Mo Vali
Mo Vali

Mo Vali
AI for Science

mv487@cam.ac.uk

I am a Physics PhD candidate at the Cavendish Laboratory, University of Cambridge, co-supervised by Diana Fusco, and Pietro Liò. I build machine learning methods and techniques for noisy, high-dimensional scientific data, from surface-enhanced Raman spectroscopy (SERS) to healthcare data. My work centres on:

  • Representation learning and tokenisation of scientific signals. Exploring how continuous, functional, time-series-like data such as vibrational spectra can be represented as sequences for language models. Designed and benchmarked text-, bitstream- and audio-style tokenisers for spectrum-to-structure prediction, with controlled input-format experiments improving molecule prediction 4.5-fold. This builds on SERS methods for isolating low-amplitude biological signals, including a previously uncharacterised indole derivative in E. coli (ACS Sensors, 2026).
  • Inverse design and simulation-based inference. Applying neural posterior estimation with normalising flows to an ODE model of phage–bacteria competition, identifying environments that produce target fitness landscapes.
  • Mechanistic interpretability of activation probes. Training linear probes for safety-relevant concepts such as harmful requests or deception in open LLMs (Qwen3-8B), then using attribution graphs to identify the internal features driving probe predictions and predict distribution shifts that may cause failure. Testing these predictions with experimental SERS data and synthetic DFT Raman spectra, where the sim-to-real gap has measurable ground truth.
  • Weakly supervised cross-modal retrieval. Aligning SERS and LC-MS data using multiple-instance contrastive learning, attention pooling and optimal transport, to retrieve the mass-spectrometry counterpart of unknown bacterial metabolites.
  • Multimodal modelling at scale. Fusing tabular, imaging, sequence and spectral data with attention-based models for IVF outcome prediction, supported by end-to-end pipelines across tens of thousands of records (AIME 2026).
  • Interpretable segmentation for biology. Developing RamanSeg, a prototype-based deep segmentation model with a bespoke activation-overlap loss, achieving 80.9% Dice on 21-channel hyperspectral data, with post-hoc feature attribution (arXiv, 2026).
  • Uncertainty-aware clinical machine learning. Building interpretable prediction pipelines for healthcare, combining multimodal learning, uncertainty quantification and model diagnostics for clinically relevant outcomes.

Alongside my PhD I have raised £225K in research funding, including a £150K grant from ai@cam, the University's flagship AI mission.

If you'd like to talk about any of this - or about squash, or 19th-century Russian novels - feel free to drop me an email.

News

2026 New preprint T-MOXAI - a hierarchical framework for temporal multimodal data (arXiv).
2026 Our SERS work identifying a previously uncharacterised indole derivative in E. coli metabolism is out in ACS Sensors.
2026 New preprint RamanSeg - interpretability-driven deep learning on Raman spectra for cancer diagnosis (arXiv).
2026 Awarded best paper at AIME 2026 for multimodal deep learning methods for healthcare.
2024 1st prize for best presentation at the RCOG World Congress, out of 1,109 eposters.
2023 Started my Physics PhD at the Cavendish Laboratory, University of Cambridge.

Things I care about

Selected Publications

  1. T-MOXAI: A Hierarchical Explainability Framework for Temporal Multimodal Data
    A. Inha, M. Vali et al.
    arXiv, 2026
    arXiv
  2. Nanoplasmonic SERS reveals a previously uncharacterised indole derivative in E. coli metabolism
    M. Vali et al.
    ACS Sensors, 2026
    Journal
  3. RamanSeg: Interpretability-driven Deep Learning on Raman Spectra for Cancer Diagnosis
    C. Tomy, M. Vali et al.
    arXiv, 2026
    arXiv
  4. AutoWrinkleID: a machine learning pipeline for biofilm wrinkle identification and quantitative analysis
    M. Murgia, M. Vali et al.
    bioRxiv, 2026
    bioRxiv
  5. Prediction of Embryo Transfer Outcomes Using Multi-Modal Deep Learning Methods
    M. Vali et al.
    British Journal of Obstetrics and Gynaecology (AIME 2026)
    Journal
  6. Broadband coherent Raman platform for stimulated Raman histology
    Talone et al.
    Proc. SPIE — Advanced Chemical Microscopy for Life Science and Translational Medicine, 2024
    DOI

Selected Invited Talks

Student Supervision

Things I'm reading

Affiliations

Misc