When studying non-symmetric nonlocal operators $$ {\cal L} f(x) = \int_{{\bf R}^d} \left( f(x+z)-f(x)-\nabla f(x)\cdot z 1_{\{|z|\leq 1\}} \right) \frac{κ(x, z)}{|z|^{d+α}} d z , $$ where $0<α<2$ and $κ(x, z)$ is a function on ${\bf R}^d\times {\bf R}^d$ that is bounded between two positive constants, it is customary to assume that $κ(x, z)$ is symmetric in $z$. In this paper, we study heat kernel of ${\cal L}$ and derive its two-sided sharp bounds without the symmetric assumption $κ(x,z)=κ(x,-z)$. In fact, we allow the kernel $κ$ to be time-dependent and also derive gradient estimate when $β\in(0\vee (1-α),1)$ as well as fractional derivative estimate of order $θ\in(0,(α+β)\wedge 2)$ for the heat kernel, where $β$ is the Hölder index of $x\mapstoκ(x,z)$. Moreover, when $α\in(1,2)$, the drift perturbation with drift in Kato's class is also considered. As an application, when $κ(x,z)=κ(z)$ does not depend on $x$, we show the boundedness of nonlocal Riesz's transorfmation: for any $p>2d/(d+2α)$, $$ \| {\cal L}^{1/2}f\|_p\asymp \|Γ(f)^{1/2}\|_p, $$ where $Γ(f):=\frac{1}{2}{\cal L} (f^2)-f {\cal L} f$ is the carré du champ operator associated with ${\cal L}$, and ${\cal L}^{1/2}$ is the square root operator of ${\cal L}$ defined by using Bochner's subordination. Here $\asymp$ means that both sides are comparable up to a constant multiple.