Prediction Quality Attributes For Mixtures From Single Component Data John Strong, PhD Sean Garner, PhD Global Pharmaceutical Sciences AbbVie, North Chicago, IL CPDG Sep 27th, 2016
AbbVie: Our reach is global 170+ Countries 12 Manufacturing Sites 6 R&D Sites Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 2 2
Our name was launched in 2013 but our story begins in 1888 125 years of patient care Addressing complex and serious diseases Making a remarkable impact on patients lives Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 3 3
We offer medicines to treat the world s challenging diseases Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 4 4
John s Bio BS Mech Eng., Portland State 1991 M.S. Chem Eng, Yale 1993 Ph.D. Biochem Eng, UMBC 1997 Joined Abbott 1997, AbbVie 2013 Married w/ 1 son in Lindenhurst, IL Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 5
Drug Product Development (DPD) the material From API molecule to solid dosage form (tablet, capsule) Multi scale Multi discipline Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 6
DPD development cycle Reduction in effort and cost if more predictive approaches can be used early on with empirical confirmation later. Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 7
Two philosophies for DPD formulation/process design Just Measure It! Quick material sparing experimental approaches exist for predicting manufacturability (e.g., flow, tableting) Employing prior knowledge, test candidate formulations for manufacturability using trusted methods. Experimentally observe which ones work, discard the rest. Model It First! Utilize database of relevant material properties and a suite of predictive models using individual component data Put in silico formulation in the hands of the formulators. Generate development space guidance for further formulation development. Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 8
The AbbVie DPD approach Embracing model based prediction wherever applicable and valueadded, as early as possible The intent is to meet the vision of QbD by interrogating a comprehensive database of relevant material properties to yield inputs to relevant models for multicomponent blends: Major effort (over 2 years) to develop the database & user interface Mean values for excipient material properties (from multi lot repeat testing) are provided to predictive models Estimates of lot to lot excipient variability can be calculated and propagated through models to obtain confidence intervals on predictions MATLAB implementation makes most of these models available to formulators through user interface Examples: Powder blend flow Blend tabletability Blend uniformity FEM tablet predictions Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 9
Powder blend flow
Motivation Good powder flowability is essential to pharmaceutical processes. Powder flowability affects: Mixing/blending Segregation Feeding/transfer/hopper flow Coating Fluidization Capsule filling Die filling/tableting Unacceptable content uniformity Variable tablet/capsule weights Variable potency Complicates/prevents manufacturing Reduces manufacturing efficiency Complicates product/process development Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 11
Motivation Powder flow (along with most powder behavior) is extremely difficult to model and predict. Modeling has been purely empirical which limits understanding and predictive capability. Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 12
Quantifying flow Flow Function Coefficient: ff : consolidation stress : unconfined yield stress ff c is an excellent metric to quantify a powder s flowability in unconfined flow situations: Jenike Classification System Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 13
Theory what influences flow at particle level? Adhesive force between two particles: 16 1 2 3 Particle Property Particle size, d p Hamaker constant, A Asperity size/surface roughness, d asp Effect on F ad with increase in property Chen et al., AIChE Journal 54 (2008) 104 121. Possible effect on powder flow Separation distance, l 0 Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 14
Theory what influences flow at particle level? Adhesion force alone does determine powder cohesiveness! Inter particle cohesion parameter (Granular Bond Number): Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 15
Theory what influences flow at particle level? Inter particle cohesion parameter (Granular Bond Number): Particle Property Effect on F ad with increase in property Effect on Bo g with increase in property Possible effect on powder flow Particle size, d p Particle Density, ρ p N/A Hamaker constant, A Asperity size/surface roughness, d asp Separation distance, l 0 Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 16
Establishing Bo g ff c Relationship Material Granular Bond Number, Bo g APAP 7.99 10 2 1.40 Avicel PH 105 4.60 10 2 2.00 Avicel PH 101 1.18 10 1 4.53 Avicel PH 102 3.74 6.28 Avicel PH 200 3.86 10 1 13.0 Flow Function Coefficient, ff c APAP (1 wt% R972) Avicel PH 105 (1 wt% R972) Avicel PH 101 (1 wt% R972) Avicel PH 102 (1 wt% R972) Avicel PH 200 (1 wt% R972) 1.08 10 1 5.27 4.40 8.30 1.27 10 1 14.1 3.03 10 2 12.3 3.71 10 3 15.1 Higher interparticle cohesiveness (high B og ) correlates to poor flow (low ff c ) Capece, Ho, Strong, Gao (2015) Prediction of powder flow performance using a multi component granular Bond Number, Powder Tech. 286:561 671 Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 17
