CLASS: Enhancing Cross-Modal Text-Molecule Retrieval Performance and Training Efficiency

TL;DR

This work addresses cross-modal text–molecule retrieval by introducing CLASS, a Curriculum Learning-based framework that jointly optimizes sample difficulty scheduling and adaptive training intensity. By quantifying per-sample difficulty through cross-modal similarities and progressively presenting easier samples before harder ones, CLASS improves retrieval performance while reducing training time. The approach is back-bone agnostic and demonstrated on AMAN and ORMA using the ChEBI-20 dataset, achieving superior metrics and notable efficiency gains. Overall, CLASS offers a practical, generalizable mechanism to enhance multimodal alignment and training efficiency in molecular retrieval tasks.

Abstract

Cross-modal text-molecule retrieval task bridges molecule structures and natural language descriptions. Existing methods predominantly focus on aligning text modality and molecule modality, yet they overlook adaptively adjusting the learning states at different training stages and enhancing training efficiency. To tackle these challenges, this paper proposes a Curriculum Learning-bAsed croSS-modal text-molecule training framework (CLASS), which can be integrated with any backbone to yield promising performance improvement. Specifically, we quantify the sample difficulty considering both text modality and molecule modality, and design a sample scheduler to introduce training samples via an easy-to-difficult paradigm as the training advances, remarkably reducing the scale of training samples at the early stage of training and improving training efficiency. Moreover, we introduce adaptive intensity learning to increase the training intensity as the training progresses, which adaptively controls the learning intensity across all curriculum stages. Experimental results on the ChEBI-20 dataset demonstrate that our proposed method gains superior performance, simultaneously achieving prominent time savings.

Figures (5)

• Figure 1: An overview of cross-modal text-molecule retrieval tasks, involving text-molecule retrieval task and molecule-text retrieval task. The text-molecule retrieval task refers to retrieving the correct molecule using text as a query, while the molecule-text retrieval task is to retrieve the corresponding text description for the molecule. The red text indicates the ground-truth retrieval result.
• Figure 2: Overview of CLASS. All the molecules are represented by white spheres for H, red for O, gray for C, and yellow for P. Initially, the multimodal encoder (①) encodes $z_i$ and $z_j$, and then inputs them into the sample difficulty quantification (②) to calculate the similarity between samples, quantifying the difficulty of sample $z_i$ based on the number $\mathcal{N}_{i}$ of confusing samples. Thereafter, the sample scheduler (③) based on a curriculum learning strategy introduces training samples via an easy-to-hard paradigm. Finally, the adaptive intensity learning (④) dynamically adjusts the model's training intensity to control the global training process of the model.
• Figure 3: The ratio of training sample used in ORMA and our CLASS (ORMA) during total training epochs for both two retrieval tasks.
• Figure 4: The ratio of training sample used in AMAN and our CLASS (AMAN) during total training epochs for both two retrieval tasks.
• Figure 5: Cases of our model and AMAN in the cross-modal text-molecule retrieval task. The red box highlights the candidate retrieved by our model that exactly corresponds to the ground truth.