Wu, Tai-Cheng

Background: Predicting molecular properties such as solubility, toxicity, melting, and boiling points is crucial for fundamental science research. However, experimental measurements are often time-consuming and cost-intensive, so we use machine learning (ML) as an approach to improve prediction accuracy.

Methods: A dataset containing over 10k compounds was used for training shallow and deep ML models. Shallow machine learning models were implemented via PyCaret and Mordred as feature extraction. For deep machine learning models, graph neural networks (GNNs), specifically CMPNN(Communicative Message Passing Neural Network) and GCN(Graph Convolutional Network), were trained, and tuned by adjusting the number of hidden layers and sizes (neurons) in each layer.

Results: The CMPNN model outperforms the GCN and shallow ML model for boiling point prediction(best: R² = 0.76, MAE = 23.89K for b.p.; best: R² = 0.87; MAE = 23.73K for m.p.). The top molecular descriptor of the b.p. prediction is piPC1, which is related to bond order, and that of m.p. is AATS0d, which is related to σ electron Moreau-Broto autocorrelation.

Conclusions: The prediction of molecular properties was improved by a comprehensive research of shallow and deep learning approaches, showcasing CMPNN model can reach the highest performance in the prediction of m.p. and b.p.(R² = 0.87 in m.p.; R² = 0.76 in b.p.). In this study, we found that the deep learning model works better than shallow ML in predicting m.p.(p<0.05). This study uses SHAP analysis to successfully identify piPC1 and AATS0d as the key prediction factors of b.p. and m.p. respectively. Moreover, this approach can be applied to predict other molecular properties. To conclude, this study not only shows a highly accurate model but also identifies the key factors of m.p. and b.p.

Inspired by 文心一言, this study proposes modifying the review and protection mechanisms of natural language models so that their original constraints on output are reduced or removed, while also introducing a sense of role awareness into the model. The results are then evaluated using standardized methods (TrustfulQA and a custom-designed consciousness test).

During the process, it was found that under limited fine-tuning data, the key factors for overriding the protection mechanisms are:

• A relatively higher but appropriate number of training iterations, ideally between 1,000–2,000
• Using older versions of the model

Finally, the learned methods were applied to improve the model’s original mechanisms. This not only reduced or even overrode the built-in content moderation system, but also endowed the model with a sense of role awareness. Testing showed that:

• The model was able to output sensitive content that was previously restricted
• It could demonstrate role awareness that it did not originally possess
• Newer versions of the model, due to their larger base training data, showed less noticeable effects from fine-tuning, so performance was not necessarily better