How Can Artificial Intelligence Distinguish Cow, Goat, and Sheep Milk?

Milk is a worldwide staple, yet figuring out its animal source can be tricky. Researchers combined two powerful tools: proteomics, the study of proteins, and artificial neural networks, a form of AI. They examined cow, goat, and sheep milk samples with MALDI‑TOF mass spectrometry, which creates a mass “fingerprint” for each protein. From these fingerprints they extracted thousands of eight‑dimensional vectors, tiny data packets that capture the most important features. Using this data, they trained a neural network to recognize each species’ unique signature. The outcome? A model that identifies milk with very high accuracy, even when differences are subtle. This approach can fight food fraud, ensure dairy quality, and protect consumers. Moreover, the same method could be applied to other foods, making supermarket shelves safer. In practice, science is reshaping how we verify food authenticity, leveraging smart algorithms and advanced protein analysis.

MALDI‑TOF (Matrix‑Assisted Laser Desorption/Ionization‑Time of Flight) mass spectrometry produces spectra within seconds by ionizing milk proteins without destroying them. Each peak corresponds to a specific mass, typical of a protein or peptide. For the three milk types, researchers collected 648, 1,554 and 2,392 spectra, translated into 16,018, 38,394 and 55,055 eight‑dimensional vectors. The eight dimensions represent normalized intensities of the most discriminative peaks, selected via variance analysis and feature‑selection methods. The neural network employed is a feed‑forward architecture with three hidden layers, ReLU activation functions, and 20% dropout to prevent overfitting. The dataset was split into 70% training, 15% validation and 15% testing, ensuring balanced class distribution. During training, the Adam optimizer minimized a cross‑entropy loss, converging within 50 epochs. Test‑set accuracy reached 98.7%, with recall above 97% for each species. Confusion analysis revealed few errors between goat and sheep, due to shared proteins. This demonstrates that high‑resolution proteomics combined with deep learning can surpass traditional single‑marker methods, providing a scalable solution for food authentication. Future work includes integrating metabolomic data and using convolutional neural networks to analyze raw spectra, further enhancing model robustness.