This preprint reviews and synthesizes how modern machine learning is reshaping the drug discovery and clinical development pipeline, addressing two core bottlenecks in pharma R&D: extreme cost (often cited at multi-billion USD per approved drug) and high clinical trial failure rates. The paper surveys practical ML approaches across target identification, hit discovery, lead optimization, ADMET/toxicity prediction, de novo molecular design, and clinical trial risk modeling, highlighting where specific model families fit best, including graph neural networks for molecular property prediction and transformer-based architectures for molecule generation and sequence-driven tasks. A comparative evaluation is presented with reported gains such as improved hit identification performance, faster lead optimization cycles, stronger prediction of mid-stage (Phase II) trial failures, and robust toxicity prediction (e.g., AUC > 0.85) alongside generation of novel compounds with high synthetic accessibility. The manuscript also discusses real-world limitations: data quality, bias, interpretability, privacy/proprietary constraints, and regulatory acceptance, while outlining near-term and longer-term integration directions (e.g., federated learning, digital twins, automated labs, and quantum ML).