[View in browser.](https://mjy.github.io/presentations/2019/SemanticDataModelsInAnatomy/index.html) This talk was written in [impress.js source](https://impress.js.org). Source is at [https://github.com/mjy/presentations/tree/main/2019/SemanticPhenotypeModelling](https://github.com/mjy/presentations/tree/main/2019/SemanticDataModelsInAnatomy) Other talks from the workshop are collected at [https://www.researchgate.net/project/Workshop-Semantic-Data-Models-in-Anatomy](https://www.researchgate.net/project/Workshop-Semantic-Data-Models-in-Anatomy) ----- # A list * 10 things that have consequences for how we model anatomy * Focus on requirements of taxonomy * Not all doom and gloom ----- # 0 - An example item ----- # 0 - ... and its Consequences * One * Two * `*` A consequence that requires no action * `?` A (more) poorly thought out point ----- # 1 - Taxonomists have published in _natural language_ for over **200 years** ----- # 1 - Consequences * * No model is needed? * A model must support what they do (diagnose species) * A model must let taxonomists flow naturally from NL observations to a formal representation * A majority(?) of statements in the model will be lossy ----- # 2 - Life is complex ----- # 2 - Consequences * Giant empty matrix, with few links * Linking nodes must be very carefully thought out * Models must isolate labels from concepts * RDF labels are not enough to identify "the same" nodes in disparate named graphs * Model must support evolving refinement ----- # 3 - Life is observed once, described, then ignored _Most anatomical descriptions will never be revised_ ----- # 3 - Consequences * Integrating previously described species will require NL processing * Let's be real- for most species descriptions will not be redone in a "native" semantic format, we simply don't have the time/resources * ? While our model's semantics must evolve, the statements/observations behind them won't * ? Our model must have "versioning" to reference the NL algorithm that were used * Our model must differentiate NL processed statements from "native" statements ----- # 4 - Taxonomists present _species_ descriptions ----- # 4 - Consequences * Is it important to provide a model that does more than what taxonomists want to do? * Few, though growing numbers of taxonomists, uniquely identify the specimens in their study * If we can't identify specimens, how are we going to identify their parts * An _instance anatomy_ sensu our workshop has never been published ----- # 5 - Reference ontologies (for gross anatomy) do not exist for most of life ----- # 5 - Consequences * ? Instance anatomies can't be merged at finer levels of granularity * ? Search/filter will need to be done on values, and therefor be variously unsatisfying * We must tackle the problem from top down and bottom up ----- # 6 - Converting human NL to a model is lossy process ----- # 6 - Consequences * The model should emphasize _minimizing_ loss of meaning * No one model will fix this issue * Similarly, conversion between representation models will be lossy, also suggesting minimizing loss is a goal- how to ensure this with model semantics? ----- # 7 - Humans can't agree ----- # 7 - Consequences * Semantics need to be fuzzy enough to draw conclusions across independent observations * It is unlikely we can have 1 graph of observations (e.g. instance anatomy) per entity being described * Merging/syncing data from the same, or different models remains, as always (sigh), the hardest problem ----- # 8 - Model organisms are described differently ----- # 8 - Consequences * Models must account from difference from "normal"/"wild type" type statements * We must work hard to escape from this relative approach less it persist into a more general usage ----- # 9 - All models need interfaces ----- # 9 - Consequences * Interfaces bias what and how models get used * Semantic models could be completed buried behind the symbolic representations that are used to capture data * Should our model be built to pre-adapt attributes/properties to the "visual" interface that will capture their instances? ----- # 10 - People want to use semantic phenotypes for AI, VR, and other buzzworthy things ----- # 10 - Consequences * Data may need specific attributes to make them useful for AI and other approaches * We should talk with Jim/the [SCATE](https://scate.phenoscape.org/) project * We need 3D coordinates for anatomy terms * See our [vronto](https://github.com/bioip/vronto) project ----- # 11 - ? Data are always generated with a purpose in mind ----- # 11 - Consequences * ? Published descriptions are not "inaccurate" (sensu our discussion in the workshop), they have fulfilled their purpose (and been accepted by a community of peers) * ? Even if we model data, we can't escape the baggage that is its original purpose? * ? Even in a universal model, some argue that data derived for one purpose are not suitable for another purpose ----- # 12 - URIs/IRIs ----- # 12 - Consequences * ? It is hard to _maintain_ and generate _resolvable_ URIs at scale * ? A universal model requires unique ids, this comes with significant issues such as services that ensure minted URIs are indeed unique * It is almost certain that we need centrally managed data "lakes"/"oceans"/repositories that our data can find their way to ----- # Conclusion * It is important that we look hard at our underlying premises * Do they reflect reality? * Do they reflect how work in biodiversity is actually done? * Do they "scale to biodiversity"? * Do they reflect realistic applications of technology? * Keep data representation and data production issues isolated * A litany of opportunity requires a pluralist approach to encouraging the use of semantics