acids

Liquid chromatography coupled with LC-MS/MS analysis was performed at the DynaMo MS-Analytics Facility at the University of Copenhagen (Denmark). A detailed description of the procedure can be found in Violi et al. (2020). In short, prior to analysis, the samples were further diluted 10-fold in deionized water and spiked with internal standards, including a 13C,15N-labeled amino acid mix (Merck Sigma-Aldrich Cell Free Amino Acid Mixture–13C,15N 767964-1EA). Chromatography was performed on a 1290 Infinity II UHPLC system (Agilent Technologies). Separation was achieved on a Zorbax Eclipse XDB-C18 column (150 × 3.0 mm, 1.8 μm, Agilent Technologies). Formic acid (0.05%, v/v) in water and acetonitrile (supplied with 0.05% formic acid, v/v) were employed as mobile phases A and B, respectively. The elution profile r was: 0.00–0.75 min, 3% B; 0.75–4.10 min, 3–65% B; 4.10–4.20 min, 65–100% B; 4.20–4.95 min, 98% B; 4.95–5.00 min, 100–3% B; and 5.0–6.0 min 3% B. The mobile phase flow rate was 400 μl/min. The column temperature was maintained at 40 °C. The liquid chromatography was coupled to an Ultivo Triplequadrupole mass spectrometer (Agilent Technologies) equipped with a Jetstream electrospray ion source operated in positive ion mode. The instrument parameters were optimized by infusion experiments with pure standards. The ion spray voltage was set to +2500 V in positive ion mode. Dry gas temperature was set to 325 °C and dry gas flow to 11 l/min. Sheath gas temperature was set to 350 °C and sheath gas flow to 12 l/min. Nebulizing gas was set to 40 psi. Nitrogen was used as dry gas, nebulizing gas and collision gas. Multiple reaction monitoring was used to monitor precursor ion → fragment ion transitions. Multiple reaction monitoring transitions and other parameters such as fragmentor voltage and collision energies were optimized using reference standards. Both Q1 and Q3 quadrupoles were maintained at unit resolution. Multiple reaction monitoring (MRM) transitions, fragmentor voltage and collision energies are detailed in File S2 in the Supplementary material. Mass Hunter Quantitation Analysis for QQQ software (Version 10, Agilent Technologies) was used for data processing. Quantification of the individual AA was performed from the ¹³C, ¹⁵N-labeled AAs used as internal standards.

PLS regression model development

NIR spectral data were correlated with AA concentrations determined via LC-MS/MS to develop predictive models using partial least squares (PLS) regression implemented in the “pls” package, version 2.8-5 (Liland et al., 2024), in R (R Core Team, 2024). The dataset was split into a training set and an independent test set by randomly selecting 70% of the data for training and 30% for testing. After splitting the data in a training set and test set, the training set was used to train PLS prediction models. During model building, a random tenfold cross-validation was performed three times. The model complexity of the final PLS prediction model was optimized by selecting the number of components that maximized prediction accuracy while minimizing overfitting based on the results from the random tenfold cross-validation. To enhance robustness, model tuning was performed using a randomized grid search approach, selecting the optimal number of PLS components based on the $R^2$ and lowest root RMSE. The final PLS regression model was trained using all training data and then applied to the test set. The model performance of the final PLS model was evaluated using key performance metrics, including the coefficient of determination ($R^2$) for the test set predictions, standard deviation of predictions (SDP), standard error of predictions (SEP), residual predictive deviation (RPD = SDP/SEP), bias and slope (Williams et al., 2017).

To identify the spectral regions most influential for each prediction, a Variable Importance in Projection (VIP) function was implemented using the predictor loadings, scores, and Y-loadings extracted from the PLS regression model. VIP scores were then computed for each wavelength by weighing each predictor’s contribution to the PLS model according to the sum of squared PLS loadings, scaled by the proportion of variance explained in the response variable. This ensures that wavelengths with a higher influence on the model receive higher VIP scores, allowing identification of key spectral regions. Additionally, residual plots were generated from test set predictions to assess model performance. All analyses were conducted in R using the pls package. The PLS loadings for all prediction models are shown in File S3 in the Supplementary material, and the residuals in File S4 in the Supplementary material.

