Selenophosphate synthetase 1 is an essential protein with roles in regulation of redox homoeostasis in mammals.

Selenophosphate synthetase (SPS) was initially detected in bacteria and was shown to synthesize selenophosphate, the active selenium donor. However, mammals have two SPS paralogues, which are designated SPS1 and SPS2. Although it is known that SPS2 catalyses the synthesis of selenophosphate, the function of SPS1 remains largely unclear. To examine the role of SPS1 in mammals, we generated a Sps1-knockout mouse and found that systemic SPS1 deficiency led to embryos that were clearly underdeveloped by embryonic day (E)8.5 and virtually resorbed by E14.5. The knockout of Sps1 in the liver preserved viability, but significantly affected the expression of a large number of mRNAs involved in cancer, embryonic development and the glutathione system. Particularly notable was the extreme deficiency of glutaredoxin 1 (GLRX1) and glutathione transferase Omega 1 (GSTO1). To assess these phenotypes at the cellular level, we targeted the removal of SPS1 in F9 cells, a mouse embryonal carcinoma (EC) cell line, which affected the glutathione system proteins and accordingly led to the accumulation of hydrogen peroxide in the cell. Furthermore, we found that several malignant characteristics of SPS1-deficient F9 cells were reversed, suggesting that SPS1 played a role in supporting and/or sustaining cancer. In addition, the overexpression of mouse or human GLRX1 led to a reversal of observed increases in reactive oxygen species (ROS) in the F9 SPS1/GLRX1-deficient cells and resulted in levels that were similar to those in F9 SPS1-sufficient cells. The results suggested that SPS1 is an essential mammalian enzyme with roles in regulating redox homoeostasis and controlling cell growth.

Mechanism for feature learning in neural networks and backpropagation-free machine learning models.

Understanding how neural networks learn features, or relevant patterns in data, for prediction is necessary for their reliable use in technological and scientific applications. In this work, we presented a unifying mathematical mechanism, known as average gradient outer product (AGOP), that characterized feature learning in neural networks. We provided empirical evidence that AGOP captured features learned by various neural network architectures, including transformer-based language models, convolutional networks, multilayer perceptrons, and recurrent neural networks. Moreover, we demonstrated that AGOP, which is backpropagation-free, enabled feature learning in machine learning models, such as kernel machines, that a priori could not identify task-specific features. Overall, we established a fundamental mechanism that captured feature learning in neural networks and enabled feature learning in general machine learning models.