Overview of the Biological Language Modeling Project

i. Discern property conservation of amino acids (e.g. discriminating hydrophobic amino acids facing outside vs. inside of transmembrane helices)

ii. Identify the most salient motifs for structural stability and for governing conformational changes, building on our existing results for rhodopsin and leveraging context-dependent statistical grammars or similar approaches.

iii. Discriminate between functional-relevant and structurally-relevant motifs.

iv. IMPACT: New strategies to help diagnose and eventually treat conformational diseases associated with transmembrane proteins.

v. EVALUATION: site-directed mutagenesis experiments.

i. Extension to supersecondary structures via long-range probabilistic linguistic models

ii. Explore b-sheet transmembrane proteins (incl b-barrels)

iii. BIOLOGICAL IMPACT: Predictive models for b-sheets and selected b-sheet based supersecondary structures

iv. BIOLOGICAL EVALUATION: Reduce prediction errors by up to 50% for existing prediction tasks, and predict structures for which there are no present prediction results.

v. COMPUTATIONAL IMPACT: New predictive algorithms and techniques such as multi-layer conditional random fields (CRFs)

vi. COMPUTATIONAL EVALUATION: Performance measured on new problems in relations text extraction and understanding

i. Based on multi-dimensional position-conditional properties

ii. Expansion from HIV and GPCR families to other families including kinases and nuclear receptors

iii. Further expansion to broader datasets including all possible protein families, and structurally homologous protein families

iv. IMPACT: Detect very remote homologs.

v. EVALUATION: Identification of heretofore unknown homologs, to be validated via biological experimentation.

i.Developing new algorithms to extract multi-type expressions of genes from micro-array data, DNA sequences, protein-protein interaction databases and gene ontology, and to induce regulatory networks based on the multi-types of evidence

ii. Adapting hierarchical classification techniques to the motif/gene identification and function analysis based on multi-abstraction-level
representations of protein sequences

iii. Improving the efficiency of machine learning algorithms for automated induction of very large regulatory networks, e.g., with thousands of genes

iv. IMPACT: The resulting regulatory networks and gene/motif classes would help biologists to discover new and interesting patterns, e.g. leading to deeper understanding of the mechanisms regulating oncogenes and their potential disruption.

v. EVALUATION: Identification of regulatory networks and gene classes (functions), to be validated first via biology databases (TRANSFAC, GO, SCPD...), and later possibly via collaboration with personnel in the U Pitt Cancer Center.