Machine learning-aided design of composite mycotoxin detoxifier material for animal feed

The development of food and feed additives involves the design of materials with specific properties that enable the desired function while minimizing the adverse effects related with their interference with the concurrent complex biochemistry of the living organisms. Often, the development process is heavily dependent on costly and time-consuming in vitro and in vivo experiments. Herein, we present an approach to design clay-based composite materials for mycotoxin removal from animal feed. The approach can accommodate various material compositions and different toxin molecules. With application of machine learning trained on in vitro results of mycotoxin adsorption–desorption in the gastrointestinal tract, we have searched the space of possible composite material compositions to identify formulations with high removal capacity and gaining insights into their mode of action. An in vivo toxicokinetic study, based on the detection of biomarkers for mycotoxin-exposure in broilers, validated our findings by observing a significant reduction in systemic exposure to the challenging to be removed mycotoxin, i.e., deoxynivalenol (DON), when the optimal detoxifier is administrated to the animals. A mean reduction of 32% in the area under the plasma concentration–time curve of DON-sulphate was seen in the DON + detoxifier group compared to the DON group (P = 0.010).

. Performance ranking of sepiolite, Na-montmorillonite and Mg-rich smectite in terms of adsorption and efficiency. The experimental settings for the uptake of DON, OTA, T2, FUM and ZEN were fixed to 1kg/t of inclusion rate of MDT, 3 µg/ml of toxin concentration. The pH during the adsorption experiment was fixed to 3 while the desorption pH was 6.5. AFB1 uptake was assessed with 2kg/t of inclusion rate of MDT, 4 µg/ml of toxin concentration. The pH during both adsorption and desorption experiments was fixed to 5.  Figure S6. RF-based synergy-capturing assessment of efficiency in a series of hybrid SEP/MONT materials. The red area is assigned to the positive synergistic effect. The experimental settings for the uptake of DON, OTA, T2, FB1 and ZEN were fixed to 2 kg/t of inclusion rate of MDT, 2 µg/ml of toxin concentration. The pH during the adsorption experiment was fixed to 3 while the desorption pH was 6.5. The main values and the MAE are given. Figure S7. Efficiency prediction of a list of regulated and yet un-regulated mycotoxins. In orange the group of trichothecenes, aflatoxins in blue, fumonisins in purple, zearalenone in green and ochratoxins in magenta. The experimental setting was fixed to 2kg/t of inclusion rate of MDT, 2 µg/ml of toxin concentration. The pH during the adsorption experiment was fixed to 3 while the desorption pH was 6.5. Figure S8. Graphical assessment of support vector machine, multiple layer perceptron and knearest neighbors predicting Ads(%) and Eff(%).

S1. Experimental material characterization
Specific surface area was estimated by Brunauer-Emmett-Teller (BET) theory on the adsorption branch of the physisorption isotherms. The isotherms were measured in a Micromeritics Gemini V surface area and pore size analyzer under N2 gas flow at 77K. The samples were previously degassed at 124°C for 18h. The pH of a 10 wt% suspension of raw clay in distilled water was measured with Crison Basic 20 pH-meter, after stirring for 10 minutes at 25°C. The cation exchange capacity (CEC) was estimated quantifying the concentration of colored Cu 2+ -triethylenetetramine complex by spectrophotometry. The exchange solution (1 l) was prepared dissolving in deionized water 1. The starting amount of mycotoxin (toxin concentration) and binder (inclusion rate) used for filling the databases, as well as the adsorption and desorption pH are selected among the ranges shown in Table S5.

S2. Non-tree models architecture
Support vector machine and k-nearest neighbors were implemented using Scikit-learn library. Support vector machine was optimized on the base of k-fold cross validation finding the best hyperparameter, i.e., kernel='poly', gamma='scale', degree=3. Multiple layer perceptron (MLP) was implemented in TensorFlow and Keras framework. The network was built with 6 dense layers with ReLU activation function, 100 neurons in the hidden layers, Adaptive Moment Estimation (Adam) as optimizer and 300 epochs. The performances were evaluated by R 2 score, mean absolute value (MAE), p-value and accuracy score. The latter is added to classify how often our regressors make predictions closed to the real value by considering the estimated experimental error as tolerance grade. The accuracy is defined as ratio of number of correct prediction and the total predictions corresponding to the testset size. The correct prediction was defined considering the estimated experimental error (5.7) and obtained when the absolute value of the difference between the predicted and real target is less the 5.7.

