Iccs9th - Do Protein Targets segregates?

Do Targets Segregate?

Andrea Zaliani

A. Zaliani 9thICCS 2

Aim• Bioinformaticians were able to segregate protein

targets by several means from 1D to 3D and 4D• We have potent means to perform same analysis from

ligand standpoint:o Fingerprint (e.g. 2D,3D, interactionFP, etc)o Shape Descriptorso Grid

• Do we appreciate their peculiarities?• Would our structural knowledge grow, if we knew some

frequent target-directing structural pattern?


Start – Method• Plenty of late work trying to link protein structures,

functions and cavities to ligands (and vice versa) through similarity concepts

• I would here stress not new methods but what we have already in our hands to boost ideas with couple of applications with freely available software (like KNIME, R)

• FP = Do we appreciate their peculiarities enough?• Can we look into statistical models? If yes, do we?


Different FingerPrint (FP) for different

scopes

• Can FP explain us this? FP Type Tan-Distance

MW,LogP,HA(CDK)… 0.000Layered(RDKit) 0.082AtomPairs (RDKit) 0.098Indigo(GGA) 0.190Morgan(RDKit) 0.302FeatMorgan(RDKit) 0.348ErG* 0.375

Similarity ≈ 0.62-65*N. Stiefl et al. JCIM.,(2006), 46(1)208; N. Stiefl et al. JCIM, (2006), 46(2)587


ErG = pharmacophore-fingerprintDevelopment of ErG (Extended reduced Graph), a 2D-pharmacophoric similarity tool for virtual screening ErG is much less substructure-dependent so that:

•Opens opportunities in library design (scaffold-hopping)•Multiple-to-one correspondence of chemical substructures to pharmacophoric patterns ‘abstract’•Similarity searching & ‘scaffold-hopping’ documented•FP interpretable as each bit corresponds to the count of pharmacophore pair distances in graph

•Atom types [6] generate pairs [21] x max_distance [15] = 315 bits

Graph

N

N

Ac

D+

Ac

D+

Hf

Ac

D+

Hf

Ar Ar

Charge / H-Bonding

Hydrophobic endcaps

Abstract ring forms

Ac

D+

Hf

Ar Ar

N

N

Ac

D+

Ac

D+

Hf

Ac

D+

Hf

Ar Ar

Charge / H-Bonding

Hydrophobic endcaps

Abstract ring forms

Ac

D+

Hf

Ar Ar

RDF vectorization

AcAcd1,AcAcd2,…,AcDod4,…,ArHfd4,…..,+-d15Cpd_A,0, 0, …,1, …,1, …,0


Experiment plan - Dataset

• From a literature database select a relevant random subset (ca.17K) literature compounds showing at least one activity (pEx50>6) towards a precise target among class families like GPCR-A, Kinases, Proteases or NHR

• Data are high quality in terms of consistency• Less than 5% of entire Pharma Database of Evolvus• To check homogeneity all vs. all similarity evaluation

with TanDistance under different FP…..


Liceptor Database

Targets Annotated• GPCR’s• Ion- Channels• CNS Transporters• Kinases• Proteases• Phosphatases

Client Proprietary Targets

Small Molecule Ligand Database Features

Liceptor database can be customized with client specified additional fields and custom data annotation

• 3.2 Million Structures• > 1000 Targets• Global Patents• Med Chem. Journals• Data annotated from 1967 • Multiple Target Data• 2D Structures• Molecular Descriptors• IC50 and Unified Values• Therapeutic Indications


Pharmacophore-based FP better

RDKit FP RDKit Feature Morgan FP


Experiment plan - Dataset


Experiment plan – Classification Model• Partition Tree model generated• Platform (KIN, GPCRA, NHR, PROT) can be

predicted with 15 ErG distances only• If shuffled on Y, models generated with ave

errors ranging 63-77% (100x)• External predictions at 82,6%


Target Family Classification Model


Learn from missclassified• 15 Distances enough to segregate 17K

compounds in four classes• From model some insights can be extracted:• Example KIN relevant features:

i. Presence of Ar-NH(OH) [DoArd1>0]ii. Absence of a-aminoacid signature

AcDod3 =0iii. Need of AcArd3 >0 if i. applies or =1

6H-Benzo[c]chromen-6-one derivatives as selective ERβ agonistsBioorganic & Medicinal Chemistry Letters 16, (6), 2006, Pages 1468-1472