Amino acid profiles in BSF larvae

View Full Size

Average concentrations (mg/g of sample) of amino acids along with the predictive performance of partial least squares regression across the amino acids expressed in units of mg/g of sample and evaluated with different metrics

Citation: Journal of Insects as Food and Feed 2026; 10.1163/23524588-bja10381

• Download Figure
• Download figure as PowerPoint slide

Table 2 presents the descriptive statistics of AA concentrations (mg/g of sample) measured by LC–MS/MS for black soldier fly larvae reared on seventeen different diets. Notable differences in both mean concentrations and variability as indicated by the coefficient of variation (CV) were observed among the AAs. For example, alanine was the most abundant AA (6.50 ± 3.52 mg/g) and had a CV of 54.16%, while asparagine was present at a much lower mean level (0.06 ± 0.08 mg/g) yet exhibited a high CV of 124.7%. In contrast, essential AAs such as lysine and leucine were present at moderate concentrations (1.34 and 1.25 mg/g, respectively) with moderate CVs (61.52 and 56.08%, respectively). The total AA (tAA) had a mean concentration of 34.95 ± 17.29 mg/g with a CV of 49.46%, indicating substantial variability across diets.

PLS regression model performance

The raw spectral NIR data used for the prediction of AA are shown in Figure 2. Table 2 summarizes the predictive performance of the various AA traits, which varied considerably among AAs. An analysis of the relationship between mean AA concentration and model performance ($R^2$) revealed a general trend where traits with higher concentrations tended to exhibit stronger predictive accuracy (Figure 3). For instance, glutamate and alanine, which had relatively high mean concentrations (2.41 mg/g and 6.50 mg/g, respectively), also demonstrated moderate to high $R^2$ values (0.88 and 0.65, respectively). However, certain exceptions were observed; for example, proline, despite its relatively high concentration (5.61 mg/g), exhibited only moderate predictability ($R^2=0.54$).

Overall, the prediction models successfully captured spectral variations for most AAs. The PLS prediction models successfully captured spectral variations for a selection of AAs. Based on an $R^2\mathrm{>}0.6$ and an RPD > 1.5, the best predictions were obtained for alanine, asparagine, glutamate, histidine, and total tAA. The best predictions were for glutamate ($R^2=0.88$, RPD = 2.66) and Asparagine ($R^2=0.78$, RPD = 2.5). Figure 4 shows the calibration and validation plots for the AA with the highest three $R^2$ values, namely asparagine ($R^2=0.78$), glutamate ($R^2=0.88$) and histidine ($R^2=0.68$). The calibration and validation plots for all AAs can be found in Files S5 and S6 in the Supplementary material.

Variable importance and robustness

View Full Size

Raw near infrared spectra of live black soldier fly larvae reared on seventeen different diets and harvested on four consecutive days.

Citation: Journal of Insects as Food and Feed 2026; 10.1163/23524588-bja10381

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Relationship between mean amino acid concentration in black soldier fly larvae and partial least squares regression model performance ($R^2$).

Citation: Journal of Insects as Food and Feed 2026; 10.1163/23524588-bja10381

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Partial least squares regression model calibration (top) and validation (bottom) plots for the prediction of asparagine, glutamate and histidine content of black soldier fly larvae.

Citation: Journal of Insects as Food and Feed 2026; 10.1163/23524588-bja10381

• Download Figure
• Download figure as PowerPoint slide

View Full Size

A sample of annotated variable importance in projection plots for lysine, methionine, leucine and total amino acids (tAA) based on the partial least squares regression models used to predict amino acid content of live black soldier fly larvae.

Citation: Journal of Insects as Food and Feed 2026; 10.1163/23524588-bja10381

• Download Figure
• Download figure as PowerPoint slide

Our VIP analyses performed on each PLS regression model identified consecutive wavelength ranges where the VIP score exceeded a threshold (≥1.0) for each AA. A sample of annotated VIP plots are shown in Figure 5. File S7 in the Supplementary material presents plots of VIP scores across the wavelength spectrum for all studied AAs. In addition, File S8 in the Supplementary material presents an overview of the most prominent intervals, spanning from the visible to near-infrared regions, and indicates the number AAs with elevated VIP scores within these intervals, as well as the maximum VIP score observed.

