In humans, the >800 G protein-coupled receptors (GPCRs) are responsible for transducing diverse chemical stimuli to alter cell state, and are the largest class of drug targets. Their myriad structural conformations and various modes of signaling make it challenging to understand their structure and function. Here we developed a platform to characterize large libraries of GPCR variants in human cell lines with a barcoded transcriptional reporter of G-protein signal transduction. We tested 7,800 of 7,828 possible single amino acid substitutions to the beta-2 adrenergic receptor (β2AR) at four concentrations of the agonist isoproterenol. We identified residues specifically important for β2AR signaling, mutations in the human population that are potentially loss of function, and residues that modulate basal activity. Using unsupervised learning, we resolve residues critical for signaling, including all major structural motifs and molecular interfaces. We also find a previously uncharacterized structural latch spanning the first two extracellular loops that is highly conserved across Class A GPCRs and is conformationally rigid in both the inactive and active states of the receptor. More broadly, by linking deep mutational scanning with engineered transcriptional reporters, we establish a generalizable method for exploring pharmacogenomics, structure and function across broad classes of drug receptors.

Deep learning models can accurately map genomic DNA sequences to associated functional molecular readouts such as protein-DNA binding data. Base-resolution importance (i.e. "attribution") scores inferred from these models can highlight predictive sequence motifs and syntax. Unfortunately, these models are prone to overfitting and are sensitive to random initializations, often resulting in noisy and irreproducible attributions that obfuscate underlying motifs. To address these shortcomings, we propose a novel attribution prior, where the Fourier transform of input-level attribution scores are computed at training-time, and high-frequency components of the Fourier spectrum are penalized. We evaluate different model architectures with and without attribution priors trained on genome-wide binary or continuous molecular profiles. We show that our attribution prior dramatically improves models' stability, interpretability, and performance on held-out data, especially when training data is severely limited. Our attribution prior also allows models to identify biologically meaningful sequence motifs more sensitively and precisely within individual regulatory elements. The prior is agnostic to the model architecture or predicted experimental assay, yet provides similar gains across all experiments. This work represents an important advancement in improving the reliability of deep learning models for deciphering the regulatory code of the genome.

Seventeen patients with advanced or recurrent salivary gland cancer were treated with cisplatin, doxorubicin, and 5-fluorouracil combination chemotherapy (PAF). Two patients achieved a complete response and four patients achieved a partial response, for an overall response rate of 35%. Six of the nine patients who received PAF in the neoadjuvant setting did not respond and proceeded to surgery and/or radiation therapy. No difference in response rate was found between those patients treated for recurrent disease v those treated with neoadjuvant chemotherapy. All three patients with adenocarcinoma responded. The response duration in patients with metastatic or recurrent disease ranged from 6 to 15 months. The PAF regimen was delivered primarily in the outpatient setting and was associated with acceptable toxicity. PAF demonstrates activity in salivary gland malignancies, and further evaluation of this combination seems warranted.

We are only just beginning to catalog the vast diversity of cell types in the cerebral cortex. Such categorization is a first step toward understanding how diversification relates to function. All cortical projection neurons arise from a uniform pool of progenitor cells that lines the ventricles of the forebrain. It is still unclear how these progenitor cells generate the more than 50 unique types of mature cortical projection neurons defined by their distinct gene-expression profiles. Moreover, exactly how and when neurons diversify their function during development is unknown. Here we relate gene expression and chromatin accessibility of two subclasses of projection neurons with divergent morphological and functional features as they develop in the mouse brain between embryonic day 13 and postnatal day 5 in order to identify transcriptional networks that diversify neuron cell fate. We compare these gene-expression profiles with published profiles of single cells isolated from similar populations and establish that layer-defined cell classes encompass cell subtypes and developmental trajectories identified using single-cell sequencing. Given the depth of our sequencing, we identify groups of transcription factors with particularly dense subclass-specific regulation and subclass-enriched transcription factor binding motifs. We also describe transcription factor-adjacent long noncoding RNAs that define each subclass and validate the function of Myt1l in balancing the ratio of the two subclasses in vitro. Our multidimensional approach supports an evolving model of progressive restriction of cell fate competence through inherited transcriptional identities.

In humans, the 813 G protein-coupled receptors (GPCRs) are responsible for transducing diverse chemical stimuli to alter cell state, and are the largest class of drug targets. Their myriad structural conformations and various modes of signaling make it challenging to understand their structure and function. Here we developed a platform to characterize large libraries of GPCR variants in human cell lines with a barcoded transcriptional reporter of G-protein signal transduction. We tested 7,800 of 7,828 possible single amino acid substitutions to the beta-2 adrenergic receptor ( β 2 AR) at four concentrations of the agonist isoproterenol. We identified residues specifically important for β 2 AR signaling, mutations in the human population that are potentially loss of function, and residues that modulate basal activity. Using unsupervised learning, we resolve residues critical for signaling, including all major structural motifs and molecular interfaces. We also find a previously uncharacterized structural latch spanning the first two extracellular loops that is highly conserved across Class A GPCRs and is conformationally rigid in both the inactive and active states of the receptor. More broadly, by linking deep mutational scanning with engineered transcriptional reporters, we establish a generalizable method for exploring pharmacogenomics, structure and function across broad classes of drug receptors. Fig. 6 A Conserved Extracellular Tryptophan-Disulfide "Structural Latch" in Class A GPCRs is Mutationally Intolerant and Conformation-Independent. A.Sequence conservation of extracellular loop 1 (ECL1) and the extracellular interface of TM3 (202 GPCRs). B. Left: Depiction of the interaction of W99 23x50 , G102 3x21 , and C106 3x25 in ECL1 of the β 2 AR. Top Right: Conservation of the structure of the ECL1 region across functionally different class A GPCRs. Bottom Right: Activity of all 19 missense variants assayed for each of the three conserved residues, with the mean activity (mutational tolerance) shown as a blue bar. The shaded bars represent the mean mutational tolerance ± 1 SD of residues in the tolerant cluster 6 (green) and the intolerant clusters 1 and 2 (red). C. A hydrogen bond network between mutationally intolerant positions W99 23x50 , G102 3x21 , and C106 3x25 . Representative examples of the structural latch are shown. D. This structural latch is maintained in both the inactive and active state structures for the β 2 AR (inactive: 2RH1, active: 3P0G), the M2 muscarinic receptor (inactive: 3UON, active: 4MQS), the kappa-opioid receptor (inactive: 4DJH, active: 6B73), and the mu-opioid receptor (inactive: 4DKL, active: 5C1M)