First Attempt - MCC .6

The first trained model with a high accuracy/recall. Low loss and relatively okay at identifying good proteins.

Positives:

• Detecting "Evolutionary Deviance": It knows exactly what a healthy human p53 or Insulin sequence looks like. If you change a single letter to a cancer mutation, it spots the "smell" of a non-functional sequence.
  -Catching "Biological Gibberish": It easily identifies frame-shifts or random noise because they don't follow the "grammar" of the human proteins it studied.

Limitations:

• Non-Human/Ancient Sequences: If you feed it a perfectly functional protein from a deep-sea vent or an ancient bacteria, it might flag it as "Toxic" simply because it doesn't look "Normal" compared to the human data.
• Fake-but-Stable (Hallucinations): As you saw with the hydrophobic/repetitive tests, if a "fake" sequence mimics the patterns of a real protein (even if it's biologically useless), the model can be tricked into giving it a high score.