S3. In vivo trials.
An in vivo trial was conducted with 8 healthy broiler chickens, females and males equally divided, of the same breed (Ross 308, commercial supplier, Aye, Belgium), the same age (4 weeks at the beginning of the treatment) and about the same body weight (BW) at arrival. The in vivo study was conducted at CER-Groupe, a GLP (Good Laboratory Practice) compliant test site (Marloie, Belgium). The animal study was approved by the Ethical Committee of CER-Groupe. Randomization was performed at arrival, based on the sex and BW of the broiler chickens in such a way that 2 groups with 4 birds (2 males and 2 females) each were formed with about the same average BW/group. Both groups were housed in the same pen of 3 m². The housing conditions were according to the EU guidelines and the Belgian guidelines 1,2 . The light schedule was a 18h/6h light/dark scheme. An acclimatization period of 8 days was respected (day 1-8). Throughout the acclimatization period, the chickens were allowed access to feed and water ad libitum. The chickens received commercial broiler chicken feed ACTI POUSSIN (batch 028590, SCAR Büllingen, Belgium). At least eight hours before the treatment (day 8), the feed was withdrawn, but water was available. Animals were fed again 4 h post administration. The feed was analysed by a multi-mycotoxin LC-MS/MS method (liquid chromatography-tandem mass spectrometry) and was found to contain low levels of DON (61.3 µg/kg) and OTA (1.6 µg/kg). These contamination levels were well within the acceptance criteria of the EU (2006/576/EC) 3 . Analytical standard of DON was obtained from Fermentek (Jerusalem, Israel), and a stock solution was prepared in ethanol at a concentration of 10 mg/ml.
The treatment consisted of a single oral bolus administration with either DON or DON in combination with SEP/MONT/AC detoxifier (0.500 mg DON/kg BW, corresponding to the maximum EU guidance level of 5 mg/kg DON in feed, and 0.4 g detoxifier/kg BW, corresponding to an inclusion rate of 0.4% in the feed), administered as oral capsules directly in the crop and using a cross-over study design respecting a one-day wash-out period between treatments. The capsules contained cellulose and the appropriate volume of DON stock solution was added, whether or not combined with the detoxifier. Capsules were closed after 20 min in order to ensure evaporation of the solvent.
Repetitive blood samples (+/-0.5 ml) were taken from the vena metatarsalis plantaris superficialis (leg vein). Intravenous (IV) catheters were placed in order to ensure continuous access to the blood vein. The time points of blood sampling were 0 h (just before administration) and 0.08, 0.25, 0.5, 0.75, 1, 1.5, 2, 4, and 8 h (post administration, p.a.). The blood samples were centrifuged within 2 hours after collection (±3,000 g, 10 min, 4°C First, a Shapiro-Wilk test for normality was performed, including inspection of QQ-plots. If the data were not normally distributed, a log-transformation of the data was performed. If OK, a paired sample t-test was performed. Data for elimination half-life (T1/2e) were not normally distributed, even after log-transformation. Therefore, a nonparametric related-samples Wilcoxon signed rank test was performed. Moreover, the relative oral bioavailability ((AUC0-¥ mycotoxin+detoxifier/ AUC0-¥ mycotoxin)*100) was evaluated as marker for efficacy of the detoxifier. In general, two treatments are considered bioequivalent or thus not different from one another if the 90% confidence interval (CI) of the ratio of a log-transformed exposure measure (AUC) falls completely within the range 80-125%, as it is assumed that differences in exposure up to 20% are not relevant. If the CI falls completely out this specified range the treatments are considered not bioequivalent, and hence a significant effect of the detoxifier in reduction of systemic exposure can be concluded.