The highest VIP values were observed in the near-infrared between 1875–1895 nm, where nearly all AAs displayed high model sensitivity. Additional NIR regions around 1725–1745 nm and 1695–1705 nm also showed consistently high VIP values across many AAs. The interval between 2300–2366 nm further displayed moderate importance, being relevant for several AAs though with lower peak VIP values than the dominant 1875–1895 nm window. In the visible range, moderate-to-high importance was observed between 405–445 nm, indicating that features in both the NIR and visible regions were informative for multiple AAs simultaneously. Regions in the first overtone range (∼1375–1395 nm) also contributed to the prediction of certain AAs, albeit with lower maximum VIP scores.

PLSR models to predict AA content

Our study presents the first successful prediction of AA content from NIR spectra of live BSF larvae, marking a novel advancement in insect nutritional analysis. The predictive performance of our PLS regression models was to some extent linked to the concentration and variability of individual AAs. Higher concentration AAs, such as alanine (6.50 mg/g) and glutamate (2.41 mg/g), yielded robust models with relatively high coefficients of determination ($R^2=0.65$ and 0.88, respectively) and moderately good to good ratio of performance to deviation (RPD = 1.70 and 2.66, respectively). This relationship between concentration and model performance aligns with findings in other food matrices, where homogenized samples with higher analyte levels, such as milk AAs (McDermott et al., 2016) and feedstuff fatty acid analyses (Fontaine et al., 2001), tend to produce more reliable NIR predictions.

Only two models had an RPD > 2, namely Asparagine (RPD = 2.50) and Glutamate (RPD = 2.66), whereas thirteen AA had an RPD < 1.5, suggesting poor model performance. The trait tAA exhibited moderate ratio of performance to deviation (RPD = 1.69), but a moderate to high $R^2$ (0.65), moderate slope (0.59), and a strong bias (1.00). These results suggest overfitting or underfitting, meaning that while the model for Glutamine appeared to perform well within the training set, its ability to generalize to new samples was poor. Conversely, Proline, despite its relatively high concentration (5.61 mg/g), showed only moderate predictability ($R^2=0.54$) and a high standard error (SECV = 1.64; SEP = 2.85), indicating spectral complexities or overlapping absorbance patterns that limited accurate AA quantification. These cases underscore the need for a multifaceted evaluation of model performance to distinguish genuinely strong predictions from those that may be misleading due to over- or underfitting.

While our work is novel in its focus on AA prediction from live BSF larvae, research in related areas is rapidly advancing. Recent studies have reported promising attempts to predict detailed fatty acid composition from live BSF larvae using NIR spectroscopy (Alagappan et al., 2025) and hyperspectral imaging (Cruz-Tirado et al., 2025). Furthermore, in live mealworms, NIR has been used to predict moisture content, protein content (Kröncke and Benning, 2022), and AA content (Kröncke et al., 2023). These findings reinforce the growing potential of non-destructive spectral methods to assess complex nutritional profiles in intact insect samples. Larger datasets, advanced chemometric techniques, or hybrid modeling approaches integrating spectral and biological data could further refine prediction accuracy. Collectively, our results and those reported in the broader literature demonstrate that, although predicting the detailed nutritional composition of live larvae presents challenges, NIR spectroscopy remains a promising and rapidly evolving tool for non-destructive nutritional assessment in both traditional food systems and emerging insect production.

