A multimodal fine-funded foundation model to support precision medicine of mesothelioma

Open to UK applicants only

Supervisors:

• Professor Dean Fennell df132@le.ac.uk
• Professor Hiuyu Zhou
• Professor Hongyu Zhou

Project Description:

BACKGROUND

Mesothelioma is a lethal cancer caused by asbestos, with less than 10% surviving more than 5 years. Precision medicine although feasible [1], is in its infancy [2]. New approaches to personalise therapy are needed to ensure patients acquire the best chance of clinical benefit, avoiding futile and potentially toxic therapy. Identifying patients who are likely to respond to a given treatment is challenging, often requiring complex multiomic interrogation. A foundation model (FM) is a large scale machine learning model, trained on vast and diverse datasets, that can be adapted to conduct specialised tasks through fine tuning using multimodal data, including the possibility of robust prediction of drug response from routine haematoxylin and eosin (H&E) slides used in routine diagnosis. FMs represent a new frontier in Artificial Intelligence that could revolutionise theranostics.

EXPERIMENTAL PLAN 

 We will install a local instance of prov-gigapath [3] which uses a novel vision transformer architecture and has been pretrained on 1.3 billion image tiles using self supervised learning (Dinov1.2), from 171K slides, as a base for fine tuning a mesothelioma foundation model utilising approximately 1000 deeply annotated mesotheliomas, derived from multiple cohort studies; MEDUSA , CONFIRM, MIST, VIM, and NERO. H&E+Multiomic data comprising whole exome, whole genome, methylome, transcriptome and proteome, will be integrated with radiological imaging and clinical (demographic, outcome) datasets, to fine tune a prov-gigapath to mesothelioma specific tasks.  We will prospectively fine-tune this foundation model in the SELECT phase II trial umbrella and evaluate its capability of diagnosing specific molecular alterations (MTAP deletion, BAP1, NF2  mutation) as well as its ability to forecast therapeutic responses. Performance will be measured by receiver operating characteristics and precision recall analyses. 

IMPLICATIONS FOR FUTURE RESEARCH

Training required for the development and deployment of a mesothelioma FM will encompass bioinformatics in multiomic pipelines and machine learning, and should be transformative, enabling robust H&E-based therapeutic stratification to accelerate precision medicine and improve patient outcomes.

References:

[1] Fennell, A.Bzura Nature Cancer 2024, Vol.3(8) , 902
[2] S. M. Janes, D. Alrifai and D. A. Fennell, N Engl J Med 2021 , 385(13) , 1207
[3] M. Zhang,…..D.A Fennell, Nat Commun 2024 Vol. 15 (1), 7187
[4] H. Xu, et al, Nature 2024 Vol. 630(8015) ,181