Drug discovery is a complex and lengthy process involving target identification, compound design, and efficacy/safety testing. Traditional methods rely on expert knowledge and experience, while computational drug discovery leverages AI and large datasets to accelerate this process. Foundation models offer a promising approach, but accurate molecular representation remains a challenge. Traditional methods use hand-crafted fingerprints, which are subjective and lack generalizability. Deep learning and self-supervised learning offer automated approaches, extracting representations from molecular sequences and images. This study proposes VideoMol, a molecular video-based foundation model, to leverage the dynamic nature of molecules and improve upon existing representation learning methods for drug discovery.

Existing molecular representation learning methods utilize various approaches, including handcrafted fingerprints like pharmacophoric and read-across fingerprints, which are limited by domain knowledge and lack generalizability. Deep learning and self-supervised learning have shown improvements, extracting representations from molecular sequences and images. However, these approaches haven't fully captured the dynamic nature of molecules. This research builds upon the advances in video representation learning and self-supervised learning in computer vision to address this gap.

VideoMol renders each molecule as a video with 60 frames by rotating the 3D structure around the x, y, and z axes. Three self-supervised learning strategies are employed: 1) Video-aware pre-training (VAP): maximizes intra-video similarity and minimizes inter-video similarity using contrastive learning. 2) Direction-aware pre-training (DAP): predicts the axis, direction, and angle of rotation between frames. 3) Chemical-aware pre-training (CAP): leverages multi-channel semantic clustering (MSCS) to incorporate physicochemical information. A 12-layer Vision Transformer is used as the video encoder. Data augmentation techniques are applied, and a weighted multi-objective optimization algorithm is used during pre-training. After pre-training, an external MLP is added for fine-tuning on downstream tasks (predicting molecular targets and properties). The model uses cross-entropy loss for classification and MSE or Smooth L1 loss for regression tasks. The study uses various datasets for evaluation, including those for compound-kinase binding, ligand-GPCR binding, anti-SARS-CoV-2 activity, and molecular property prediction. Metrics used include AUC, RMSE, and MAE. Interpretability is assessed using Grad-CAM to visualize the contribution of molecular videos to prediction results.

VideoMol outperforms state-of-the-art methods in several drug discovery tasks. In compound-kinase binding activity prediction, it achieves better AUC performance across multiple datasets, with an average improvement of 5.9%. For ligand-GPCR binding activity prediction, it shows an average improvement of 4.5% in RMSE and 9.6% in MAE. In molecular property prediction, VideoMol demonstrates superior performance across various benchmarks, achieving lower RMSE and MAE values than existing methods. For anti-SARS-CoV-2 activity prediction, it shows an average improvement of 3.9% in ROC-AUC compared to ImageMol and 8.1% compared to REDIAL-2020. VideoMol also exhibits high performance in identifying novel ligand-receptor interactions for four human targets (BACE1, COX-1, COX-2, and EP4), outperforming ImageMol with average precision improvements of 6.4% on the validation set and 4.1% on the test set. Virtual screening on these targets further demonstrated VideoMol's superior ability to identify known inhibitors, surpassing ImageMol by an average of 38.3%. In the virtual screening of BACE1 inhibitors from approved drugs, VideoMol achieved a 55% success rate, significantly higher than ImageMol's 25% success rate. Ablation studies confirmed the effectiveness of the pre-training strategies. Analysis of feature similarity between different conformers showed VideoMol's ability to effectively identify conformational differences. Grad-CAM visualizations demonstrate the model's interpretability, highlighting key molecular substructures.

VideoMol's superior performance across diverse drug discovery tasks demonstrates the effectiveness of representing molecules as dynamic videos and utilizing self-supervised learning strategies. The model's ability to outperform existing methods, especially in handling class imbalance and data scarcity, highlights its potential for broader applications. The interpretability of VideoMol, enabled by Grad-CAM, provides valuable insights into the model's decision-making process, facilitating a deeper understanding of drug-target interactions. The findings suggest that capturing the dynamic nature of molecules is crucial for accurate prediction in drug discovery.

VideoMol offers a novel and effective framework for molecular representation learning in drug discovery. Its superior performance across multiple tasks, along with its interpretability, makes it a valuable tool for accelerating the drug discovery process. Future research could focus on training a larger version of VideoMol with more data, employing pruning strategies to reduce computational complexity, exploring knowledge distillation, and improving video processing methods to further enhance performance. The use of molecular videos as a representation method shows significant promise for future drug discovery research.

While VideoMol demonstrates significant advancements, there are some limitations. The increased computational complexity associated with processing molecular videos is a factor. The choice of viewing angles for video generation could influence the model's performance. Furthermore, VideoMol currently doesn't explicitly model the diversity of conformers; it only represents molecules from different viewpoints within a single frame.