Pharmacoinformatics Research Group Department of Pharmaceutical Chemistry Topliss batchwise scheme reviewed in the era of Open Data Lars Richter, Gerhard F. Ecker Dept. of Pharmaceutical Chemistry [email protected] pharminfo.univie.ac.at Topliss batchwise scheme Topliss substituent proposals Topliss ranking schemes subst. π 3,4-Cl2 4-Cl 4-CH3 4-OCH3 H 1Topliss 1 σ -σ π+σ Es scheme new substituent selection 1 1 5 1 2-5 π 2 2 4 2 2-5 3-CF3, 4-Cl; 3-CF3, 4-NO2; 4CF3; 2,4-C12; 4-c-C5H9; 4-cC6H11; 4-CH(CH3)2; 4-C(CH3)3; 3,4-(CH3)2; 4-O(CH3),CH3; 4OCH2Ph; 4-N(C2H5) σ 3-CF3, 4-Cl; 3-CF3, 4-NO2; 4CF3; 2,4-C12; 4-c-C5H9; 4-cC6H11 3 4-5 4-5 4 5 3 2 1 3 et al. J Med Chem 1977 3 5 4 2-5 2-5 1 -σ π+σ 4-N(C2H5)2; 4-N(CH3)2; 4-NH2; 4-NHC4H9; 4-OH; 4OCH(CH3)2; 3-CH3,4-OCH3 3-CF3, 4-Cl; 3-CF3, 4-NO2; 4CF3; 2,4-C12; 4-c-C5H9; 4-cC6H11 Topliss batchwise scheme Series of five phenyl-substituted propafenone derivatives measured against P-Glycoprotein substituion EC50 rank 3,4-Cl2 0.150 5 4-Cl 0.132 4 4-CH3 0.063 2 4-OCH3 0.045 1 H 0.079 3 Which compound should be synthesized next? Topliss batchwise scheme Topliss propafenone dataset substituent EC rank proposals 3,4-Cl 0.150 5 substituion Topliss ranking schemes 50 2 subst. π σ -σ π+σ Es scheme -σπ 3,4-Cl2 1 1 5 1 2-5 4-Cl 2 2 4 2 2-5 4-CH3 4-OCH3 H 3 4-5 4-5 4 5 3 2 1 3 3 5 4 2-5 2-5 1 σ -σ π+σ 4-Cl substituent 0.132 selection 4 new 4-CH3 0.063 2 3-CF3, 4-Cl; 3-CF3, 4-NO2; 44-OCH3 0.045 1 CF3; 2,4-C12; 4-c-C5H9; 4-cC6H11; 4-CH(CH3)2; 4-C(CH3)3; H 0.079 3 3,4-(CH3)2; 4-O(CH3),CH3; 4OCH2Ph; 4-N(C2H5) 3-CF3, 4-Cl; 3-CF3, 4-NO2; 4CF3; 2,4-C12; 4-c-C5H9; 4-cC6H11 4-N(C2H5)2; 4-N(CH3)2; 4-NH2; 4-NHC4H9; 4-OH; 4OCH(CH3)2; 3-CH3,4-OCH3 3-CF3, 4-Cl; 3-CF3, 4-NO2; 4CF3; 2,4-C12; 4-c-C5H9; 4-cC6H11 Topliss batchwise scheme propafenone dataset -σ substituion EC50 rank 3,4-Cl2 0.150 5 4-Cl 0.132 4 4-CH3 0.063 2 4-OCH3 0.045 1 H 0.079 3 • 4-N(CH3)2 derivative was synthesized and tested • no affinity increase 4-N(CH3)2 How often do Topliss schemes (π, σ, -σ, π+σ, Es) occur in large databases? How useful do Topliss schemes prove in activity optimization? www.openphacts.org www.openphacts.org How often do Topliss patterns occur? 1. Return 3,4-dichloro substituted compounds in postgresql ChEMBL 20 using RDKit cartridge 9312 cpds 540 x 2a. For each 3,4-Cl2 substituent check for availablity of 4-Cl, 4-OCH3, 4-CH3 and H substitutions 200 series 3. Check for each compound series for bioactivity data (pChEMBL) measured in - same target in same assay SQL query - activity type = IC50 or Ki - plus, if available, activity for new subst. selection 3nM 5nM 8nM 9nM 10nM 1 2 3 4 5 new substitution selection 1108 bioactivity data for additional substituents Raw data output after mining ChEMBL 200 series new substitution selection 1108 bioactivity data for additional substituents 3nM 5nM 8nM 9nM 10nM 1 2 3 4 5 How often do Topliss patterns occur? subst. π σ -σ π+σ Es 3,4-Cl2 1 1 5 1 2-5 4-Cl 2 2 4 2 2-5 4-CH3 3 4 2 3 2-5 4-OCH3 4-5 5 1 5 2-5 H 4-5 3 3 4 1 13 7 3 2 34 # of series 57 of 200 