What machine learning is good at here
Machine learning in drug design is QSAR's powerful descendant: it learns patterns from large datasets using flexible models — random forests, gradient boosting, neural networks, graph models that read the molecule directly. Built on the same cheminformatics representations (descriptors, fingerprints, molecular graphs), ML can capture relationships too tangled for a linear equation. Its most reliable wins are in property prediction.
ADMET prediction is the everyday workhorse: models that flag likely solubility, permeability, metabolic stability, or hERG liability before a molecule is ever made. These predictions are imperfect but cheap, so they let you filter thousands of ideas to a sane shortlist. The discipline is honest book-keeping: track a model's accuracy on held-out compounds, respect its applicability domain, and never let a confident-looking number override clear experimental data.
Filling in missing structures, inventing new ones
Two ML-adjacent tools extend structure-based design. When no experimental structure of your target exists, a homology model (or now a deep-learning structure prediction) builds an approximate 3D model from a related protein's known structure — good enough for hypothesis-making if a close template exists, unreliable if not. And binding site prediction scans a surface to suggest where a pocket — even an allosteric one — might sit, pointing you at druggable cavities you might otherwise miss.
The most eye-catching frontier is generative chemistry and de novo design: models that propose entirely new molecules optimized toward a target profile, rather than only scoring molecules you supply. They can explore vast chemical space and suggest non-obvious scaffolds. But they must be constrained, or they invent things that are unstable, unsynthesizable, or absurd. Pairing a generator with hard filters — synthesizability, predicted ADMET, docking — is what turns a clever idea machine into a useful one.
Folding it all back into the cycle
No single tool in this track designs a drug. They earn their keep by accelerating the design–make–test cycle: each iteration generates ideas, predicts their behavior, makes the most promising, measures the truth, and feeds that truth back to sharpen the next round of models. Computation's real role is to make each loop cheaper and smarter — testing ten ideas in silico for every one you commit to the bench.
And the goal is never potency alone. Real candidates must satisfy multiparameter optimization — balancing affinity, selectivity, solubility, permeability, metabolic stability, and safety at once. The whole computational toolkit you have met — docking, QSAR, MD, FEP, ML, ADMET prediction — exists to navigate that many-dimensional trade-off faster than synthesis-and-test alone ever could.