LM-GVP: an extensible sequence and structure informed deep learning framework for protein property prediction.

Proteins perform many essential functions in biological systems and can be successfully developed as bio-therapeutics. It is invaluable to be able to predict their properties based on a proposed sequence and structure. In this study, we developed a novel generalizable deep learning framework, LM-GVP, composed of a protein Language Model (LM) and Graph Neural Network (GNN) to leverage information from both 1D amino acid sequences and 3D structures of proteins. Our approach outperformed the state-of-the-art protein LMs on a variety of property prediction tasks including fluorescence, protease stability, and protein functions from Gene Ontology (GO). We also illustrated insights into how a GNN prediction head can inform the fine-tuning of protein LMs to better leverage structural information. We envision that our deep learning framework will be generalizable to many protein property prediction problems to greatly accelerate protein engineering and drug development.

Retrospective clinical trial experimentally validates glioblastoma genome-wide pattern of DNA copy-number alterations predictor of survival.

Modeling of genomic profiles from the Cancer Genome Atlas (TCGA) by using recently developed mathematical frameworks has associated a genome-wide pattern of DNA copy-number alterations with a shorter, roughly one-year, median survival time in glioblastoma (GBM) patients. Here, to experimentally test this relationship, we whole-genome sequenced DNA from tumor samples of patients. We show that the patients represent the U.S. adult GBM population in terms of most normal and disease phenotypes. Intratumor heterogeneity affects ≈ 11 % and profiling technology and reference human genome specifics affect <1% of the classifications of the tumors by the pattern, where experimental batch effects normally reduce the reproducibility, i.e., precision, of classifications based upon between one to a few hundred genomic loci by >30%. With a 2.25-year Kaplan-Meier median survival difference, a 3.5 univariate Cox hazard ratio, and a 0.78 concordance index, i.e., accuracy, the pattern predicts survival better than and independent of age at diagnosis, which has been the best indicator since 1950. The prognostic classification by the pattern may, therefore, help to manage GBM pseudoprogression. The diagnostic classification may help drugs progress to regulatory approval. The therapeutic predictions, of previously unrecognized targets that are correlated with survival, may lead to new drugs. Other methods missed this relationship in the roughly 3B-nucleotide genomes of the small, order of magnitude of 100, patient cohorts, e.g., from TCGA. Previous attempts to associate GBM genotypes with patient phenotypes were unsuccessful. This is a proof of principle that the frameworks are uniquely suitable for discovering clinically actionable genotype-phenotype relationships.

GSVD- and tensor GSVD-uncovered patterns of DNA copy-number alterations predict adenocarcinomas survival in general and in response to platinum.

More than a quarter of lung, uterine, and ovarian adenocarcinoma (LUAD, USEC, and OV) tumors are resistant to platinum drugs. Only recently and only in OV, patterns of copy-number alterations that predict survival in response to platinum were discovered, and only by using the tensor GSVD to compare Agilent microarray platform-matched profiles of patient-matched normal and primary tumor DNA. Here, we use the GSVD to compare whole-genome sequencing (WGS) and Affymetrix microarray profiles of patient-matched normal and primary LUAD, USEC, and OV tumor DNA. First, the GSVD uncovers patterns similar to one Agilent OV pattern, where a loss of most of the chromosome arm 6p combined with a gain of 12p encode for transformation. Like the Agilent OV pattern, the WGS LUAD and Affymetrix LUAD, USEC, and OV patterns are correlated with shorter survival, in general and in response to platinum. Like the tensor GSVD, the GSVD separates these tumor-exclusive genotypes from experimental inconsistencies. Second, by identifying the shorter survival phenotypes among the WGS- and Affymetrix-profiled tumors, the Agilent pattern proves to be a technology-independent predictor of survival, independent also of the best other indicator at diagnosis, i.e., stage. Third, like no other indicator, the pattern predicts the overall survival of OV patients experiencing progression-free survival, in general and in response to platinum. We conclude that comparative spectral decompositions, such as the GSVD and tensor GSVD, underlie a mathematically universal description of the relationships between a primary tumor's genotype and a patient's overall survival phenotype, which other methods miss.

Mathematically universal and biologically consistent astrocytoma genotype encodes for transformation and predicts survival phenotype.

DNA alterations have been observed in astrocytoma for decades. A copy-number genotype predictive of a survival phenotype was only discovered by using the generalized singular value decomposition (GSVD) formulated as a comparative spectral decomposition. Here, we use the GSVD to compare whole-genome sequencing (WGS) profiles of patient-matched astrocytoma and normal DNA. First, the GSVD uncovers a genome-wide pattern of copy-number alterations, which is bounded by patterns recently uncovered by the GSVDs of microarray-profiled patient-matched glioblastoma (GBM) and, separately, lower-grade astrocytoma and normal genomes. Like the microarray patterns, the WGS pattern is correlated with an approximately one-year median survival time. By filling in gaps in the microarray patterns, the WGS pattern reveals that this biologically consistent genotype encodes for transformation via the Notch together with the Ras and Shh pathways. Second, like the GSVDs of the microarray profiles, the GSVD of the WGS profiles separates the tumor-exclusive pattern from normal copy-number variations and experimental inconsistencies. These include the WGS technology-specific effects of guaninecytosine content variations across the genomes that are correlated with experimental batches. Third, by identifying the biologically consistent phenotype among the WGS-profiled tumors, the GBM pattern proves to be a technology-independent predictor of survival and response to chemotherapy and radiation, statistically better than the patient's age and tumor's grade, the best other indicators, and MGMT promoter methylation and IDH1 mutation. We conclude that by using the complex structure of the data, comparative spectral decompositions underlie a mathematically universal description of the genotype-phenotype relations in cancer that other methods miss.

A higher-order generalized singular value decomposition for comparison of global mRNA expression from multiple organisms.

The number of high-dimensional datasets recording multiple aspects of a single phenomenon is increasing in many areas of science, accompanied by a need for mathematical frameworks that can compare multiple large-scale matrices with different row dimensions. The only such framework to date, the generalized singular value decomposition (GSVD), is limited to two matrices. We mathematically define a higher-order GSVD (HO GSVD) for N≥2 matrices D(i)∈R(m(i) × n), each with full column rank. Each matrix is exactly factored as D(i)=U(i)Σ(i)V(T), where V, identical in all factorizations, is obtained from the eigensystem SV=VΛ of the arithmetic mean S of all pairwise quotients A(i)A(j)(-1) of the matrices A(i)=D(i)(T)D(i), i≠j. We prove that this decomposition extends to higher orders almost all of the mathematical properties of the GSVD. The matrix S is nondefective with V and Λ real. Its eigenvalues satisfy λ(k)≥1. Equality holds if and only if the corresponding eigenvector v(k) is a right basis vector of equal significance in all matrices D(i) and D(j), that is σ(i,k)/σ(j,k)=1 for all i and j, and the corresponding left basis vector u(i,k) is orthogonal to all other vectors in U(i) for all i. The eigenvalues λ(k)=1, therefore, define the "common HO GSVD subspace." We illustrate the HO GSVD with a comparison of genome-scale cell-cycle mRNA expression from S. pombe, S. cerevisiae and human. Unlike existing algorithms, a mapping among the genes of these disparate organisms is not required. We find that the approximately common HO GSVD subspace represents the cell-cycle mRNA expression oscillations, which are similar among the datasets. Simultaneous reconstruction in the common subspace, therefore, removes the experimental artifacts, which are dissimilar, from the datasets. In the simultaneous sequence-independent classification of the genes of the three organisms in this common subspace, genes of highly conserved sequences but significantly different cell-cycle peak times are correctly classified.