Learn from missclassified• 15 Distances enough to segregate 17K

compounds in four classes• From model some insights can be extracted:• Example KIN relevant features:

i. Presence of Ar-NH(OH) [DoArd1>0]ii. Absence of a-aminoacid signature

AcDod3 =0iii. Need of AcArd3 >0 if i. applies or =1


Classification Model – What to learn• 15 Distances enough to segregate 17K

compounds in four classes• From model some insights can be extracted:• PROTEASE Target relevant features:

i. Presence of AA signature AcDoD3

ii. Presence of AcArd3

iii. Absence/Presence of max 1 HfArd4

Hf

Ar


Classification Model – How do we use this• We can try to use these as smarts query into PDB

http://www.pdb.org/pdb/search/advSearch.do • PROTEASE Target relevant features:

i. Presence of AA signature AcDoD3ii. Presence of AcArd3iii. Presence of max 1 HfArd4

• Results of query after removal of non polypeptide, solvents, chain duplicates

• 101 complexes of which 53% correct proteases

• If only i.&iii. Were used, then 1141 hits found with 738 protease complexes (65%) retrieved


Single Family Classification Models• Each Target Family could also be modeled through

classification• KNIME offers several functions for:

o Data preparationo Training/Test split with stratification on populationo Data reduction performed with an exhaustive retrograde selectiono Cross-validation with 100X Leave-10%-outo Shuffled-Y 100 classification models built for negative testo Performance statistics given on 25% external test set


Classification Model – NHR

HX

Ar



Ave. Distance Profiles

A. Zaliani 9thICCS 190,00 0,10 0,20 0,30 0,40 0,50 0,60 0,70 0,80 0,90 1,00



Classification Model – Kinase







Classification Model – GPCRA






26

Lessons learned here• QC-based database essential • 2D Pharmacophoric FP approach is enough but has to be

“understood”• Making FP less cryptic help understanding potentialities and

limits• Targets do segregate. Ligands help us realizing this, the more

the more precise• Pharmacophoric Graph Space is immensely less problematic

than chemical space• Provocation: how big is graph space of IP?

A. Zaliani 9thICCS


Limitations• Question: you find what you already know?• Question: Do abstraction help us? • Every FP method is ok, provided that teaches us

something• Promiscuity reduction is not the only final aim

(controlled promiscuity might be a need)• Graph distances might be too general• 2D Pharmacophoric fingerprinting to be improved


Future work• 3D distances (3Dtriangles) could easily implemented• Combinations of ligand FP and cavity FP could be

really a breakthrough to have a grip on multi-pharmacology

• FP Weighting for atomic de-solvation contribution is, for me, KEY

• Agonist/antagonist split• pEX50 >6 will provide different pictures?


Acknowledgements

Prof. M. Berthold

Greg LandrumNik Stiefl

Aniket Ausekar, CEOVikram Palshikar

Rashmi Jain

Mike Bodkin


Appendix


Approach to Polypharmacology• Pharmacophore target family mapping using Neural Networks (Kohonen)• Cpds mapped together with annotated actives from different sources (MDDR, UBI, etc.)• Clustering method to suggest pharmacophore similarity (Ext.Reduced Graphs

fingerprint)SOM Binary ErG on 9444 cpds with pIC50>8

pIC50_8_SOM8_8_1M_Z (x value)0 1 2 3 4 5 6 7

0

1

2

3

4

5

6

7

Protease GCPRa Kinases NHR Transporter

Neuron 7,3775cpds from

different families

NN

S

OO

N N 2425712pIC50(PR)=8.79

N

Cl

N

N O

O450207pIC50(NPY_V)=8.79

Presentations & Public Speaking

Iccs9th - Do Protein Targets segregates?