The MD5 pre-image resistance theoretically is broken bu not practically. Instead of the generic preimage attack with 2128 complexity, there is an attack that requires 2123.4 complexity. MD5, however, is not practically broken. You cannot reach the 123-bit search space in a meaningful time. Consider the collective power of BitCoin miners; in 2020/1/17 they reached 126.1314 Exahashes per second, that makes
- 267 SHA256 double hashes in a second, and
- 292 SHA256 double hashes in a year. So you need 221 years with that power.
What you asking is against the design principle of the hash functions; the avalanche criteria. That is changing one bit of input will flip approximately half of the output bits, randomly.
If you can find even a single bit better than the generic pre-image attack, then you have a theoretical weakness. you can use this attack like divide and conquer technique to reduce the overall pre-image attack complexity.