Titulo Estágio
Generative Adversarial Network for Drug Design
Áreas de especialidade
Sistemas Inteligentes
Local do Estágio
Laboratory of Artificial Neural Networks (LARN)
Enquadramento
In the recent century, the impact that molecules such as penicillin had in human society is only comparable to the monumental effort (in terms of either time, money or luck) that their finding demanded. Even nowadays, a new material requires an estimated time of 20 years before it is ready to be commercialized. This fact has inspired molecular discovery. For molecular discovery, we need to navigate chemical space, the set of all possible compounds. Navigating such a space is hard; the space is incredibly large and discrete, small changes in a molecule can change its properties radically. Nowadays, the feasibility of inverse design has been accelerated partly due to new developments from the artificial intelligence community: Generative Adversarial Networks(GAN), text sequence generation models and Reinforcement Learning (RL).
This proposal proposes the use of state-of-the-art Deep Learning methods to develop a computational model that accurately generates novel molecules that not only have the predicted activity against a target but also include multiple pharmacological objectives. This model presents an Objective-Reinforced Generative Adversarial Networks for inverse-design chemistry that extends the conventional single-objective reinforcement learning (RL) methods. Recent applications of RL for tuning generative methods have presented promising results. However, in tasks like molecules generation, it is wanted to optimize some different properties at the same time (e. g. diversity, drug-likeness, and synthesizability).
The model is composed of three key elements: a Generator G, a Discriminator D and a Reinforcement component R. On one side, the discriminative network determine whether a molecule is likely to come from the initial distribution (positive) or not (negative). The reinforcement provides a quality metric that will quantify the desirability of a given molecule. Finally, the generator has the task of generating molecules that maximizes the objective function, which is a linear combination of the discriminator component and the reinforcement component, parametrized by a tunable parameter. In this way, the adversarial approach is meant to keep the generation of molecules similar to the initial distribution of data, while the reinforcement learning biases this generation towards the maximization of the reward.
Objetivo
The main objective of this proposal is to develop a deep generative method of objective-reinforced Generative Adversarial Networks which can perform molecular generation biased towards some desired chemical properties. The main goals of this proposal are:
1. Construct a data set for the descriminator model;
2. Construct a data set for the GAN model;
3. Perform Data Pre-Processing, Normalization and Scaling;
4. Select appropriate ML algorithms for building the generative and
discriminative model;
5. Select appropriate ML algorithms for building the reinforcement model;
6. Perform sampling and model evaluation and validate the overall model
with real data sets.
7. Integrate the implemented components reusable platform.
Plano de Trabalhos - Semestre 1
1st semester
• Overview of drug discovery, including target identification, lead
discovery, and lead optimization;
• Overview of machine learning techniques, namely Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU) and Generative Adversarial Networks(GANs);
• Propose initial deep generative model combined with reinforcement
learning workflow and prepare the first case study
• Prepare the intermediate report.
Plano de Trabalhos - Semestre 2
2nd semester
• Select, and pre-process a collection of large datasets for experiments;
• Study, and select, Machine Learning (ML) algorithms for building the
generative model to create valid novel drug molecules;
• Study and select feature selection algorithms and objective reinforcement
learning model machine for generating chemically feasible drug molecule
• Analyse experimental results: e.g., study parameter values; compare
performance of the reduced datasets vs. previous results, etc.;
• Prepare a research paper and the final version of the thesis.
Condições
This work will be carried out in the Laboratory of Neural Networks (LARN) of CISUC, where there will be regular supervision and feedback on behalf of the supervisors. Familiarity with machine learning and data mining algorithms and software tools are essential. Participating students will acquire valuable knowledge and experience with model building and data science by mining massive datasets, which skills are currently in high demand for various technology employers due to the relevance to various applications
Orientador
Maryam Abbasi, Joel P. Arrais, Bernardete Ribeiro
maryam@dei.uc.pt 📩