Signature codes are widely used for user identification (UI) and channel estimation (CE) in wireless networks due to their high spectral efficiency. To design an effective signature matrix exploiting prior information, we propose an end-to-end machine-learning-aided signature code (ML-SC) system. The ML-SC system comprises an encoder based on binarized neural networks (BNNs) and a decoder based on a modified trainable iterative soft-thresholding algorithm (TISTA). The BNNs efficiently handle the trainable discrete signature matrix, while the modified TISTA enables support for a large number of users with enhanced performance under Rayleigh fading channels. Our simulation results demonstrate that the proposed ML-SC system maintains scalability across different matrix dimensions. With this scalability advantage, the ML-SC achieves consistent performance improvements. The signature matrix generated by the ML-SC system, which we refer to as the ML-signature matrix, yields superior decoding performance compared to randomly generated and deterministic binary matrices, demonstrating an effective SNR gain of approximately 2.5–5 dB compared to conventional approaches. We also verify that the proposed ML-signature matrix demonstrates strong compatibility with various prevalent decoding methods. Furthermore, we confirm significant enhancement of the restricted isometry constant (RIC) for the ML-signature matrix, which provides theoretical support for the observed performance improvements. Copyright © 2026 The Institute of Electronics, Information and Communication Engineers.