From SMILES to Enhanced Molecular Property Prediction: A Unified Multimodal Framework with Predicted 3D Conformers and Contrastive Learning Techniques

Journal of Chemical Information and Modeling, Год журнала: 2024, Номер unknown

Опубликована: Дек. 6, 2024

We present a novel molecular property prediction framework that requires only the SMILES format as input but is designed to be multimodal by incorporating predicted 3D conformer representations. Our model captures comprehensive features leveraging both sequential character structure of and three-dimensional spatial conformers. The employs contrastive learning techniques, utilizing InfoNCE loss align embeddings, along with task-specific functions, such ConR for regression SupCon classification. To address data imbalance, we incorporate feature distribution smoothing (FDS), common challenge in drug discovery. evaluated through multiple case studies, including SARS-CoV-2 docking score prediction, using MoleculeNet sets, kinase inhibitor JAK-1, JAK-2, MAPK-14 custom sets curated from PubChem. results consistently outperformed state-of-the-art methods, FDS significantly improving tasks enhancing classification performance. These findings highlight flexibility robustness our model, demonstrating its effectiveness across diverse tasks, promising applications discovery analysis.

Язык: Английский

---

Knowledge-enhanced Relation Graph and Task Sampling for few-shot molecular property prediction

Information Sciences, Год журнала: 2025, Номер unknown, С. 122357 - 122357

Опубликована: Май 1, 2025

Язык: Английский

---

Advancing Image Generation with Denoising Diffusion Probabilistic Model and ConvNeXt-V2: A novel approach for enhanced diversity and quality

Computer Vision and Image Understanding, Год журнала: 2024, Номер 247, С. 104077 - 104077

Опубликована: Июль 14, 2024

Язык: Английский

---

Adapting differential molecular representation with hierarchical prompts for multi-label property prediction

Briefings in Bioinformatics, Год журнала: 2024, Номер 25(5)

Опубликована: Июль 25, 2024

Abstract Accurate prediction of molecular properties is crucial in drug discovery. Traditional methods often overlook that real-world molecules typically exhibit multiple property labels with complex correlations. To this end, we propose a novel framework, HiPM, which stands for Hierarchical Prompted Molecular representation learning framework. HiPM leverages task-aware prompts to enhance the differential expression tasks representations and mitigate negative transfer caused by conflicts individual task information. Our framework comprises two core components: Representation Encoder (MRE) Task-Aware Prompter (TAP). MRE employs hierarchical message-passing network architecture capture features at both atom motif levels. Meanwhile, TAP utilizes agglomerative clustering algorithm construct prompt tree reflects affinity distinctiveness, enabling model consider multi-granular correlation information among tasks, thereby effectively handling complexity multi-label prediction. Extensive experiments demonstrate achieves state-of-the-art performance across various datasets, offering perspective on learning.

Язык: Английский

---

Orthrus: Towards Evolutionary and Functional RNA Foundation Models

bioRxiv (Cold Spring Harbor Laboratory), Год журнала: 2024, Номер unknown

Опубликована: Окт. 12, 2024

In the face of rapidly accumulating genomic data, our understanding RNA regulatory code remains incomplete. Pre-trained foundation models offer an avenue to adapt learned representations biological prediction tasks. However, existing are trained using strategies borrowed from textual or visual domains, such as masked language modelling next token prediction, that do not leverage domain knowledge. Here, we introduce Orthrus, a Mamba-based model pre-trained novel self-supervised contrastive learning objective with augmentations. Orthrus is by maximizing embedding similarity between curated pairs transcripts, where formed splice isoforms 10 organisms and transcripts orthologous genes in 400+ mammalian species Zoonomia Project. This training results latent representation clusters sequences functional evolutionary similarities. We find generalized mature isoform significantly outperform on five mRNA property tasks, requires only fraction fine-tuning data so.

Язык: Английский

---

Multi-granularity physicochemical-inspired molecular representation learning for property prediction

Expert Systems with Applications, Год журнала: 2024, Номер unknown, С. 126115 - 126115

Опубликована: Дек. 1, 2024

Язык: Английский

---