OpenFold: retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization.

AlphaFold2 revolutionized structural biology with the ability to predict protein structures with exceptionally high accuracy. Its implementation, however, lacks the code and data required to train new models. These are necessary to (1) tackle new tasks, like protein-ligand complex structure prediction, (2) investigate the process by which the model learns and (3) assess the model's capacity to generalize to unseen regions of fold space. Here we report OpenFold, a fast, memory efficient and trainable implementation of AlphaFold2. We train OpenFold from scratch, matching the accuracy of AlphaFold2. Having established parity, we find that OpenFold is remarkably robust at generalizing even when the size and diversity of its training set is deliberately limited, including near-complete elisions of classes of secondary structure elements. By analyzing intermediate structures produced during training, we also gain insights into the hierarchical manner in which OpenFold learns to fold. In sum, our studies demonstrate the power and utility of OpenFold, which we believe will prove to be a crucial resource for the protein modeling community.

Standard machine learning approaches outperform deep representation learning on phenotype prediction from transcriptomics data.

BACKGROUND: The ability to confidently predict health outcomes from gene expression would catalyze a revolution in molecular diagnostics. Yet, the goal of developing actionable, robust, and reproducible predictive signatures of phenotypes such as clinical outcome has not been attained in almost any disease area. Here, we report a comprehensive analysis spanning prediction tasks from ulcerative colitis, atopic dermatitis, diabetes, to many cancer subtypes for a total of 24 binary and multiclass prediction problems and 26 survival analysis tasks. We systematically investigate the influence of gene subsets, normalization methods and prediction algorithms. Crucially, we also explore the novel use of deep representation learning methods on large transcriptomics compendia, such as GTEx and TCGA, to boost the performance of state-of-the-art methods. The resources and findings in this work should serve as both an up-to-date reference on attainable performance, and as a benchmarking resource for further research. RESULTS: Approaches that combine large numbers of genes outperformed single gene methods consistently and with a significant margin, but neither unsupervised nor semi-supervised representation learning techniques yielded consistent improvements in out-of-sample performance across datasets. Our findings suggest that using l2-regularized regression methods applied to centered log-ratio transformed transcript abundances provide the best predictive analyses overall. CONCLUSIONS: Transcriptomics-based phenotype prediction benefits from proper normalization techniques and state-of-the-art regularized regression approaches. In our view, breakthrough performance is likely contingent on factors which are independent of normalization and general modeling techniques; these factors might include reduction of systematic errors in sequencing data, incorporation of other data types such as single-cell sequencing and proteomics, and improved use of prior knowledge.

OpenProteinSet: Training data for structural biology at scale.

Multiple sequence alignments (MSAs) of proteins encode rich biological information and have been workhorses in bioinformatic methods for tasks like protein design and protein structure prediction for decades. Recent breakthroughs like AlphaFold2 that use transformers to attend directly over large quantities of raw MSAs have reaffirmed their importance. Generation of MSAs is highly computationally intensive, however, and no datasets comparable to those used to train AlphaFold2 have been made available to the research community, hindering progress in machine learning for proteins. To remedy this problem, we introduce OpenProteinSet, an open-source corpus of more than 16 million MSAs, associated structural homologs from the Protein Data Bank, and AlphaFold2 protein structure predictions. We have previously demonstrated the utility of OpenProteinSet by successfully retraining AlphaFold2 on it. We expect OpenProteinSet to be broadly useful as training and validation data for 1) diverse tasks focused on protein structure, function, and design and 2) large-scale multimodal machine learning research.

Cav1.2 L-type Ca²âº channels mediate cocaine-induced GluA1 trafficking in the nucleus accumbens, a long-term adaptation dependent on ventral tegmental area Ca(v)1.3 channels.

