Multimodal Representation\nLearning via Graph Isomorphism\nNetwork for Toxicity Multitask Learning

Toxicity is paramount for comprehending compound properties,\nparticularly\nin the early stages of drug design. Due to the diversity and complexity\nof toxic effects, it became a challenge to compute compound toxicity\ntasks. To address this issue, we propose a multimodal representation\nlearning model, termed multimodal graph isomorphism network (MMGIN),\nto address this challenge for compound toxicity multitask learning.\nBased on fingerprints and molecular graphs of compounds, our MMGIN\nmodel incorporates a multimodal representation learning model to acquire\na comprehensive compound representation. This model adopts a two-channel\nstructure to independently learn fingerprint representation and molecular\ngraph representation. Subsequently, two feedforward neural networks\nutilize the learned multimodal compound representation to perform\nmultitask learning, encompassing compound toxicity classification\nand multiple compound category classification simultaneously. To test\nthe effectiveness of our model, we constructed a novel data set, termed\nthe compound toxicity multitask learning (CTMTL) data set, derived\nfrom the TOXRIC data set. We compare our MMGIN model with other representative\nmachine learning and deep learning models on the CTMTL and Tox21 data\nsets. The experimental results demonstrate significant advancements\nachieved by our MMGIN model. Furthermore, the ablation study underscores\nthe effectiveness of the introduced fingerprints, molecular graphs,\nthe multimodal representation learning model, and the multitask learning\nmodel, showcasing the model’s superior predictive capability\nand robustness.