GPMO: Gradient Perturbation-Based Contrastive Learning for Molecule Optimization

Xixi Yang, Li Fu, Yafeng Deng, Yuansheng Liu, Dongsheng Cao, Xiangxiang ZengIJCAI 2023 · DOI

Abstract

Molecule optimization (MO) is a critical phase in drug discovery, focusing on refining lead compounds to improve ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) properties. While sequence-to-sequence models have shown promise, they frequently suffer from "exposure bias," where small errors during inference cascade into invalid molecular structures. We present GPMO, a novel framework that utilizes gradient-based perturbations to create challenging positive and negative examples for contrastive learning. Our results demonstrate that GPMO significantly outperforms existing baselines in multi-property optimization tasks, achieving a 96.72% success rate in satisfying multiple constraints simultaneously.

Introduction

In drug discovery, the transition from a lead compound to a clinical candidate requires balancing multiple competing chemical properties. Deep learning models often frame this as a machine translation problem, converting a SMILES string of a molecule with suboptimal properties into one with improved characteristics. However, the standard training paradigm, "teacher-forcing," creates a gap between training and real-world usage. During inference, the model relies on its own previous (and potentially erroneous) outputs, leading to "exposure bias." This often results in the generation of chemically invalid or structurally irrelevant molecules.

The GPMO Methodology

GPMO addresses exposure bias by integrating contrastive learning with a Transformer-based encoder-decoder architecture. Unlike traditional contrastive learning that relies on random sampling or domain-specific heuristics to find similar (positive) or dissimilar (negative) molecules, GPMO uses gradient-based perturbations to generate "hard" training examples.

The GPMO framework architecture, illustrating the contrastive learning strategy with positive and negative perturbations and the gradient-based process for generating hard examples.

The framework employs two primary types of perturbations:

  1. Negative Perturbations: A small mathematical nudge is applied to the hidden state of the target molecule to create a "hard negative." This molecule appears nearly identical to the target in the embedding space but is technically incorrect, forcing the model to learn precise chemical boundaries.
  2. Positive Perturbations: "Hard positives" are generated by nudging the hidden state to maximize similarity while maintaining a distribution consistent with valid molecular structures via KL-divergence loss.

The model is further steered by specific condition tokens (e.g., Delta_LogD(0.3, 0.5]), allowing for precise control over the degree of optimization.

Experimental Setup

The model was pre-trained on the MOSES dataset to learn the fundamental grammar of SMILES strings. For the optimization task, we utilized a dataset of matched molecular pairs (MMPs) labeled with three key ADMET properties: LogD (lipophilicity), Solubility, and Clint (metabolic stability). Performance was measured based on chemical validity, structural consistency (MMP relationship), and the ability to satisfy all property constraints simultaneously ("All" metric).

Results and Analysis

GPMO demonstrated superior performance across all benchmarks. Notably, it achieved a 96.72% success rate on the "All" metric, a nearly 10% improvement over the previous state-of-the-art model, Modof. The model also exhibited high data efficiency, outperforming most baselines even when trained on only 10% of the available data.

Example of iterative molecule optimization at Round 5, showing the successful evolution of a lead compound with improved LogD and stable ADMET properties.

A key strength of GPMO is its ability to perform iterative optimization. In multi-round tests, the model successfully evolved molecules over five rounds, steadily improving target properties like LogD while preserving the core molecular scaffold.

Conclusion

GPMO provides a robust, domain-knowledge-free approach to molecule optimization. By leveraging gradient perturbations to combat exposure bias, it sets a new standard for multi-property molecular design. Future work will focus on expanding the property scope to include toxicity and synthetic accessibility, further bridging the gap between AI-driven design and laboratory synthesis.

Summary updated

BibTeX
@inproceedings{yang2023gpmo,
  author = {Xixi Yang and Li Fu and Yafeng Deng and Yuansheng Liu and Dongsheng Cao and Xiangxiang Zeng},
  title = {{GPMO: Gradient Perturbation-Based Contrastive Learning for Molecule Optimization}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.24963/ijcai.2023/549},
}
Figure