Research & projects

B.Sc. Physics thesis, Chair of Theoretical Biophysics, Technical University of Munich, 2026

Fine-tuning and fusion

Boltz-2’s affinity head was never trained on peptides. I fine-tuned it, trained three other models on the same data, and fused all four into one score.

0.573

Pearson r of the fused score on unseen receptors

0.74 Å

median centroid error with the peptide encoded as residues

7.3–11.8×

faster affinity scoring with Boltzina, my Boltz-2 fork

• Boltz-2
• Graph neural networks
• Protein language models
• Score fusion

Read the project →