Bartosz Janusz

RePaRank predicts antibody-antigen interfaces using a proximity ranking task, outperforming benchmarks in paratope prediction and being competitive in epitope prediction without external resources.

This study benchmarks inverse folding methods for antibody design to assess their effectiveness in predicting binder sequences given a structure.

The paper creates a dataset of antibody-antigen complexes with mutations and evaluates affinity prediction methods, finding that current models poorly predict binding effects compared to empirical methods.

PairedAbNGS pools 58 single-cell bioprojects of natural paired human and mouse antibodies, revealing heavy/light V gene pairing biases beyond usage frequencies and conserved CDR3-FWR2 inter-chain contacts.

The paper introduces nanoFOLD, an inverse folding method for designing nanobodies, and applies it to binder enrichment and novel design.

RIOT annotates immunoglobulin sequences by handling nucleotide and amino acid data, using an open germline database, and improving speed and accuracy over existing tools.

The paper introduces nanoBERT, a transformer model for predicting amino acids in nanobody sequences, showing its advantage over general models and gene-specific statistics.

The paper shows that pre-training deep learning models on physically accurate antibody structures improves the prediction of physically plausible CDR-H3 loop geometries.