series (29%) extracted from ChEMBL 20 follow a Topliss pattern 200 series 3nM 5nM 8nM 9nM 10nM 1 2 3 4 5 distribution of 200 series π σ -σ π+σ Es others How useful do Topliss prove in activity optimization? Topliss pattern # of series substituent selection [1] more active [2] percent age π 13 29 9 31 % σ 7 9 1 11 % -σ 3 5 1 20 % π+σ 2 2 1 50 % scheme π 3-CF3, 4-Cl; 3-CF3, 4-NO2; 4CF3; 2,4-C12; 4-c-C5H9; 4-cC6H11; 4-CH(CH3)2; 4C(CH3)3; 3,4-(CH3)2; 4O(CH3),CH3; 4-OCH2Ph; 4N(C2H5) σ 3-CF3, 4-Cl; 3-CF3, 4-NO2; 4CF3; 2,4-C12; 4-c-C5H9; 4-cC6H11 -σ [1] For each series, bioactivity for substituents, proposed by Topliss new substituent selection were collected from ChEMBL 20, if available. [2] Check whether proposed substituents lead to more active cpds Topliss approach seems to have difficulties for series following the σ scheme in activity optimization for the series found in ChEMBL. new substituent selection π+σ 4-N(C2H5)2; 4-N(CH3)2; 4-NH2; 4-NHC4H9; 4-OH; 4OCH(CH3)2; 3-CH3,4-OCH3 3-CF3, 4-Cl; 3-CF3, 4-NO2; 4CF3; 2,4-C12; 4-c-C5H9; 4-cC6H11 poor performance of -σ is in agreement with propafenone data How useful do Topliss prove in activity optimization? propafenone dataset -σ substituion EC50 rank 3,4-Cl2 0.150 5 4-Cl 0.132 4 4-CH3 0.063 2 4-OCH3 0.045 1 H 0.079 3 target Type Topliss proposal for propafenone dataset, 4-N(CH3)2, did not show activity gain. Are there -σ series in ChEMBL with bioactivity data for 4-N(CH3)2 substitution? 4-OCH3 (nM) 4-N(CH3)2 (nM) EC50 45 82 Alpha-1a adrenergic receptor (ChEMBL) Ki 0.3 0.8 µ-opioid receptor (ChEMBL) Ki 0.50 63 P-Glycoprotein Also in the two cases of ChEMBL the -σ proposal 4-N(CH3)2 failed to increase activity. Topliss batchwise scheme propafenone aryloxy non Topliss substituion EC50 rank 3,4-Cl2 0.522 5 4-Cl 0.190 4 4-CH3 0.063 1 4-OCH3 0.180 3 H 0.079 2 Ranking pattern 5 4 1 3 2 in this dataset can‘t be assigned to an existing Topliss scheme How often does the pattern 5 4 1 3 2 occur in ChEMBL? In general, which other, non Topliss pattern occur frequently in ChEMBL? Which non Topliss pattern occur in ChEMBL? subst. new1 3,4-Cl2 1 new2 new3 5 5 aryloxy 5 4-Cl 2 2 3 4 4-CH3 4 4 1 1 The pattern found in aryloxy dataset, does not occur in ChEMBL However: High similarity to new3 distribution of 200 series 4-OCH3 3 1 4 3 H 5 3 2 2 # series 6 4 4 0 Do we find an underlying physicochemical driving force in the new3 pattern? Can we extrapolate to aryloxy dataset? π σ -σ π+σ Es new1 new2 new3 Correlation analysis within new3 series target name pattern # of cpds in series [1] r (π) r (σ ) Prostanoid EP 1 rec 53142 5+8 -0.81** Adenosine A3 rec 53142 5+8 -0.54* Adenosine A3 rec 53142 5+8 -0.67** Chymase 53142 5 + 13 -0.49** P-Glycoprotein 54132 5 [1] Next to the 5 datapoints from 3,4-Cl2, 4-Cl, 4-OCH3, 4-CH3 and H, bioactivity data from other substituents listed in Topliss et al 1977 were selected for correlation analysis. r (vdw_area) ** p < 0.05 , * p < 0.10 Correlation analyses were undertaken to calculate the Pearson correlation coefficient (r) between