Research

Fragment-based drug discovery

Data-driven strategies for fragment library design

Molecular networks

Statistical inference of molecular-network properties

Random-matrix frameworks

Machine learning with filters derived from random-matrix theory

Organic electronics

Multiscale modelling of organic semiconductors and electronic devices

Visualization of a physicochemical reaction "pharmacophore"
Fragment-based drug discovery is an effective and efficient strategy for generating target-optimized drugs. Its success rests in part on careful design of the underlying fragment library. We aim to formulate design rules of what constitutes a “good” fragment by harnessing information across diverse target groups. In doing so, we are developing statistical methods that combine structural modelling, machine learning and statistical filtering to identify relevant patterns in binding data, while making optimal use of the unique structuring of the data. The resulting algorithms and models will provide deeper insight into molecular reactivity, improve the composition of fragment libraries, and boost the in-silico aspects of fragment-based drug discovery.
Visualization of the bipartite graph of olfactory receptors and aroma sensations as inferred by a SOAP kernel
Molecular networks are ensembles of molecules with emergent, collective properties that arise from a set of (at times indirect) interaction pathways between the individual components. They appear across physics, chemistry and biology, for example in the context of energy-harvesting materials, chemical reactions, drug-drug interactions, or biological receptor activity. We develop techniques to model these systems and their properties from the molecular structure upwards. The techniques may be entirely physics-driven whenever the system complexity can be simplified to a degree that direct simulation using quantum-chemical, atomistic or particle-based approaches is possible. However, if the relevant microscopic mechanisms are unknown, or if the system complexity exceeds computational bounds, a data-driven approach may be the only viable route to derive structure-property relationships and thus predict the outcome of new molecular combinations and structures. We develop strategies to derive such statistical approaches, to combine them with physical prior knowledge about interaction mechanisms and pathways, and extract scientific insight from the model predictions.
Inference of mechanism of action across antimicrobial combinations using a filtered projection metric.
Real data is generally affected by noise due to experimental and computational error, statistical fluctuations, mislabelling of data, as well as spurious covariances due to finite sampling. We investigate approaches how to effectively remove sampling bias from data and identify relevant signal components by offsetting the empirical distribution of the high-dimensional features against a predefined null distribution, where properties of the latter can be predicted using random-matrix theory. By embedding these filtering steps in a machine-learning architecture, the resulting random-matrix frameworks are applicable to diverse settings in a generic manner. We are particular interested in molecular informatics and computational biology, including drug discovery and protein function prediction.
Bottom-up workflow for deriving structure-property relationships in organic solar cells, from molecular structure to device characteristics
Organic semiconductors are materials at the interface between hard and soft matter. As such, they combine structural variability, processibility and mechanical flexibility with the ability to efficiently transport charge and energy. This unique set of properties makes them a promising class of materials for electronic devices, including organic solar cells and light-emitting diodes. Our objective is to understand their function at the microscopic scale - a prerequisite for the rational design and optimization of both molecular components and device composition. To this end, we develop multiscale simulation protocols to study the interplay between molecular architecture, supramolecular organization and electronic structure, and relate these microscopic and mesoscopic system properties to the macroscopically observable performance measures - i.e., for solar cells, the open-circuit voltage, but also short-circuit current and fill factor.

Software

BenchML

Develop, optimize, deploy machine-learning pipelines

soap++

Machine-learning library for molecular and materials modelling

librmt

Random-matrix frameworks for statistical modelling

VOTCA

Toolkit for charge- and energy-transport simulations

... to be released soon ... Visit BenchML on GitHub
... started off as a library for atomistic and molecular convolutional descriptors and kernels for machine learning. Next to the successful SOAP descriptor and variants thereof, it implements nonlinear filtering strategies, graph kernels and attribution routines. The library has a C++ core with python bindings, allowing for fast execution and simple extension. Visit SOAP++ on GitHub
... is a python toolkit for the design and deployment of random-matrix frameworks for machine learning. The core of the library implements filters and transformations informed by random-matrix theory. Models are flexibly constructed as pipelines acting on a state object that records data input and output. Visit librmt on GitHub
... (or: Versatile Object-oriented Toolkit for Coarse-graining Applications) started as a C++ software suite for post-processing of molecular-dynamics data. Its two major components are VOTCA-CSG (a library for systematic coarse-graining of particle systems, V. Rühle et al.) and VOTCA-CTP (a library for charge- and energy-transport simulations), recently joined by VOTCA-XTP (an extension for quantum-chemical calculations developed in the group of B. Baumeier). Visit VOTCA on GitHub

Teaching

Teaching

NAN

Computer simulation methods

Monte-Carlo, molecular dynamics and a little bit of machine learning

Login

About

J. D. Bernal with a random-network model of a liquid (ca. 1950)
Computer simulation methods in chemistry and physics This a course on how to solve the many-body problem of statistical mechanics using molecular-dynamics and Monte-Carlo simulations, taught in collaboration with Dr Rosana Collepardo. It is directed at 4th-year students of the Natural Sciences Tripos. Download slides (Part I) Download slides (Part II) - Coming soon
Computer experiments with floating-point numbers. Present affiliation:

Dr Carl Poelking
Department of Chemistry
University of Cambridge
Lensfield Road, CB2 1EW Cambridge
carl.poelking@floatlab.io
Internal pages for users that still remember their passwords Go to login page

Publications

Visit Google Scholar profile

Copyright 2019 | Carl Poelking