One of the main difficulties to enable the future scaling of IoT networks is the issue of massive connectivity. Recently, Modern Random Access protocols have emerged as a promising solution to provide massive connections for IoT. One main protocol of this family is Irregular Repetition Slotted Aloha (IRSA), which can asymptotically reach the optimal throughput of 1 packet/slot. Despite this, the problem is not yet solved due to lower throughput in non-asymptotic cases with smaller frame sizes. In this paper, we propose a new variant of IRSA protocol named Deep-Learning and Sensing-based IRSA (DS-IRSA) to optimise the performance of IRSA in short frame IoTs, where a sensing phase is added before the transmission phase and users' actions in both phases are managed by a deep reinforcement learning (DRL) method. Our goal is to learn to interact and ultimately to learn a sensing protocol entirely through Deep Learning. In this way, active users can coordinate well with each other and the throughput of the whole system can be well improved. Simulation results show that our proposed scheme convergence quickly towards the optimal performance of almost 1 packet/slot for small frame sizes and with enough minislots and can achieve higher throughput in almost all cases.