physicochemical features π, σ , vdw_area and the respective bioactivity data. Statistically significant negative vdw_area correlations indicate that new3 pattern & aryloxy bind to a tight pocket Discover the ranking globe How to look at the ranking space globally? There are 120 (5!) ranking possibilites (patterns) (1,2,3,4,5), (2,1,3,4,5), (1,3,2,4,5), … (5,4,3,2,1) Calculation of Spearman’s rank correlation distance matrix for 120 possibilities (R function corDist) Spherical MDS to represent the distance matrix on the surface of a sphere (R function smacofSphere), Kruksal-Stress = 0.15 Each point represents a pattern (e.g. 1,2,3,4,5) similar patterns are in vincinity to each other Frequency contour map Color coding based on frequency of patterns. Red = high frequency Blue = low frequency Map analysis Frequency contour map Color coding based on frequency of patterns. -σ • Only three –σ pattern in ChEMBL • In the investigated cases, poor predictability of –σ scheme aryloxy steric island π and σ continent trench Red = high frequency Blue = low frequency *Es steric island π and σ continent • surrounded by Es pattern • lies in area with negative vdw_area correlation Only Topliss patterns (π, σ, π+σ, Es ) and rankings patterns with four or more series (new1, new2, new3) are schown. Van der Waals contour map Color coding based on vdw_area correlations with bioactivity. Only series with activity data for five additional derivatives (e.g. 4-CF3, 4-OH ...) are used in correlation analysis (n>=10). Resulting correlations with p > 0.1 were omitted. The remaining coefficients were used for color coding. Red ... positive correlation Blue ... negative correlation Summary & Outlook • Open medicinal chemistry data such as those in ChEMBL allow analysis of complex SAR patterns • Connecting these data with data from pathways and diseases like implemented in the Open PHACTS Discovery Platform will open up completely new possibilities for linking chemical SAR patterns to biological endpoints • Quality of data is key for the analysis (assays) Next steps • Look for X-ray structures of complexes • Analyse with respect to target classes Pharmacoinformatics Research Group Department of Pharmaceutical Chemistry Pharmacoinformatics Research Group Department of Pharmaceutical Chemistry SQL query: get all 3,4-Cl2 compounds RDKit Chemoinformatics toolkit 2014.03 SMILES RDKit cartridge Data processing in python ChEMBL 20 postgreSQL > 13 000 000 activities 200 series Pharmacoinformatics Research Group Department of Pharmaceutical Chemistry -> 120 ranking possibilies are created -> Spearman ranking distance matrix calculated -> Spherical MDS is undertaken -> X,Y,Z coordinates are exported as CSV file Coordinates.csv Python data preprossesing Spherical MDS in R software 2D - EquidistantCylindrical Projections 3D - Orthographic Basemap toolkit • provides list of globe projections • create contour maps Pharmacoinformatics Research Group Department of Pharmaceutical Chemistry For each series bioactivity data for 3,4-Cl2, 4-Cl, 