Molecular docking, scoring, and virtual screening play an increasingly important role in computer-aided drug discovery. Scoring functions (SFs) are typically employed to predict the binding conformation (docking task), binding affinity (scoring task), and binary activity level (screening task) of ligands against a critical protein target in a disease's pathway. In most molecular docking software packages available today, a generic binding affinity-based (BA-based) SF is invoked for all three tasks to solve three different, but related, prediction problems. The limited predictive accuracies of such SFs in these three tasks has been a major roadblock toward cost-effective drug discovery. Therefore, in this work, we develop BT-Score, an ensemble machine-learning (ML) SF of boosted decision trees and thousands of predictive descriptors to estimate BA. BT-Score reproduced BA of out-of-sample test complexes with correlation of 0.825. Even with this high accuracy in the scoring task, we demonstrate that the docking and screening performance of BT-Score and other BA-based SFs is far from ideal. This has motivated us to build two task-specific ML SFs for the docking and screening problems. We propose BT-Dock, a boosted-tree ensemble model trained on a large number of native and computer-generated ligand conformations and optimized to predict binding poses explicitly. This model has shown an average improvement of 25% over its BA-based counterparts in different ligand pose prediction scenarios. Similar improvement has also been obtained by our screening-based SF, BT-Screen, which directly models the ligand activity labeling task as a classification problem. BT-Screen is trained on thousands of active and inactive protein-ligand complexes to optimize it for finding real actives from databases of ligands not seen in its training set. In addition to the three task-specific SFs, we propose a novel multi-task deep neural network (MT-Net) that is trained on data from the three tasks to simultaneously predict binding poses, affinities, and activity levels. We show that the performance of MT-Net is superior to conventional SFs and on a par with or better than models based on single-task neural networks.