Preliminary evaluation of the Transfer Learning capabilities of MOTL for multi-omics cancer survival analysis

Le résumé fourni par la source

The high-dimensionality of multi-omics datasets combined with the multiplicity of their interactions make their analysis challenging, especially in the field of prognosis. In an insightful study (Herrmann et al. 2021) where 13 survival analysis methods were compared on 18 cancer data-sets analyzed separately, they highlight the limit of molecular data for the improvement of a prognosis based only on clinical data. To tackle this limit, we added joint Dimension Reduction (jDR) methods in a follow-up work (Le Goff et al. 2025) and managed to show a statistical improvement, aggregated across all cancers, in favor of these methods against models based on clinical data only. Yet, the main factor driving prediction performances is the number of patients rather than the chosen methodology. Here, we will focus on a paradigm to virtually increase the number of samples, Transfer Learning (TL). In order to learn a general "cancer knowledge" that will be transferred to a targeted cancer and ease its survival analysis, the pre-trained model will be learnt on a pan-cancer multi-omics dataset. We start by evaluating TL in the context of jDR methods as they seem to perform best in this context, raising MOTL (Hirst et al. 2025), built on the jDR method MOFA (Argelaguet et al. 2018), as the only candidate. MOFA and MOTL were compared on the 7 smallest datasets analyzed by Herrmann et al. (2021), as they are the most likely to benefit from TL. Preliminary results need to be consolidated but mainly highlight one cancer in favor of MOTL and one against (only significant results after correction). Attempts to interpret these results enhance the need to determine a procedure to select the best subset of cancers out of which the pre-trained model will be learnt for the transfer to be optimal on a designated targeted cancer.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.