Dice Question Streamline Icon: https://streamlinehq.com

Self-supervised objectives for molecular foundation models

Identify suitable self-supervised pretraining objectives for molecular foundation models that serve as molecular analogues of next-token prediction, such as masked-atom prediction, dynamics forecasting, or structure denoising.

Information Square Streamline Icon: https://streamlinehq.com

Background

The paper raises the challenge of defining self-supervised objectives for molecular foundation models that align with physical structure and dynamics. Unlike language, molecules involve continuous configurations and temporal evolution, complicating direct analogues of next-token prediction.

Candidate objectives include predicting masked atoms or features, forecasting future dynamics segments, and denoising noisy structures, but the choice and formulation remain unsettled.

References

Key open questions: Self-supervised objective: What is the "next-token prediction" for molecules? Predicting masked atoms? Forecasting dynamics? Denoising structures?

— Learning Biomolecular Motion: The Physics-Informed Machine Learning Paradigm (2511.06585 - Deshpande, 10 Nov 2025) in Section 7, Future Directions—Physics-Grounded Foundation Models