A Multimodal Semi-Supervised Learning Framework for Pharmaceutical Cocrystals Prediction

Area of Science:

• Solid-state chemistry
• Pharmaceutical development
• Computational chemistry

Background:

• Cocrystal formation enhances drug properties like solubility and bioavailability.
• Identifying coformer pairs for cocrystal formation is challenging and uncertain.
• Existing experimental data is biased towards successful cocrystals, limiting machine-learning model accuracy.

Purpose of the Study:

• To reframe cocrystal prediction as a learning problem with missing negative information.
• To develop a conservative strategy for identifying molecular pairs unlikely to form cocrystals.
• To improve the reliability and generalization of cocrystal prediction models.

Main Methods:

• Leveraging multiple molecular descriptions (structural, electronic, physicochemical) to identify reliable "negative" examples.
• Using agreement between descriptions to exclude implausible coformer pairs.
• Mitigating data imbalance and fine-tuning a pretrained graph attention network with pseudonegative examples.

Main Results:

• A data-centric strategy significantly improves cocrystal prediction reliability and generalization.
• The approach enhances the accuracy of machine-learning models compared to existing methods.
• Successfully identifies molecular pairs unlikely to form cocrystals, providing reliable negatives.

Conclusions:

• Correcting for missing negative information is critical for realistic computational screening.
• This method offers a more useful approach for guiding experimental cocrystal discovery.
• Enhances the predictive power of machine learning in pharmaceutical solid-state chemistry.