#ECCB18 is on. I was supposed to give 1 of the 3 accepted talks from our team at @4Catalyzer. But I couldn't make it bc of ongoing green card application. I will briefly highlight our 3 works here. (thread) >>>
I was going to present "Functional Annotation of Genes Through Integration of Disparate Data Sources with Deep Learning"
We developed an end-to-end trainable framework which performs data integration and functional prediction through a deep learning framework.
We developed an end-to-end trainable framework which performs data integration and functional prediction through a deep learning framework.
Model integration speeds up the convergence and improves the performance of the functional annotation tasks.
Our framework improves the quality of functional annotations by integrating different data sources through cross-modality encoders without handcrafting features.
Our framework improves the quality of functional annotations by integrating different data sources through cross-modality encoders without handcrafting features.
Cross-modality encoding is an alternative way to provide strong priors and mitigate the data limitations by integrating different data sources. The magic is the compressing available genomic, proteomic, and other interrelated datasets through an information bottleneck.
Our approach naturally incorporates incomplete observation at inference time. Therefore, unlike multimodal deep learning models which do not scale with the number of modalities, ours easily scales to a very large number of modalities.
The router is pre-trained via a knowledge base and it learns how to aggregate all information available to enhance the signal of the individual models. Every time a new model is connected to the system, the knowledge base learns richer priors.
Another work from our team is lead by an awesome ML scientist Zhizhuo Zhang called: "Re-inventing Pair-wise Alignment through Deep Learning"
Zhizhuo designed a new architecture called PARN by tweaking LSTM cells and it learns how to align sequences.
Zhizhuo designed a new architecture called PARN by tweaking LSTM cells and it learns how to align sequences.
The problem with the classical alignment algorithms is their disability to automatically adapt to specific biases in your sequence of interest. For example, structural variants, homopolymer bias etc. PARN *learns* how to align instead of forcing it to use pre-defined penalties.
Another nice aspect with PARN is, you do not need to show the correct alignment while you're training. Just some measure of similarity is sufficient and that enables to plug-in this module to down-stream differential learning algorithms for end-to-end training.
Our 3rd project is lead and implemented by the wizard of bioinformatics and information theory expert, Minh Duc Dao, called: "A deep learning model to improve the speed and accuracy of genome assembly"
The model is called Quorum which learns accurate and fast consensus calling.
The model is called Quorum which learns accurate and fast consensus calling.
Consensus calling is very important and we need fast and accurate consensus callers for genome assembly.
After multiple sequence alignment, simple majority rule doesn't work everywhere. But the trick to speed up is, you can easily see where you can get along with majority rule and apply the more complex algorithm to loci where it doesn't work.
In conclusion, Quorum speeds up the assembly pipeline without sacrificing the accuracy.
The talks are in the first session on Monday #ECCB18. And you can also check out the posters:
PARN project by Zhizhuo Zhang: P_Ap076
Quorum project by Minh Duc Cao: P_Ap070
PARN project by Zhizhuo Zhang: P_Ap076
Quorum project by Minh Duc Cao: P_Ap070
