Motivation: Bringing a new drug to the market is expensive and time-consuming. To cut the costs and time, computer-aided drug design (CADD) approaches have been increasingly included in the drug discovery pipeline. However, despite traditional docking tools show a good conformational space sampling ability, they are still unable to produce accurate binding affinity predictions. This work presents a novel scoring function for molecular docking seamlessly integrated into DockingApp, a user-friendly graphical interface for AutoDock Vina. The proposed function is based on a random forest model and a selection of specific features to overcome the existing limits of Vina’s original scoring mechanism. A novel version of DockingApp, named DockingApp RF, has been developed to host the proposed scoring function and to automatize the rescoring procedure of the output of AutoDock Vina, even to nonexpert users. Results: By coupling intermolecular interaction, solvent accessible surface area features and Vina’s energy terms, DockingApp RF’s new scoring function is able to improve the binding affinity prediction of AutoDock Vina. Furthermore, comparison tests carried out on the CASF-2013 and CASF-2016 datasets demonstrate that DockingApp RF’s performance is comparable to other state-of-the-art machine-learning- and deep-learning-based scoring functions. The new scoring function thus represents a significant advancement in terms of the reliability and effectiveness of docking compared to AutoDock Vina’s scoring function. At the same time, the characteristics that made DockingApp appealing to a wide range of users are retained in this new version and have been complemented with additional features.
The development of algorithms, based on machine learning techniques, supporting (or even replacing) human judgment must take into account concepts such as data bias and fairness. Though scientific literature proposes numerous techniques to detect and evaluate these problems, less attention has been dedicated to methods generating intentionally biased datasets, which could be used by data scientists to develop and validate unbiased and fair decision-making algorithms. To this end, this paper presents a novel method to generate a synthetic dataset, where bias can be modeled by using a probabilistic network exploiting structural equation modeling. The proposed methodology has been validated on a simple dataset to highlight the impact of tuning parameters on bias and fairness, as well as on a more realistic example based on a loan approval status dataset. In particular, this methodology requires a limited number of parameters compared to other techniques for generating datasets with a controlled amount of bias and fairness.
24 March 2025 in Neural Computing and Applications
Knowledge distillation is recognized as a valuable model compression strategy that alleviates the computational burden of large language models while preserving performance. This strategy involves training a smaller model utilizing both real data and predictions from a more cumbersome model. Traditional distillation methods, however, are often compromised by exposure bias, which results from reliance on next-step prediction training loss. This bias emerges when models are tested in free-running mode, differing from their training regime and leading to a progressive drift in input distributions between testing and training phases. An analogous issue, known as ‘distributional shift’, has been effectively addressed in imitation learning through various methodologies. Therefore, this paper specifically tailors an imitation learning-based solution to a traditional knowledge distillation framework which inherently considers both real data and the teacher’s predictions as dual sources of expert demonstrations. The effectiveness of this approach is demonstrated over five different test datasets, where it outperforms traditional benchmarks across all evaluation metrics. Specifically, it achieves superior results in perplexity, multi-token generation, and G-Eval score, indicating improvements in both predictive accuracy and alignment with human judgment in text quality. These results underscore the potential of this approach to effectively address exposure bias in large language model distillation.