AMPA receptor (AMPAR) plasticity at glutamatergic synapses in the mesoaccumbal dopaminergic pathway has been implicated in persistent cocaine-induced behavioral responses; however, the precise mechanism underlying these changes remains unknown. Utilizing cocaine psychomotor sensitization, we have examined phosphorylation of GluA1 at key residues serine 845 (S845) and S831, as well as GluA1 cell surface levels in the nucleus accumbens (NAc) of cocaine-preexposed mice and the role of brain-specific Ca(v)1.2 and Ca(v)1.3 L-type Ca²âº channels (LTCCs), therein. We found higher basal levels of S845 phospho-GluA1 (P-GluA1) and cell surface GluA1 in the NAc following protracted withdrawal from cocaine exposure, changes that occur independently of LTCCs. In contrast, we found that a cocaine challenge that elicits expression of the cocaine-sensitized response increases S831 P-GluA1 that further increases surface GluA1 beyond the higher basal levels. Intra-NAc pharmacological manipulations indicate that the Ca(v)1.2-activated CaM kinase II (CaMKII) mediates cocaine-induced increase in S831 P-GluA1 and that both Ca(v)1.2-activated CaMKII and extracellular signal-regulated kinase 2 (ERK2) mediate the increase in GluA1 cell surface levels specific to the sensitized response. Experiments using adenoassociated viral vectors expressing Ca(v)1.3 and ERK2 siRNA further indicate that recruitment of the Ca(v)1.2 pathway in the NAc is dependent on ventral tegmental area Ca(v)1.3 LTCCs and ERK2. Together, these results identify candidate pathways that mediate cocaine-induced AMPAR plasticity in the NAc and provide a mechanism linking LTCCs and GluA1 plasticity to cocaine-induced persistent behavioral changes.

FAAH genetic variation enhances fronto-amygdala function in mouse and human.

Cross-species studies enable rapid translational discovery and produce the broadest impact when both mechanism and phenotype are consistent across organisms. We developed a knock-in mouse that biologically recapitulates a common human mutation in the gene for fatty acid amide hydrolase (FAAH) (C385A; rs324420), the primary catabolic enzyme for the endocannabinoid anandamide. This common polymorphism impacts the expression and activity of FAAH, thereby increasing anandamide levels. Here, we show that the genetic knock-in mouse and human variant allele carriers exhibit parallel alterations in biochemisty, neurocircuitry and behaviour. Specifically, there is reduced FAAH expression associated with the variant allele that selectively enhances fronto-amygdala connectivity and fear extinction learning, and decreases anxiety-like behaviours. These results suggest a gain of function in fear regulation and may indicate for whom and for what anxiety symptoms FAAH inhibitors or exposure-based therapies will be most efficacious, bridging an important translational gap between the mouse and human.

Behavioral characterization of cereblon forebrain-specific conditional null mice: a model for human non-syndromic intellectual disability.

A nonsense mutation in the human cereblon gene (CRBN) causes a mild type of autosomal recessive non-syndromic intellectual disability (ID). Animal studies show that crbn is a cytosolic protein with abundant expression in the hippocampus (HPC) and neocortex (CTX). Its diverse functions include the developmental regulation of ion channels at the neuronal synapse, the mediation of developmental programs by ubiquitination, and a target for herpes simplex type I virus in HPC neurons. To test the hypothesis that anomalous CRBN expression leads to HPC-mediated memory and learning deficits, we generated germ-line crbn knock-out mice (crbn(-/-)). We also inactivated crbn in forebrain neurons in conditional knock-out mice in which crbn exons 3 and 4 are deleted by cre recombinase under the direction of the Ca(2+)/calmodulin-dependent protein kinase II alpha promoter (CamKII(cre/+), crbn(-/-)). crbn mRNA levels were negligible in the HPC, CTX, and cerebellum (CRBM) of the crbn(-/-) mice. In contrast, crbn mRNA levels were reduced 3- to 4-fold in the HPC, CTX but not in the CRBM in CamKII(cre/+), crbn(-/-) mice as compared to wild type (CamKII(cre/+), crbn(+/+)). Contextual fear conditioning showed a significant decrease in the percentage of freezing time in CamKII(cre/+), crbn(-/-) and crbn(-/-) mice while motor function, exploratory motivation, and anxiety-related behaviors were normal. These findings suggest that CamKII(cre/+), crbn(-/-) mice exhibit selective HPC-dependent deficits in associative learning and supports the use of these mice as in vivo models to study the functional consequences of CRBN aberrations on memory and learning in humans.