Predicting drug approvals: The Novartis data science and artificial intelligence challenge.

We describe a novel collaboration between academia and industry, an in-house data science and artificial intelligence challenge held by Novartis to develop machine-learning models for predicting drug-development outcomes, building upon research at MIT using data from Informa as the starting point. With over 50 cross-functional teams from 25 Novartis offices around the world participating in the challenge, the domain expertise of these Novartis researchers was leveraged to create predictive models with greater sophistication. Ultimately, two winning teams developed models that outperformed the baseline MIT model-areas under the curve of 0.88 and 0.84 versus 0.78, respectively-through state-of-the-art machine-learning algorithms and the use of newly incorporated features and data. In addition to validating the variables shown to be associated with drug approval in the earlier MIT study, the challenge also provided new insights into the drivers of drug-development success and failure.

A large-scale, in vivo transcription factor screen defines bivalent chromatin as a key property of regulatory factors mediating Drosophila wing development.

Transcription factors (TFs) are key regulators of cell fate. The estimated 755 genes that encode DNA binding domain-containing proteins comprise ∼ 5% of all Drosophila genes. However, the majority has remained uncharacterized so far due to the lack of proper genetic tools. We generated 594 site-directed transgenic Drosophila lines that contain integrations of individual UAS-TF constructs to facilitate spatiotemporally controlled misexpression in vivo. All transgenes were expressed in the developing wing, and two-thirds induced specific phenotypic defects. In vivo knockdown of the same genes yielded a phenotype for 50%, with both methods indicating a great potential for misexpression to characterize novel functions in wing growth, patterning, and development. Thus, our UAS-TF library provides an important addition to the genetic toolbox of Drosophila research, enabling the identification of several novel wing development-related TFs. In parallel, we established the chromatin landscape of wing imaginal discs by ChIP-seq analyses of five chromatin marks and RNA Pol II. Subsequent clustering revealed six distinct chromatin states, with two clusters showing enrichment for both active and repressive marks. TFs that carry such "bivalent" chromatin are highly enriched for causing misexpression phenotypes in the wing, and analysis of existing expression data shows that these TFs tend to be differentially expressed across the wing disc. Thus, bivalently marked chromatin can be used as a marker for spatially regulated TFs that are functionally relevant in a developing tissue.

Rationally designed turn promoting mutation in the amyloid-ß peptide sequence stabilizes oligomers in solution.

Enhanced production of a 42-residue beta amyloid peptide (Aß(42)) in affected parts of the brain has been suggested to be the main causative factor for the development of Alzheimer's Disease (AD). The severity of the disease depends not only on the amount of the peptide but also its conformational transition leading to the formation of oligomeric amyloid-derived diffusible ligands (ADDLs) in the brain of AD patients. Despite being significant to the understanding of AD mechanism, no atomic-resolution structures are available for these species due to the evanescent nature of ADDLs that hinders most structural biophysical investigations. Based on our molecular modeling and computational studies, we have designed Met35Nle and G37p mutations in the Aß(42) peptide (Aß(42)Nle35p37) that appear to organize Aß(42) into stable oligomers. 2D NMR on the Aß(42)Nle35p37 peptide revealed the occurrence of two ß-turns in the V24-N27 and V36-V39 stretches that could be the possible cause for the oligomer stability. We did not observe corresponding NOEs for the V24-N27 turn in the Aß(21-43)Nle35p37 fragment suggesting the need for the longer length amyloid peptide to form the stable oligomer promoting conformation. Because of the presence of two turns in the mutant peptide which were absent in solid state NMR structures for the fibrils, we propose, fibril formation might be hindered. The biophysical information obtained in this work could aid in the development of structural models for toxic oligomer formation that could facilitate the development of therapeutic approaches to AD.

The predicted structure of the headpiece of the Huntingtin protein and its implications on Huntingtin aggregation.

We have performed simulated tempering molecular dynamics simulations to study the thermodynamics of the headpiece of the Huntingtin (Htt) protein (N17(Htt)). With converged sampling, we found this peptide is highly helical, as previously proposed. Interestingly, this peptide is also found to adopt two different and seemingly stable states. The region from residue 4 (L) to residue 9 (K) has a strong helicity from our simulations, which is supported by experimental studies. However, contrary to what was initially proposed, we have found that simulations predict the most populated state as a two-helix bundle rather than a single straight helix, although a significant percentage of structures do still adopt a single linear helix. The fact that Htt aggregation is nucleation dependent infers the importance of a critical transition. It has been shown that N17(Htt) is involved in this rate-limiting step. In this study, we propose two possible mechanisms for this nucleating event stemming from the transition between two-helix bundle state and single-helix state for N17(Htt) and the experimentally observed interactions between the N17(Htt) and polyQ domains. More strikingly, an extensive hydrophobic surface area is found to be exposed to solvent in the dominant monomeric state of N17(Htt). We propose the most fundamental role played by N17(Htt) would be initializing the dimerization and pulling the polyQ chains into adequate spatial proximity for the nucleation event to proceed.

Simulating oligomerization at experimental concentrations and long timescales: A Markov state model approach.

Here, we present a novel computational approach for describing the formation of oligomeric assemblies at experimental concentrations and timescales. We propose an extension to the Markovian state model approach, where one includes low concentration oligomeric states analytically. This allows simulation on long timescales (seconds timescale) and at arbitrarily low concentrations (e.g., the micromolar concentrations found in experiments), while still using an all-atom model for protein and solvent. As a proof of concept, we apply this methodology to the oligomerization of an Abeta peptide fragment (Abeta(21-43)). Abeta oligomers are now widely recognized as the primary neurotoxic structures leading to Alzheimer's disease. Our computational methods predict that Abeta trimers form at micromolar concentrations in 10 ms, while tetramers form 1000 times more slowly. Moreover, the simulation results predict specific intermonomer contacts present in the oligomer ensemble as well as putative structures for small molecular weight oligomers. Based on our simulations and statistical models, we propose a novel mutation to stabilize the trimeric form of Abeta in an experimentally verifiable manner.

Ensemble molecular dynamics yields submillisecond kinetics and intermediates of membrane fusion.

Lipid membrane fusion is critical to cellular transport and signaling processes such as constitutive secretion, neurotransmitter release, and infection by enveloped viruses. Here, we introduce a powerful computational methodology for simulating membrane fusion from a starting configuration designed to approximate activated prefusion assemblies from neuronal and viral fusion, producing results on a time scale and degree of mechanistic detail not previously possible to our knowledge. We use an approach to the long time scale simulation of fusion by constructing a Markovian state model with large-scale distributed computing, yielding an understanding of fusion mechanisms on time scales previously impossible to simulate to our knowledge. Our simulation data suggest a branched pathway for fusion, in which a common stalk-like intermediate can either rapidly form a fusion pore or remain in a metastable hemifused state that slowly forms fully fused vesicles. This branched reaction pathway provides a mechanistic explanation both for the biphasic fusion kinetics and the stable hemifused intermediates previously observed experimentally. Our distributed computing and Markovian state model approaches provide sufficient sampling to detect rare transitions, a systematic process for analyzing reaction pathways, and the ability to develop quantitative approximations of reaction kinetics for fusion.