Biases are needed for the dark scalation routine.
Bias is a fixed pedestal, that accounts for the electronic signal (or "noise"). Since it is very close to zero signal, it is dominated by readout noise. So, to get the true underlying signal, we need a large amount of biases. It is true that they are included in darks, but you'll end with a much better electronic characterization using several dozens of bias, which are cheap to do. Having dozens of darks is impractical in most cases.
Darks, on the other hand, measures the thermal current, the signal that is generated only by thermical effects. It also suffers from noise, but it is more farther than the readout noise. Darks are not fixed. The dark current depends on the exposure time (mostly linear) and from the temperature (T^4). Also, there are some weird pixels that do not follow the linearity rule, and have either too large darks current, or too low; these are the hot and cold pixels. In many cases, there are defective pixels, where no data is recoverable.
If the dark current do not precisely match the lights, one may rescale them. But, to do so, we need to subtract the fixed pattern measured by the bias.
Also, bias are needed to apply the flat frames. They also include the bias factor, and since flats are multiplicative effects, we need to apply that calibration to get proper master flats.
To sumarize, the main equation for calibration is: (lights - bias - k*(darks-bias))/(flats-bias)
Do not try to take shortcuts. We need the three calibration frames to get the most of our light frames.
