GPT Models to Predict Liquid-liquid Phase Separation Propensity of Peptides

Introduction: Liquid-liquid phase separation (LLPS) of biological macromolecules is a crucial phenomenon in the formation of intracellular membrane-less organelles. Due to a lack of physical barriers, the liquid condensates can exchange their components rapidly with the surroundings. Several proteins possess LLPS properties with intrinsically disordered regions (IDRs) comprising low-complexity domains (LCDs). There have been very few reports on the use of AI-based models to predict peptide sequences with LLPS properties. In this work, we utilize specialized transformer neural network-based language models, specifically GPT models, to decipher the molecular grammar of Intrinsically Disordered Proteins (IDPs). Further, we confirm the LLPS-forming ability of some of the peptides that were predicted to form LLPS through rigorous experiments.

Materials and

Methods: Using a GPT model that can generate LLPS-prone sequences, 2000 peptide sequences were generated, each with the potential to form LLPS. Four peptides with the highest propensity (LLPS+) were selected and synthesized using standard Fmoc chemistry and purified by HPLC. To assess LLPS, peptides were dissolved in 1X phosphate-buffered saline (PBS) at pH 7.4 and the formation of droplets was confirmed by optical microscopy and turbidity measurements at 600 nm. Further, the effect of ionic strength and pH on LLPS was monitored by measuring absorbance at 600 nm.

Results, Conclusions, and Discussions: Results &

Discussion: The results of our GPT model predicted that Prot 83, Prot 45, Prot 182, and Prot 28 have LLPS+ propensity (Fig. 1A). Optical microscopy data showed that in agreement with our predictive GPT model, all four peptides undergo phase separation in physiological buffers (Fig. 1B). There was considerable variation in the size of the droplets and presumably the peptides exhibit a concentration-dependent effect on droplet formation. Turbidity measurements at 600 nm also indicated differing kinetics of condensate formation between peptides (Fig. 1C). Further, variations in NaCl concentrations and pH also played profound roles in the LLPS properties of these peptides (data not shown).

Conclusions: Our results show that we have successfully developed a GPT model that can predict peptide/protein sequences with LLPS forming propensity.

Acknowledgements (Optional):