4-CH3, 4-OCH3 and 4-H is available • For the majority of the series (91%) there are bioactivity data for more substituents e.g. 4-CF3, 4-OH, 4-F, ... available. (Substituents taken from „new substituent selection“) • More than 57% of the series have activity data for five or more additional substituents. For series with 5 or more additional substituents (n>=10) correlation analysis were run: Series_8 3,4-Cl2 4-Cl 4-CH3 4-OCH3 4-H 4-CF3 4-F 4-OH 3,4-(CH3)2 4-C(CH3)3 pIC50 6.3 7.0 7.4 7.6 8 6.9 7.7 6.6 7 6.1 vdw_area 134 117 116 131 99 129 103 109 134 152 In this example: R = -0.70, p = 0.03 Series 8 with pattern 5 4 3 2 1, has R(vdw) = -0.7 Pharmacoinformatics Research Group Department of Pharmaceutical Chemistry Details to Multidimensional Scaling with First 2D MDS bad Kruksal-Stress-1 > 0.2 Second 3D MDS good Kruksal-Stress-1 = 0.11 but visualization not helpful Third Spherical MDS moderate Kruksal-Stress-1 = 0.15, good visualization ✔ get120Possibilities() ... creates a vector with 120 rankings [(1,2,3,4,5), (2,1,3,4,5) ...] corDist () ... calculates Spearman‘s rank correlation distance smacofSphere() ... runs spherical MDS, type=„ordinal“ because we have rankings, algorithm=„primal“ ... handling of ties xyz.120 ... x,y,z – coordinates of the MDS run Coordinates (xyz.120) are exported to CSV file and are the input for Basemap Pharmacoinformatics Research Group Department of Pharmaceutical Chemistry • • • • • • • Potentielle Fragen: -> Wie lange dauert so eine Suche wenn der Workflow steht ~ 1 Tag (4 Prozessoren Rechner, 8GB RAM) -> Wie werden Salze behandelt? Skript ist so geschrieben dass diese nicht berücksichtigt werden. Soll heißen es wäre potentiell möglich dass die diChloro verbindung ein Natriumsalz ist und das Methylderivat ein Kaliumsalz. Wie auch immer in den 200 Serien war dies nie zu finden und spielt somit keine Rolle. -> Wie steht es um Chiralität. Ich habe die Chiralität nicht berücksichtigt in der Query. Dies wäre möglich gewesen aber da die Codierung von Chiralitäten in ChEMBL nicht umfassend ist habe ich es nicht berücksichtigt. -> wie groß muss den Unterschied sein zwischen den Bioaktivitäten damit es als Serie anerkannt wurde? Im Topliss paper findet man rankings mit log >0.1 zwischen den Verbindungen. Wir haben darauf keine Rücksicht genommen und alle Daten verwendet (so wie es übrigens auch die Gruppe die 2014 eine ähnliche Analyse auch gemacht haben) Die Datenanalyse zeigt von den 200 serien: Haben 43 eine Differenz von mindestens „>0.1 log“ zwischen den rankings. 77 series haben 1 verstoß dieser regel, d.h. die differnz zwischen 2 rankings ist ein mal kleiner 0.1. 80 haben dann 2 oder mehr verstöße. Warum habt ihr die anderen pattern 2pi-pi^2, pi-sigma usw. nicht berücksichtigt? Die Komplexität wäre deutlich höher gewesen ohne dass es einen nennenswerten Informationsgewinn gegeben hätte. Zur Abgrenzung, die neuen pattern „new 1, new 2, new 3) fallen in keines der von Topliss postulierten pattern auch nicht in die erweiterte Auswahl (2pi- pi^2, pi-3sigma, usw.)