Structural characterisation of chemical systems in many research fields relies on one-dimensional proton Nuclear Magnetic Resonance (1H NMR) spectroscopy due to its rapid acquisition and experimental simplicity. Accurate spectral interpretation is expert-driven, manual, and difficult to scale, limiting reproducibility and applicability in high-throughput settings. Although heuristic and deep learning approaches have been explored, automated extraction of chemically meaningful information from NMR spectra remains an open challenge. This doctoral work investigates how such information is encoded in resonance signals and develops efficient, reliable, and uncertainty-aware algorithms for automated 1H NMR spectral interpretation. MuSe Net, a supervised probabilistic deep learning framework for automated signal detection and multiplet splitting pattern classification with uncertainty quantification, is presented. MuSe Net achieves accurate predictions across a wide range of spin systems and experimental conditions, remaining robust at low signal-to-noise ratios and in the presence of negative peaks. This makes it an effective component of an automated pipeline for photo-Chemically Induced Dynamic Nuclear Polarization (CIDNP) NMR data analysis. The pipeline combines spectral and molecular structure data using probabilistic and deep learning approaches to quantify binding-induced signal changes in ligand-protein spectra, enabling binding strength estimation, hit classification, and fragment scoring with minimal user intervention. Developed through industrial collaboration within Innosuisse-funded projects, MuSe Net was realised in partnership with Bruker AG, while the photo-CIDNP pipeline supports NexMR AG in the management and annotation of large fragment-screening databases. These algorithms aim to accelerate workflows and increase the accessibility of NMR spectral analysis in applied research.
In diversi ambiti di ricerca, la caratterizzazione della struttura molecolare si basa sulla Risonanza Magnetica Nucleare unidimensionale del protone (1H NMR), per la sua semplicità sperimentale e rapidità di acquisizione. Tuttavia, l’interpretazione dei dati richiede ancora annotazione manuale da parte di esperti, con limitazioni in termini di riproducibilità e scalabilità. Nonostante l’uso di approcci euristici e di deep learning, l’estrazione automatica di informazioni chimiche rilevanti dagli spettri NMR rimane una sfida aperta. Questa tesi di dottorato studia come tali informazioni siano codificate nei segnali di risonanza e sviluppa algoritmi efficienti e affidabili, in grado di quantificare l’incertezza, per l’interpretazione automatica degli spettri 1H NMR. Viene introdotta MuSe Net, una rete neurale probabilistica per il rilevamento dei segnali e la classificazione dei pattern di splitting dei multipletti, che consente un’identificazione robusta di segnali sovrapposti. Il modello fornisce predizioni accurate su un’ampia gamma di sistemi di spin e condizioni sperimentali, risultando un componente efficace di una pipeline automatizzata per l’analisi di dati NMR photo-CIDNP. La pipeline combina dati spettrali e informazioni sulla struttura molecolare mediante approcci probabilistici e di deep learning per quantificare le variazioni di segnale indotte dal legame nei sistemi ligando–proteina, consentendo la stima della forza di legame, la classificazione degli hit e lo scoring dei frammenti con intervento minimo da parte dell’utente. Nell'ambito di progetti di collaborazione industriale finanziati da Innosuisse, MuSe Net è stato realizzato in partnership con Bruker AG, mentre la pipeline photo-CIDNP supporta NexMR AG nella gestione e annotazione di ampi database di screening di frammenti. Questi algoritmi mirano ad accelerare i flussi di lavoro e a incrementare l'accessibilità dell'analisi spettroscopica NMR nella ricerca applicata.
Deep Learning Approaches to the Automated Interpretation of One-Dimensional 1H NMR Spectra
FISCHETTI, GIULIA
2026
Abstract
Structural characterisation of chemical systems in many research fields relies on one-dimensional proton Nuclear Magnetic Resonance (1H NMR) spectroscopy due to its rapid acquisition and experimental simplicity. Accurate spectral interpretation is expert-driven, manual, and difficult to scale, limiting reproducibility and applicability in high-throughput settings. Although heuristic and deep learning approaches have been explored, automated extraction of chemically meaningful information from NMR spectra remains an open challenge. This doctoral work investigates how such information is encoded in resonance signals and develops efficient, reliable, and uncertainty-aware algorithms for automated 1H NMR spectral interpretation. MuSe Net, a supervised probabilistic deep learning framework for automated signal detection and multiplet splitting pattern classification with uncertainty quantification, is presented. MuSe Net achieves accurate predictions across a wide range of spin systems and experimental conditions, remaining robust at low signal-to-noise ratios and in the presence of negative peaks. This makes it an effective component of an automated pipeline for photo-Chemically Induced Dynamic Nuclear Polarization (CIDNP) NMR data analysis. The pipeline combines spectral and molecular structure data using probabilistic and deep learning approaches to quantify binding-induced signal changes in ligand-protein spectra, enabling binding strength estimation, hit classification, and fragment scoring with minimal user intervention. Developed through industrial collaboration within Innosuisse-funded projects, MuSe Net was realised in partnership with Bruker AG, while the photo-CIDNP pipeline supports NexMR AG in the management and annotation of large fragment-screening databases. These algorithms aim to accelerate workflows and increase the accessibility of NMR spectral analysis in applied research.| File | Dimensione | Formato | |
|---|---|---|---|
|
phd_thesis_GFischetti.pdf
accesso aperto
Licenza:
Tutti i diritti riservati
Dimensione
10.15 MB
Formato
Adobe PDF
|
10.15 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/377987
URN:NBN:IT:UNIVE-377987