Abstract:Molecule optimization is an important problem in chemical discovery and has been approached using many techniques, including generative modeling, reinforcement learning, genetic algorithms, and much more. Recent work has also applied zeroth-order (ZO) optimization, a subset of gradient-free optimization that solves problems similarly to gradient-based methods, for optimizing latent vector representations from an autoencoder. In this paper, we study the effectiveness of various ZO optimization methods for optimizing molecular objectives, which are characterized by variable smoothness, infrequent optima, and other challenges. We provide insights on the robustness of various ZO optimizers in this setting, show the advantages of ZO sign-based gradient descent (ZO-signGD), discuss how ZO optimization can be used practically in realistic discovery tasks, and demonstrate the potential effectiveness of ZO optimization methods on widely used benchmark tasks from the Guacamol suite. Code is available at: https://github.com/IBM/QMO-bench.

What problem does this paper attempt to address?

This paper attempts to solve several key problems in molecular optimization, which are crucial in chemical discovery. Specifically, the paper focuses on how to use Zeroth - Order (ZO) optimization methods to optimize molecular targets, which usually have varying smoothness, sparse optimal values and other challenges. The main contribution of the paper lies in evaluating the effectiveness of different ZO optimization methods on molecular optimization problems, especially in terms of convergence speed, convergence accuracy and robustness to unfriendly landscapes of molecular objective functions. The following are the main research points of the paper: 1. **Evaluating the effectiveness of ZO optimization methods**: The paper experimentally compares several different ZO optimization methods, such as ZO - Gradient Descent (ZO - GD), Sign - based ZO - Gradient Descent (ZO - signGD) and ZO - Adaptive Momentum method (ZO - Adam), to evaluate their performance in molecular optimization tasks. 2. **Characteristics of molecular objective functions**: The paper explores the special properties of objective functions in molecular optimization problems, such as activity cliffs, sparse optimal regions and discrete jumps, which make molecular optimization very difficult. 3. **Advantages of Zeroth - Order optimization methods**: The research shows that ZO - signGD is not only superior to other methods in convergence speed, but also shows higher robustness when dealing with unfriendly landscapes of molecular objective functions. Although ZO - signGD may have lower convergence accuracy in other problems (such as adversarial attacks), it shows higher effectiveness and accuracy in molecular optimization. 4. **Practical applications and hybrid methods**: The paper discusses how to apply ZO optimization methods in actual drug discovery scenarios, including the identification of new lead molecules and the optimization of known lead molecules. In addition, a hybrid method that combines ZO optimization methods with other generative models is proposed to improve search efficiency and the ability to meet multiple design constraints. 5. **Experimental verification**: In order to verify the effectiveness of the above methods, the paper conducts experiments on three tasks in the Guacamol benchmark test suite, including perindopril_mpo, zaleplon_mpo and deco_hop. The experimental results show that ZO - signGD performs best in these tasks, especially when dealing with difficult objective function landscapes. In conclusion, through systematic experiments and analysis, this paper shows the potential of ZO optimization methods in molecular optimization, especially the superior performance of ZO - signGD when dealing with complex molecular objective functions. This provides new tools and methods for future drug discovery.

An Empirical Evaluation of Zeroth-Order Optimization Methods on AI-driven Molecule Optimization

Optimizing molecules using efficient queries from property evaluations

Structure-Based Molecule Optimization via Gradient-Guided Bayesian Update

Meta Learning for Low-Resource Molecular Optimization

Optimization of Molecules via Deep Reinforcement Learning

DeepZero: Scaling up Zeroth-Order Optimization for Deep Model Training

Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark

Evolutionary Multiobjective Molecule Optimization in an Implicit Chemical Space

Human-level molecular optimization driven by mol-gene evolution

Accelerating Black-Box Molecular Property Optimization by Adaptively Learning Sparse Subspaces

ZO-AdaMU Optimizer: Adapting Perturbation by the Momentum and Uncertainty in Zeroth-Order Optimization

Efficient molecular conformation generation with quantum-inspired algorithm

Molecule discovery and optimization via evolutionary swarm intelligence

ZO-AdaMM: Zeroth-Order Adaptive Momentum Method for Black-Box Optimization

Optimization of General Molecular Properties in the Equilibrium Geometry Using Quantum Alchemy: An Inverse Molecular Design Approach

Zeroth-Order Optimization Meets Human Feedback: Provable Learning via Ranking Oracles

GuacaMol: Benchmarking Models for De Novo Molecular Design

Graph-Based Molecular Pareto Optimisation

Diagnosing and fixing common problems in Bayesian optimization for molecule design

Q-Drug: a Framework to bring Drug Design into Quantum Space using Deep Learning