No abstract

Two atomic level knowledge-based mean force interaction potentials (KBPs), a centrosymmetric burial position term and a long-range pairwise term, were developed. These were tested by comparing multiple configurations of three structurally unrelated proteins and were found successfully to (i) discriminate native state proteins from grossly misfolded structures in inverse folding tests, (ii) rank identify, using the KBP energy/r.m.s.d. correlation, native from progressively less native-like (compact and dilated) structures generated via molecular dynamics sampling, providing an energy gradient sloping from partially unfolded structures towards near-native states in inverse refinement tests (iii) smooth the overall potential energy surface in the region of dilated non-native structures by countering local minima of the in vacuo molecular mechanical potential and (iv) serve as a local minimum detector during simulated temperature quenching studies. These atomic KBPs discriminated native from non-native structures with greater overall sensitivity than did either a residue-based pairwise interaction potential or an effective solvation potential based on atomic contact volume occupancy. The KBPs presented here are immediately useful as a tool for selecting "good refinement candidates' from an arbitrary collection of protein configurations and may play a role in dynamic computational protein refinement.

Simulations of periodic boxes of tetrafluoromethane and trifluoromethane were run to determine van der Waals parameters for fluorine and for hydrogen attached to a fluorine-bearing carbon. The simulations of CF4 were performed first to determine the optimal van der Waals radius R* and well depth E for fluorine by adjusting these parameters to reproduce the experimental molar volume and enthalpy of vaporization of CF4.The best values of R* and E were determined to be 1.75 A and 0.061 kcalimol. Using these fluorine parameters, the simulations of CHF3 were then performed to determine if the hydrogen of this molecule required a smaller R* than that used for the "normal" hydrocarbon hydrogen determined by Spellmeyer and Kollman (results in preparation). That R* was determined by running Monte Carlo simulations on methane, ethane, propane, and butane and adjusting R* and E for carbon and hydrogen to reproduce the experimental molar volume and enthalpy of vaporization. It was found that an R* of 1.21 A was optimal, significantly smaller than the R* = 1.49 A found by Spellmeyer for "normal" hydrocarbon hydrogens. This value of R* is in good agreement with the R* for the hydrogen in CHF3 derived independently using ab initio calculations and molecular mechanics on F3C-H.

No abstract

A fully functional parallel version of the molecular dynamics (MD) module of AMBER3a has been implemented. Procedures parallelized include the calculation of the long-range nonbonded Coulomb and Lennard-Jones interactions, generation of the pairlist, intramolecular bond, angle, dihedral, 1-4 nonbonded interaction terms, coordinate restraints, and the SHAKE bond constraint algorithm. As far as we can determine, this is the f i s t published description where a distributed-memory MIMD parallel implementation of the SHAKE algorithm has been designed to treat not only hydrogen-containing bonds but also all havy-atom bonds, and where "shaken" crQsslinks are supported as well. We discuss the subtasking and partitioning of an MD time-step, load balancing the nonbonded evaluations, describe in algorithmic detail how parallelization of SHAKE was accomplished, and present speedup, efficiency, and benchmarking results achieved when this hypercube adaptation of the MD module AMBER was applied to several variant molecular systems. Results are presented for speedup and efficiency obtained on the nCUBE machine, using up to 128 processors, as well as benchmarks for performance comparisons with the CRAY YMP and F'PS522 vector machines. 0 1993 by John Wiley & Sons, Inc.