Periodic Reporting for period 2 - LIMORD (Longitudinal Integrative Models for Online Relapse Detection)
Période du rapport: 2023-08-01 au 2024-07-31
The project is focused on exploiting longitudinal and heterogeneous patient data in order to detect their relapse.
In Multiple Myeloma (MM), detecting relapse remains a difficult task, with about a third of detections evaluated from painful symptoms. Clinical factors such as translocation t(4;14) or deletion of chromosome 17p are known to impact prognosis but there remains a large heterogeneity and the underlying regulation mechanisms are poorly understood.
Personalized medicine is a practice of medicine that has started emerging with the rise of genomics technologies in the early 2000s, and which aims at tailoring medical decisions to the patient’s predicted response. It is often misinterpreted as implying that unique treatments can be designed for each individual. If this idea represents the holy grail of the field, the current status rather consists in using diagnostic tests to determine which medical treatments will work best for each patient, by stratifying diseases into subgroups.
This project is an attempt to integrate high-dimensional and longitudinal sequencing and clinical data in order to provide earlier relapse detections, more relevant patient classifications, and get insight into the complex mechanisms driving disease progression and reaction to treatment.
In Multiple Myeloma (MM), detecting relapse remains a difficult task, with about a third of detections evaluated from painful symptoms. Clinical factors such as translocation t(4;14) or deletion of chromosome 17p are known to impact prognosis but there remains a large heterogeneity and the underlying regulation mechanisms are poorly understood.
Personalized medicine is a practice of medicine that has started emerging with the rise of genomics technologies in the early 2000s, and which aims at tailoring medical decisions to the patient’s predicted response. It is often misinterpreted as implying that unique treatments can be designed for each individual. If this idea represents the holy grail of the field, the current status rather consists in using diagnostic tests to determine which medical treatments will work best for each patient, by stratifying diseases into subgroups.
This project is an attempt to integrate high-dimensional and longitudinal sequencing and clinical data in order to provide earlier relapse detections, more relevant patient classifications, and get insight into the complex mechanisms driving disease progression and reaction to treatment.
The project has included the development of a theoretical probabilitstic model to help clinicians make the best possible decisions in the following of their patients. To this effect, we met with Dipti Talaulikar, Multiple Myeloma expert at ANU, analysed large cohort data, and fitted the parameters of our models to obtain a coherent and useful initial algorithm. This model has been calibrated from expert knowledge as well as a very large data set of marker data measurements for a cohort of several hundreds of multiple myeloma patients. The implementation of the algorithm is freely available online, and the work has been submitted for publication and presented at several conferences and colloquiums.
The second part of the project is dedicated to the identification of RNA modifications from direct sequencing data, and its possible application to patient samples. We have designed a tool to assess RNA integrity and sample quality and are have tested our approach on numerous available datasets, including some inhouse controlled degradation experiments to validate our findings. The tool, freely available online, allows the estimation of RNA decay rate and sample comparisons. The work has been submitted for publication and presented at several international conferences, including specialised conferences in the field of RNA decay biologists, which were not initially targeted.
We have also developed a deep-learning algorithm to identify 4 of the most prevalent RNA modifications in human mRNA (m6A, m5C, ac4C and pU). Though the tool is already available online, changes in the sequencing technology led us to reproduce all training datasets, and then adapt and re-train all our models, so we are still currently applying our tool on a large-scale conservation project.
The usage of those tools on a pilot project including the sequencing of 9 patient samples has allowed us to design better RNA extraction and library preparations for patient samples, and has allowed to obtain so insight into the modification landscape of multiple myeloma, in relation with risk factors.
The last part of the project is dedicated to our probabilistic model extension to include, in particular, learning-while-managing techniques. Our prototype algorithm is now released and we are working on the theoretical aspects of the learning framework. We have embedded our model in a Bayes-Adaptative Partially Observed Markov Decision Process and are now working on identifying the optimal resolution strategy. When this is achieved, we will start its diffusion in medical groups to showcase the added value of our approach.
The second part of the project is dedicated to the identification of RNA modifications from direct sequencing data, and its possible application to patient samples. We have designed a tool to assess RNA integrity and sample quality and are have tested our approach on numerous available datasets, including some inhouse controlled degradation experiments to validate our findings. The tool, freely available online, allows the estimation of RNA decay rate and sample comparisons. The work has been submitted for publication and presented at several international conferences, including specialised conferences in the field of RNA decay biologists, which were not initially targeted.
We have also developed a deep-learning algorithm to identify 4 of the most prevalent RNA modifications in human mRNA (m6A, m5C, ac4C and pU). Though the tool is already available online, changes in the sequencing technology led us to reproduce all training datasets, and then adapt and re-train all our models, so we are still currently applying our tool on a large-scale conservation project.
The usage of those tools on a pilot project including the sequencing of 9 patient samples has allowed us to design better RNA extraction and library preparations for patient samples, and has allowed to obtain so insight into the modification landscape of multiple myeloma, in relation with risk factors.
The last part of the project is dedicated to our probabilistic model extension to include, in particular, learning-while-managing techniques. Our prototype algorithm is now released and we are working on the theoretical aspects of the learning framework. We have embedded our model in a Bayes-Adaptative Partially Observed Markov Decision Process and are now working on identifying the optimal resolution strategy. When this is achieved, we will start its diffusion in medical groups to showcase the added value of our approach.
We expect that by the end of the project, we will have developed new and widespread tools to analyse data from direct sequencing technologies, will have showcased their applicability on patient samples, will have provided a landscape of RNA modifications in Myeloma (with a comparison between different levels of risks between patients), and will have provided a new probabilistic model for better and faster relapse detection, learning the specificity of each patients, and able to incorporate multiple omics data.