Of­f­re 113 sur 598 du 12/09/2019, 10:11


Max Planck Insti­tute of Col­loids and Inter­faces - The­ory & Bio-Sys­tems

The Max Planck Insti­tute of Col­loids and Inter­faces was foun­ded in 1992. Research in Col­loid and Inter­face Sci­ence is widely covered. Cur­rent research top­ics include com­plex car­bo­hydrate molecules, molecu­lar force sensors and motors, meso­scopic hybrid sys­tems, bio­mi­metic mem­branes and ves­icles as well as the develop­ment of car­bo­hydrate-based vac­cines and intel­li­gent bio­ma­ter­i­als

M.Sc. Thesis pro­ject: Machine learn­ing for bio­molecu­lar sim­u­la­tions

Work­ing field:

Phys­ics, bioin­form­at­ics, com­pu­ta­tional sci­ence, engin­eer­ing, math­em­at­ics


You should have solid pro­gram­ming skills, strong interest in model build­ing, and an enthu­si­astic atti­tude towards learn­ing. Good work­ing know­ledge of math­em­at­ics, and exper­i­ence in bio­molecu­lar sim­u­la­tions or machine learn­ing will be con­sidered a plus. Suit­able back­grounds for the pro­ject include phys­ics, engin­eer­ing, bioin­form­at­ics, com­puter sci­ence, and math­em­at­ics. The work­ing lan­guage will be Eng­lish.

What we of­fer:

This pro­ject explores the poten­tial of neural net­works in build­ing bio­molecu­lar mod­els for molecu­lar dynam­ics sim­u­la­tions.

At its best, molecu­lar dynamic (MD) sim­u­la­tions provide real­istic 3D videos (with atom level res­ol­u­tion) on the func­tion­ing of bio­molecules. This makes MD an impact­ful tool in bios­ciences, used by thou­sands of sci­ent­ists and facil­it­at­ing dis­cov­er­ies like the Tami­flu drug res­ist­ance of the H1N1 Swine Flu. At the core of every MD sim­u­la­tion is a force field - a model used to describe the inter­ac­tions between the atoms. The cor­rect­ness of the MD sim­u­la­tion relies entirely on the per­form­ance of the chosen force field.

Unfor­tu­nately, the cur­rent force fields are some­what flawed, and their improve­ment is held back by the com­plex­ity of the prob­lem and old fash­ioned approaches. Devel­op­ment of more effi­cient, high-through­put tools for build­ing bio­molecu­lar force fields is des­per­ately needed.

In this pro­ject, you will achieve this by har­ness­ing the meth­ods used in big data for bio­molecu­lar model devel­op­ment.

A force field is gen­er­ic­ally defined through a set of para­met­ers: a “para­meter vec­tor”. Sim­il­arly, the bio­molecu­lar struc­tures res­ult­ing from the MD sim­u­la­tion can be described as a “res­ult vec­tor”. Your start­ing point will be to treat the prob­lem by using a neural net­work that con­nects the “res­ult vec­tor” to the “para­meter vec­tor”. Once trained, the net­work will be used to pre­dict the “para­meter vec­tor” that is able to repro­duce our exper­i­ment­ally meas­ured struc­tural data. Based on the per­form­ance of this ini­tial approach other machine learn­ing meth­ods may also be explored to improve accur­acy and reduce com­pu­ta­tional cost.

Ini­tially, you will work on lipid sys­tems - an import­ant class of bio­molecules that play a cent­ral role in cell struc­ture and sig­nal­ing. The approach, however, is gen­er­al­iz­able to any bio­molecule.

The pro­ject will be car­ried out in Max Planck Insti­tute of Col­loids and Inter­faces as part of a joint col­lab­or­a­tion between Dr. Markus Miet­tinen, Dr. Angelo Val­leri­ani, and Dr. Hanne Antila.

How to ap­ply:

Inter­ested can­did­ates should send a CV, a let­ter of motiv­a­tion, and a tran­script of uni­versity record via email to by Oct. 1st, 2019.

Please include "Machine learn­ing for bio­molecu­lar sim­u­la­tions" in the email title.

The tent­at­ive start­ing date of the pro­ject is 15.11.2019.