I am excited to announce that Platform 23.12 release comes with a new “Target Prioritisation” view, focusing on displaying target-specific properties in a disease agnostic way.
The target attributes have been aggregated into four main sections — Precedence, Tractability, Doability, Safety — and individually scored as part of an Open Targets project that aims to improve our target recommendation capability, called “Target Engine”. A huge team effort!!
A traffic light system has been designed to visually inform on target prioritisation, with the aim to facilitate target recommendations.
I’m an expert in target selection in the industry and a big fan of the traffic light system and other visualization tools that help with such an important, yet sophisticated process.
I’d like to share some feedback
First, the strongest benefits:
The traffic lights is fantastic choice. It really helps.
And another wonderful thing is an opportunity to get metric sources right at the same window.
Also there some discussion points:
The most important question is about the Association Score. Unfortunately, I haven’t yet found documentation about how weights are applied. But from what I can see the question rises: is it purposefully desired outcome if the target with 748 entities at Phase IV is equal to one with 3 entities at Phase I?
Could it be caused by equal weights of such unequal parameters for potential drug target like number and phase of competitors and ortholog identity (btw why not homology)? At least visually it seems like the difference between 83 and 89% in mouse target protein identity is as important as pipeline highest clinical phase?
As one can notice clinical phase is a very important if not a key parameter in target selection. As a portfolio manager in search I’d be glad to get more detailed information. I found a great examples at Covid tools page.
As well as ability to sort by individual parameters
Are you thinking about integrating these options into the “Target Prioritisation”?
Speaking about mouse orthologs.
I see there is other species data included. At least is shown at information window. This kind of information can also be important metric individually. At least cyno homology and closest human homologs. Have you discuss it and why mouse orthology was chosen in particular?
Also I have some ideas how to take these metrics into account with greater predictive accuracy.
Several thoughts about individual metrics.
membrane protein
Is it quality metric or just a modality-related parameter? If we talk about small drug why is non-membrane protein is in ‘yellow’ zone?
mouse models
First, it seems like the best option ‘0’ is at ‘yellow’ zone by design. Is it asumed?
Second, why is based on manifestations in knockout models? This is important for a certain range of diseases, but not for all of them. Are you planning to expand this metric?
secreted protein
Is secreted protein really always better option?
How do you factor ‘NA’ data into the overall scoring? After all, the lack of drug candidates in CT and secreted forms are significantly different in meaning.
Have you thought about adding the ability to hide some of the parameters, to rank by columns, to add a display of weights (e.g. brightness or size)? It is difficult for the human brain to visually evaluate a large amount of data at once, and the attempt to “take everything into account” may be an illusion and reduce the quality of choice.
I’m also very interested in the option of interactive custom weights
Unfortunately I can only attach two screenshots to illustrate the points. it seems like it would be nice for such a powerful visualization tool to expand the limit a bit
However I would be happy to discuss service and metrics development and share other ideas and observations!
Thank you for your detailed and valuable feedback, it’s much appreciated. I will let Juan respond to the points regarding the Target Prioritisation view specifically, but I can clear up one point of confusion —
The association score that you see in the Target Prioritisation view is the score derived from the evidence presented in the Target-Disease Association view:
You can find an explanation of the scoring and weighting of the evidence in our documentation.. The association score is a reflection of the amount and quality of evidence linking the target to the disease in question. It is separate to the data you see in the Target Prioritisation view, which presents information related to the target itself (the same data is presented in a Target Profile page).
You can modify the weighting of the evidence sources contributing to the overall association score in the Associations page (click on the cog, Show Data Sources Controls).
We have included mouse orthologs as it is the most widely use preclinical species, but I agree with you that there are other relevant organisms. As we wanted to show relevant data in a manageable environment for users, we had to prioritise which information convert to columns.
We take your thoughts about it for the future.
a) Membrane and secreted protein. They are both properties of the target. For instance, in membrane protein column if the target has information and is in membrane should be green; if it has location information from our sources (Uniprot and HPA) but is not membrane protein will be yellow. If the target has no location information will be shown as white (No data).
b) mouse models with 0 value are these targets which have mouse models with no severe phenotypes reported (please, check the target prioritisation documentation to know more about how we evaluate this and understand limitations).
6, 7 and 8. I think they are clarified with Helena’s answer. We have not implemented a quantitative score for this target properties in this view.
We will be very happy if you share with us your views about more columns information and how you think we could improve the user experience for target prioritisation. I also think that @carcruz and @Annalisa_Buniello would be glad to hear.
Yes, indeed, I took the score as a separate integral evaluation of targets as potential projects — must be a professional strain
I suggest throwing such a potential extension of functionality into the backlog to think about.
And then the possibility of sorting by individual parameters is all the more interesting
I probably asked too many questions at once)
Thanks for the detailed answer to all the points!
I have a few follow-ups. But first I would like to clarify the purpose of prioritisation. In the video there is the phrase ‘to evaluate the suitability of a target from a drug-discovery perspective’. I read it as “is the target a good idea for a new drug?”. From that perspective, I guess I still believe that the number of potential drugs in the clinic is a much stronger factor than the association. If we take that as the main sorting criterion, I’d like to at least screen out phase III+. Maybe there is some video showing an example of using the obtained priorities in a real or close to real case. It would answer a lot of questions
Furthermore,
If we are talking about modality agnostic comparison, why include modality-specific metrics (e.g. membrane protein, small molecule binding, chemical probes)?
I take your interest of filtering/sorting by clinical phase as showing the specific drug modalities available across clinical phases per target.
There are ongoing efforts for workshops on user cases/examples for interacting with the prioritisation framework. In the case of some of them be publicly available in the future, I’m sure @hcornu will let you know in the community and everywhere . Meanwhile, is fantastic if you use the tool for your work and want to share your experience.
Sorry, I did not explain well on that sentence. I meant with the “drug modality agnostic” showing target properties without being constrained to an specific drug modality.
We hope this prioritisation framework improves by the time, as we incorporate more data, have more iterations with partners and the community.
