Thank you all for your suggestions.
1. Use diode.
As an interesting idea, the temperature coefficient of a silicon diode is 2mV / C and it would then not be measured, but divination from the coffee grounds.
2. Use differential amplifier.
I'll think about it, I have 1/2 buffer MCP6022 there and I'll consider how many errors would its connection insert as differential amplifier.
500mV is approximately 11% of range. I have to consider how big the error inserts 4pcs 1% resistors in the amplifier + leaning the reference voltage on the lousy amplifier in STM32F303.
What is really going on?
I have 150PSI Air pressure sensor 500mV - 4,5V, for calibration have BME280 and calibration point on atmospheric pressure.
Yes, it's not worth solving.