In public key encryption from LWE, we do the following steps
$\textbf{PKE.KeyGen($1^n$)}$ takes as input the security parameter n, samples $A \leftarrow \mathbb{Z}_p^{n \times m}$ and $\textbf{e} \leftarrow \chi^m$, and outputs a key-pair $(pk,sk)$ where $sk = s \in \mathbb{Z}^n_q$ and $pk = (A,s^T\cdot A + \textbf{e}^T)$.
$\textbf{PKE.Enc($pk = (A,b^T),b$)}$ takes as input a public key $(A,b^T)$ and a message $a \in {\{0,1}\}$, samples a short vector $r \leftarrow {\{0,1}\}^m$, and outputs a ciphertext $(A\cdot r,b^T\cdot r + a \cdot \lfloor\frac{q}{2}\rfloor)$.
If we now assume that the secret $s$ is a matrix instead of a vector, how can we deterministically encode the message $a$ so that we can recover it during decryption? The error term will also be a matrix in this case where each column is chosen independently of the other from some error distribution.