OPEN PROBLEMS IN VISIBLE LEARNING
(DCELL AND DRUGCELL)
Although artificial neural networks are powerful classifiers, their internal structures are hard to interpret. In the life sciences, extensive knowledge of cell biology provides an opportunity to design visible neural networks (VNNs) that couple the model’s inner workings to those of real systems. Towards this goal we have recently developed The DCell, a VNN embedded in the hierarchical structure of approximately 2,500 subsystems comprising a eukaryotic cell (Ma et al. Nature Methods 2018). Trained on several million genotypes, DCell accurately simulates cellular growth phenotypes and also captures the molecular pathways by which alterations to the genome lead to changes in cellular growth. Using this concept we have built a VNN simulation of a cancer cell, called DrugCell, which predicts the response of a patient’s tumor to specific drugs alongside the cellular subsystems most important for governing that drug response (Kuenzi, Park et al. Cancer Cell 2020). DrugCell thus has significant interest for use in precision medicine applications.
While DCell and DrugCell are exciting proofs of concept, they also raise numerous questions in how visible neural networks should be best developed and applied. Open questions relate to model structure, training, interpretation, clinical use, and access via the web. Below is an incomplete list of open problems. You may also be interested in proposing a question not yet on this list.
– Ideker Lab Bioinformatics Team, Fall 2021
Integrate DrugCell with few-shot transfer learning. Another previous paper by Ma et al. (Nature Cancer 2021) demonstrated an algorithm called TCRP (Transfer of Cellular Response Prediction) based on techniques from the field of transfer learning. In particular “few-shot” learning is a framework whereby a model is trained not for optimal prediction accuracy, but for optimal transferability to a new context. This paradigm mirrors the need in biomedicine to transfer results from basic research contexts, such as high throughput screens in cell lines, to clinical contexts focused on tumors in patients. Once trained in this way, a general model is obtained which can be quickly tuned to new contexts with few additional samples (“shots”). A significant research direction is to explore whether these or related methods of transfer learning can be used within DrugCell to allow more efficient transfer of the cancer simulation to different patient populations.
Category: Model training | Duration: 1-2 quarters
Evaluate the most reliable methods for interpreting visible neural network models. A key aspect of the DrugCell study is the ability to readout the molecular pathways that give rise to a drug response. One method, used in the paper, is to rank systems by the degree to which the system output is more predictive than the system input (Relative Local Improvement in Predictive Power or RLIPP). Other measures, such as saliency maps, are also of interest however. This project will identify alternative model interpretation methods and evaluate them in a comparative analysis.
Category: Interpretation | Duration: 1-2 quarters
De novo inference of model structure. Currently DCell and DrugCell take the structure of the model from a predetermined database or ontology of cell parts or functions. While this structure is instrumental to the ability to interpret the model, it may miss important aspects that can be inferred directly from data. The goal of this project is to work towards a hybrid model by which the structure of the model can be adjusted further during the learning process. Beginning with a simulation of how this process works will be critical.
Category: Model structure | Duration: 3+ quarters
Scoring of gene mutations. A primary input to DCell and DrugCell is a profile of gene mutations. Currently mutation states are represented as binary numbers (1 non-synonymous coding mutation, 0 other). However, coding alterations can differ markedly in quality (missense, nonsense or frameshift mutation; copy number abberration), magnitude as well as direction of effect (loss or gain of function). Moreover, alterations affecting important change to the non-coding genome must also be handled. This project will seek to increase model expressivity as it relates to these input features.
Category: Input/Output | Duration: 2 quarters
Incorporation of cell states beyond growth. Currently the DCell/DrugCell models translate a genetic mutation profile (input) to cell growth or proliferation (output). However, expression profiling has the potential to capture a rich set of cell states beyond proliferation, including states of resistance/sensitivity to different agents and stress, differentiation or pluripotence, and so on. Significant questions relate to where in the VNN model such states should be incorporated: as input, as output, or as direct observations of the states of internal systems? Other questions relate to how the model objective function should then be trained to incorporate these potentially multiple layers of state information. This project will attempt to incorporate expression information, addressing these and other considerations.
Category: Input/Output | Duration: 1+ quarters
Advance DrugCell towards clinical application. We are working together with clinicians at Moores Cancer Center to incorporate DrugCell as a drug recommender system in clinical trials. To prepare for clinical use, a number of aspects of the model must be critically evaluated. Are we using the right input genes? Which drugs have the most reliable predictive performance for clinical applications? Should less predictive drugs factor into model training, or should they be removed entirely? Can the model structure be pruned to a much smaller size, or replaced by a simple set of logic or rules, based on only the systems most informative for prediction?
Category: Clinical | Duration: 3+ quarters
Build a visible learning system for autism, diabetes, and other disorders. Thus far, we have explored visible machine learning systems in model species (budding yeast, DCell) and for use in prediction of cancer drug response (DrugCell). However, such models have exciting applications in many other disease areas. A significant lab interest is in understanding the genetics of neuropsychiatric disorders, thus training a DCell-like model for these diseases would be a highly significant direction. A separate lab interest is in modeling diabetes.
Category: New applications | Duration: 3+ quarters
DrugCell Oracle Website. Complete and publish the drugcell.ucsd.edu interactive server to enable use of DrugCell by both researchers and clinicians.
Category: Accessibility | Duration: 1-2 quarters
Methods enabling interpretability of compound structure. DCell and DrugCell enable an interpretation of predictions by means of a Visual Neural Network (VNN). The VNN scores protein assemblies and pathways according to their relevance to the prediction. The network which is modelling compound structure is still a black box. This project aims to enable an interpretability of the most relevant compound substructures (atoms, bonds) by applying and evaluating different approaches, including attention-based encoding in the style of PaccMann [Oskooei 2018] and graph convolutional neural networks.
Category: Interpretation | Duration: 2-3 quarters