Novor.ai – Increased precision and accuracy utilizing an AI model for de novo sequencing

Significance of the Topic

Peptide de novo sequencing enables the determination of peptide sequences without relying on existing protein databases, which is essential for immunopeptidomics, metaproteomics and other applications lacking enzyme specificity. High-resolution instruments like timsTOF generate complex spectra that demand accurate and fast de novo algorithms to uncover novel epitopes and biomarkers.

Goals and Study Overview

This work introduces Novor.ai, an AI-driven enhancement of the established Novor de novo sequencing algorithm. The primary objective was to assess its precision and recall on previously unseen MHC class I and II datasets, comparing performance against the existing BPS Novor algorithm optimized for timsTOF data.

Methodology and Used Instrumentation

• AI Model Design: Decision‐tree scoring functions were replaced by AI‐derived scoring functions, while final sequence assembly was performed by a dynamic programming algorithm that enforces precursor mass error tolerances.
• Training Data: More than 7.3 million MS2 spectra were used, sourced from NIST and MassIVE‐KB spectral libraries, the SysteMHC Atlas and a custom timsTOF MHC peptide library.
• Validation Data: MHC class I and II samples from Hoenisch‐Gravel et al. were run in triplicate; ground truth assignments were obtained via the TIMS2Rescore workflow in Bruker ProteoScape coupled with ProLuCID at 1% FDR.
• Instrumentation: Bruker timsTOF mass spectrometer with trapped ion mobility spectrometry, data processed through Bruker ProteoScape and TIMS2Rescore.

Main Results and Discussion

Novor.ai achieved an average 19% improvement in correct amino acid identifications and a 16% peptide‐level gain on the MHC-I dataset compared to BPS Novor. For MHC-II data, peptide assignment accuracy increased by 26%. Precision–recall curves and percentage‐correct metrics highlight these performance gains. Although Novor.ai operates more slowly than BPS Novor, the marked accuracy improvements justify further optimization. Direct comparisons with other AI de novo tools remain to be conducted.

Benefits and Practical Applications

• Improved confidence in immunopeptidomic analyses, enhancing tumor antigen discovery and epitope mapping workflows.
• Database‐independent sequencing suitable for metaproteomics and non‐specific enzymatic digestions.
• Valuable for laboratories using high‐resolution timsTOF platforms seeking robust de novo sequencing solutions.

Future Trends and Potential Applications

Further refinement of model parameters and architecture is expected to boost processing speed and accuracy. Expanding benchmarks to include other AI de novo sequencing tools and diverse instrument platforms will validate broader applicability. Implementation of transfer learning for non‐timsTOF datasets and integration into clinical proteomics pipelines represent promising future directions.

Conclusion

Novor.ai significantly outperforms the existing BPS Novor algorithm in both precision and accuracy for timsTOF immunopeptidomic data. Continued optimization and comprehensive benchmarking against alternative AI models will establish its role in next‐generation de novo sequencing.

Reference

• Hoenisch‐Gravel N. et al. TOFIMS mass spectrometry‐based immunopeptidomics refines tumor antigen identification. Nat Commun 14, 7472 (2023).