VIP and robustness

Our VIP analysis highlighted that the most informative spectral regions for AA prediction in BSF larvae were in the NIR range between 1875–1895 nm, followed closely by bands at 1725–1745 nm and 1695–1705 nm, where a broad set of AAs showed consistently high VIP scores. These intervals correspond primarily to overtone and combination bands of –CH, –NH and –OH functional groups, which are fundamental for the chemical characterization of AAs (Osborne, 2006; Workman and Weyer, 2012). In addition, the interval between 2300–2366 nm also displayed moderate importance for several AAs, aligning with earlier reports on BSF larvae. For example, Cruz-Tirado et al. (2025) found that optimal PLS loadings for stearic acid prediction peaked around 2240 nm, overlapping with our identified 2300–2366 nm region. Similarly, Alagappan et al. (2025) noted high calibration performance in the 870–2530 nm range, implicitly encompassing the bands we identified as critical. The visible region between 405–445 nm, likely reflecting absorptions from aromatic side chains (e.g. tryptophan, tyrosine, phenylalanine), also contributed moderately across multiple AAs. Together, these results confirm that both structural backbone vibrations and side-chain–specific absorptions drive the predictive capacity of our models, and that targeting the 1700–1900 nm and 2300–2366 nm intervals may improve model robustness and enhance rapid, non-destructive nutritional profiling for quality control and insect breeding programs.

Nutritional profile of BSF larvae

Our study represents the first comprehensive analysis of the AA profile in BSF larvae that quantifies 19 individual amino acids using high-resolution LC-MS/MS. In contrast to previous investigations (e.g. Barroso et al., 2017; De Marco et al., 2015; Spranghers et al., 2017), which predominantly focused on the total amino acid composition of processed BSF larvae meals using conventional hydrolysis-based methods, our approach directly captures the underivatized amino acids AA fraction in living larvae (Le et al., 2014; Nakano et al., 2017). This distinction is crucial, as hydrolysis releases both free and protein-bound amino acids while potentially chemically modifying the AAs and introducing degradation artifacts (Nakano et al., 2017; Violi et al., 2020). By contrast, the LC-MS/MS method quantifies only the non-bound underivatized AAs, providing a dynamic and physiologically relevant snapshot of the larvae’s metabolic state.

Future applications

The results presented in this paper as well as other studies that used spectroscopy to predict body composition of living insect larvae (Alagappan et al., 2025; Cruz-Tirado et al., 2025, Kröncke et al., 2022, 2023), offer a promising avenue for rapid quality control and selective breeding in insect production. In selective breeding, it is essential for the selection candidates to survive. Non-destructive will therefore be a valuable tool to selectively breed for improved larval body composition. The method presented in the current study would allow for selection of superior groups of larvae, for example groups of full-sib larvae. Selection based on group performance is a practically more feasible alternative to the time-consuming individual selection strategy (Sandrock et al., 2025). Furthermore, a tool for rapid quality control will be especially helpful for insect producers that use different industrial waste streams as a feed source. Fluctuations in feed composition affect BSF larval body composition (Berggreen et al., 2025), as well as their development time (Laursen et al., 2024; Spranghers et al., 2024) and survival (Spranghers et al., 2024). Using a rapid assessment tool can be used to monitor the larvae’s body composition and select the optimum time to harvest. The results in this study contribute to the development of tools to optimize BSF production, by creating new opportunities for selective breeding as well as management. While this study demonstrates the feasibility of AA prediction from live BSF larvae, future research should focus on further refining these models by incorporating larger datasets, accounting for environmental variability, and expanding their application to additional nutritional markers such as minerals. Advancements in machine learning integration may also enhance predictive performance, ultimately strengthening the role of NIR spectroscopy as a core technology for sustainable and scalable insect protein production.

Limitations

Despite promising results in the very first attempt of predicting detailed AA content in live BSF larvae in this study, several limitations warrant further investigation. First, we analyzed groups of larvae under the implicit assumption of homogeneity, potentially overlooking intra-group variability. This issue is compounded by the lack of controlled genetic structures, as larvae were not grouped by pedigree, which may have increased intra-group variability while minimizing differences between groups. Future studies should consider analyzing individual larvae and controlling the genetic background. Additionally, evaluating alternative chemometric techniques and prediction models, including deep learning, could further enhance predictive accuracy. Furthermore, the experimental diets were deliberately formulated as a controlled gradient, using the same base ingredients while varying the casein and sucrose proportions, in order to introduce sufficient variation for robust NIR model calibration and to enhance generalizability. This design enables stable multivariate calibration by ensuring that variation in both spectral and chemical properties is well represented across the calibration space. Alternative experimental designs based on sets of unrelated diets, for example introducing differences in nutrient availability rather than composition alone due to texture, or consistency effects, could be interesting for future studies to assess broader model generalization.