Phase I trials are the cornerstone of cancer drug development, and the goal of phase I dose-finding studies is to determine a safe and possibly effective dose of a new treatment. With traditional cytotoxic treatments in oncology, the primary endpoint has been the dose-limiting toxicity (DLT), which, with efficacy, is assumed to increase with the drug dose. A phase I trial seeks the highest dose with acceptable risk of DLT, called the maximum tolerated dose (MTD). The 3+3 design has remained the most popular design in practice for the past 30 years, despite that many model-based methods such as the continual reassessment method (CRM) offer clear advantages in terms of identifying the correct MTD. With recent advances in immunotherapy and molecular targeted therapy, the chances of observing the DLT became rare and infrequent, which raises critical challenges in dose-finding design. In this setting, a phase I trial must account for the cumulative effect of both severe toxicities, DLT, and mild to moderate toxicities to better capture the toxicity profile of new treatments. This dissertation presents three novel approaches to address these challenges in phase I trials in oncology. First, we propose a rule-based design called i3+3, suitable for standard single-agent phase I settings. The i3+3 design is based on flexible but straightforward rules that account for the variabilities in the observed data. In short, the i3+3 design asks clinicians to compare observed toxicity rates with a prespecified toxicity interval and make dose-escalation decisions according to simple rules. The i3+3 design possesses both the simplicity and transparency of rule-based approaches, and the superior operating characteristics seen in model-based approaches. Second, we consider a Bayesian framework based on “probability of decision” (POD) for dose-finding trial designs of MTAs, called POD-BIN. The two key features of the proposed POD-BIN design are: (\textit{i}) the posterior predictive probabilities of dose escalation decisions are considered and (\textit{ii}) time-to-toxicity for both mild toxicity (MT) and dose-limiting toxicity (DLT) are modeled simultaneously. This allows the possibility of enrolling new patients when previously enrolled patients are still being followed for toxicity, thus potentially shortening the trial length. The Bayesian decision rules in POD-BIN utilize the PODs to balance the trade-off between the need to speed up the trial and the risk of exposing patients to overly toxic doses. POD-BIN appears to be able to control the frequency of making risky decisions and, at the same time, shorten the trial duration in the simulation. Third, we develop another adaptive design under the Bayesian framework, called the CLAST design, incorporating both longitudinal counts of MTs and time-to-DLTs per patient and allowing the rates of both responses to vary flexibly with time. We adopt the Bayesian joint modeling approach, which captures the within-patient correlation between the two outcomes via an association structure. We find that using MT counts in addition to DLT observations within the first cycle increases the probability of correctly identifying the MTD, particularly when the MTD is not among the highest dose levels being considered.