Multi-modal Contrastive Learning with Negative Sampling Calibration for Phenotypic Drug Discovery

Jiahua Rao, Hanjing Lin, Leyu Chen, Jiancong Xie, Shuangjia Zheng, Yuedong YangCVPR 2025 · OpenReview

Abstract

Phenotypic drug discovery presents a promising strategy for identifying first-in-class drugs by bypassing the need for specific drug targets. Recent advances in cell-based phenotypic screening tools, including Cell Painting and the LINCS L1000, provide essential cellular data that capture biological responses to compounds. While the integration of the multi-modal data enhances the use of contrastive learning (CL) methods for molecular phenotypic representation, these approaches treat all negative pairs equally, failing to discriminate molecules with similar phenotypes. To address these challenges, we introduce a foundational framework MINER that dynamically estimates the likelihoods of sample pairs as negative pairs based on uni-modal disentangled representations. In addition, our approach incorporates a mixture fusion strategy to effectively integrate multimodal data, even in cases where certain modalities are missing. Extensive experiments demonstrate that our method enhances both molecular property prediction and molecule-phenotype retrieval accuracy. Moreover, it successfully recommends drug candidates from phenotype for complex diseases documented in the literature. These findings underscore MINER's potential to advance drug discovery by enabling deeper insights into disease mechanisms and improving drug candidate recommendations.

Introduction

Phenotypic Drug Discovery (PDD) focuses on observing how drugs alter cellular characteristics—such as morphology or gene expression—to identify therapeutic potential without requiring a predefined protein target. This approach is particularly valuable for complex diseases with poorly understood mechanisms. However, training effective AI models on this multi-modal data is challenging. Traditional Contrastive Learning (CL) methods typically employ binary weighting, treating all "negative" pairs (a molecule paired with a non-corresponding cellular response) as equally dissimilar. In biological systems, this assumption often fails; different molecules may induce similar phenotypic responses, leading to "false negatives" that introduce noise and hinder the model's ability to learn fine-grained chemical nuances.

The MINER Framework

To overcome these limitations, we propose MINER (MultI-modal Contrastive Learning with NEgative Sampling CalibRation). The framework integrates three core innovations to improve molecular and phenotypic representations.

The MINER architecture overview, showing multi-modal encoders, uni-modal denoising, calibrated contrastive learning, and MoE fusion.

1. Denoising Uni-modal Representations

MINER disentangles embeddings for each modality (molecular graphs, cell images, and gene expression) into informative features and redundant noise. An inter-correlation loss, based on Pearson correlation, ensures these components remain statistically independent, making the resulting features more robust to experimental noise.

2. Negative Sampling Calibration

Instead of binary weighting, MINER employs a Multi-Layer Perceptron (MLP) to predict an alignment probability ($\gamma$) for negative pairs. If the model identifies that a "negative" pair actually shares similar biological content, it reduces the penalty. This dynamic calibration allows the model to handle "weakly aligned" pairs more effectively than traditional CL.

3. Mixture-of-Experts (MoE) Fusion

Biological datasets often suffer from missing modalities. MINER addresses this by using a generative module to synthesize missing data and a Mixture-of-Experts layer. A "router" dynamically weights the contribution of each modality, ensuring the most reliable information is prioritized for predictions.

Experimental Evaluation

The framework was evaluated on two major benchmarks: Broad6K (6,567 molecules) and ChEMBL2K (2,355 molecules, with 75.33% missing gene expression data).

Molecular Property Prediction

MINER demonstrated superior performance in predicting molecular properties. On the ChEMBL2K dataset, it outperformed the previous best baseline (InfoCORE) by 3.6% in Average AUROC. Notably, MINER showed significant gains in identifying high-confidence hits at the 80%, 85%, and 90% AUROC thresholds.

Molecular property prediction results (AUROC) on ChEMBL2K and Broad6K datasets.

Retrieval and Drug Repurposing

In molecule-phenotype retrieval tasks, MINER achieved a 14.29% increase in Mean Reciprocal Rank (MRR) over InfoCORE for cell painting screens. Furthermore, the model was applied to rank drug candidates for Alzheimer’s Disease. 60% of the top 10 recommendations, including drugs like Noscapine and Clofibrate, were found to be supported by existing clinical literature, validating MINER's utility in real-world drug repurposing scenarios.

Conclusion

MINER addresses critical flaws in traditional contrastive learning for drug discovery by introducing negative sampling calibration and robust multi-modal fusion. By explicitly modeling noise and handling missing data through an MoE strategy, MINER provides a more nuanced and accurate framework for representing the complex relationship between molecules and biological phenotypes.

Summary updated

Abstract

Phenotypic drug discovery presents a promising strategy for identifying first-in-class drugs by bypassing the need for specific drug targets. Recent advances in cell-based phenotypic screening tools, including Cell Painting and the LINCS L1000, provide essential cellular data that capture biological responses to compounds. While the integration of the multi-modal data enhances the use of contrastive learning (CL) methods for molecular phenotypic representation, these approaches treat all negative pairs equally, failing to discriminate molecules with similar phenotypes. To address these challenges, we introduce a foundational framework MINER that dynamically estimates the likelihoods of sample pairs as negative pairs based on uni-modal disentangled representations. In addition, our approach incorporates a mixture fusion strategy to effectively integrate multimodal data, even in cases where certain modalities are missing. Extensive experiments demonstrate that our method enhances both molecular property prediction and molecule-phenotype retrieval accuracy. Moreover, it successfully recommends drug candidates from phenotype for complex diseases documented in the literature. These findings underscore MINER's potential to advance drug discovery by enabling deeper insights into disease mechanisms and improving drug candidate recommendations.
BibTeX
@inproceedings{rao2025multi,
  author = {Jiahua Rao and Hanjing Lin and Leyu Chen and Jiancong Xie and Shuangjia Zheng and Yuedong Yang},
  title = {{Multi-modal Contrastive Learning with Negative Sampling Calibration for Phenotypic Drug Discovery}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  url = {https://openreview.net/forum?id=nnG89nJIT6},
}
Figure