Winning Team: MagiXterial

Team Leader: Yifeng Tang

Team members: Qingchuan Yang, Han Xu, Zuochao Zhao, Huishuang Tan

Summary:
-Compared traditional ML methods with neural networks on predicting alloy properties from compositions
-Proposed a new stacking model which uses GBDT trees as featurizer followed by a neural network regressor
-Compared the performance of three methods on predicting alloy compositions from properties: Direct Backward NN Modelling, Iterative Searching and Optimization and Generative Model

Winning Team: AI.Rex

Team Leader: Yifeng Tang

Team members: Huishuang Tan, Hanzhang Liu, Jiayu Qin, Wanxiang Shen, Long Qian

Summary:
In chemical synthesis planning, retrosynthsis analysis tools can readily provide numerous synthetic paths, but that is beyond the capacity of experimental validation. Hence, evaluating the reaction feasibility accuractely and efficiently is vital in reducing the number of proposed paths and only retaining those have higher chance of success in wet lab. We successfully built models based on reaction SMILES or reaction graph, and achieved reasonably high test accuracy in both approches (>90%). We managed to integrate reactivity-related Quantum Mechanics descriptors into the feasibility prediction model and boosted the mode performance.

Winning Team: OMX

Team Leader: Wanxiang Shen

Team members: Wanxiang Shen

Summary:
The original intention was to establish an interpretable xAI-based omics machine learning model and algorithm flow for disease prediction, diagnosis and identification of key markers. The customized multi-channel AggMapNet outperformed state-of-the-art machine learning models, and the interpretable module of AggMapNet identified key metabolites or proteins for Covid-19 tests, and the predictions of disease severity were highly consistent with the biomarkers or biological mechanisms reported in the literature.

Winning Team: Alpha-Spec

Team Leader: Jian Song

Team members: Liqi Yan

Summary:
Deep learning technology was used to identify whether the secondary spectrum collected by the mass spectrometer were signal or noise. Alpha-Spec used an advanced deep learning module, Transformer, and the model was trained for 7 rounds. The deep learning model of Alpha-Spec achieved the recognition of spectral quality based solely on spectra (acc: ~70%). The spectra quality scoring system of the Alpha-Spec can improve DDA identification (PSM: ~3.3%).

Winning Team: TQL

Team Leader: Zheng Cao

Team members: Tiantao Liu

Summary:
The team proposed a synthesis evaluation system to analyze the synthetic accessibility and identify the possible side products.

Winning Team: CRITICAL

Team Leader: Ke Liu

Team members: Shangde Gao, Kaifan Yang, Yichao Fu

Summary:
The team built a model comprised of input layer, embedding layer, aggregation layer, attention layer and Set2Set layer to predict the stiffness tensor of materials.

Winning Team: Aimat

Team Leader: Xiangyu Zhang

Team members: Xiangyu Zhang

Summary:
The team Adapted the CGNN and GeoCGNN models to predict stiffness tensors of materials from their crystal structure