Establishing Bo g ff c Relationship (Cont.) ff c B og relationship: ff 1 α=15.7 β=0.27 (Maximum ff c constrained to 13.7) Flow function coefficient can be empirically related to the bond number. Cohesive non cohesive boundary (Bo g = 1) separates free flowing powders from all others. Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 18
Bond Number for Powder Blends ff c B og relationship: Single component 1 ff ff c B og relationship: Powder blend ff, 1, Bond number of a mixture.. Weighted harmonic mean: 1,.,, Interaction between all component of a mixture are considered. Material properties of all component in the blend are related to the bulk powder behavior. Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 19
Predicting ff c of Binary Mixtures Binary blends are well predicted. Average absolute deviation is 0.38. Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 20
Predicting ff c of Binary Mixtures Binary blends for coated (surface modified) materials are also well predicted. Average absolute deviation is 0.54. Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 21
Predicting ff c of Ternary Mixtures Average absolute deviation is 0.25 for as received materials. Average absolute deviation is 1.67 for coated (surface modified) materials. Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 22
Population dependent granular Bond Number,, Accounts for the multitude of interactions between particles of different sizes within a powder Provides prediction for how ff c changes with modifications to particle size distributions of mixture components Flow Function Coefficient, ff c (-) 20 15 10 5 Non-cohesive Cohesive As-Received Surface Modified Sieved Eq. (15): = 14.8, = 0.28 95% CI 0 10-3 10-2 10-1 10 0 10 1 10 2 10 3 10 4 10 5 Population-Dependent Granular Bond Number, Bo* g (-) Capece, Silva, Sunkara, Strong, Gao, On the Relationship of Interparticle Cohesiveness and Bulk Powder Behavior: Flowability of Pharmaceutical Powders, J. Pharm. (in press) Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 23
Population dependent granular Bond Number Flow Function Coefficient, ff c (-) 20 15 10 5 Experimental Prediction Experimental and predicted flow function coefficients for powder blends containing various loadings of APAP. All blends contain 1 wt% silica (Aerosil R972). 0 20 wt% 40 wt% 60 wt% 80 wt% 99 wt% APAP Loading Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 24
Tablet tensile strength
Introduction A Partial History Fell & Newton, 1970 Predicted tensile strength for various forms of lactose of same PSD Sheikh Salem & Fell, 1981 also showed linear relationship for some materials Leuenberger, 1982 showed deviation from linearity, proposed interaction term to account for deviations Leuenberger, 1985 predicted interaction term using solubility parameters for limited materials Jetzer, 1986 interactions occur with dissimilar materials Ilkka 1993 Proposed the following trends Plastic + plastic linear relationship between components Plastic + brittle dictated by plastic (lower yield pressure), slightly non linear Brittle + brittle dictated by higher yield pressure component (skeleton effect), non linear Kuentz & Leuenberger, 2000 Used percolation theory to explain deviations van Veen, 2000 Showed that two plastic materials can be very non linear if their densification and relaxation mechanisms are different Van Veen, 2004 Used intermediate data point to calculated tensile strength for non linear materials Michrafy, 2007 Tried additive and geometric mixing rules with Ryshkewitch Duckworth equation, no clear conclusion on which works better Etzler, 2011 Justified a binary geometric mixing rule weighted by volume fraction via analogy to van der Waals binary interactions Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 26
Tensile Strength Mixture Model Problem 1: Most drug compounds aren t very compressible, even in a simulator under slow compression. How is the pure drug tensile strength estimated? 3 Drug m n i1 i v i, i f ) ( i 1 2 Excipient 1 Excipient 2 Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 27
Tensile Strength Mixture Model Problem 1: Most drug compounds aren t very compressible, even in a simulator under slow compression. Solution: Measure tensile strength of a 50:50 drug-excipient ratio for each excipient. Then determine a pure drug tensile strength which fits the 50:50 data point. Take mean of estimates. 1 1,3 3 Drug i1 Excipient 1 Excipient 2 m 3 2,3 n i1 1 2 2 2 v i, i i i,3 v,1 1 v, D Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 28
Tensile Strength Mixture Model Problem 2: Different excipients may exhibit very different yield strength, i.e., different porosities at a given pressure. Would a single tablet relative density then be sufficiently indicative of the relative extent of deformation of each of the components? Solution:? We haven t investigated this yet! m i n i1 f ) ( i v i, i Relative Density 0.900 0.850 0.800 0.750 0.700 0.650 0.600 0.550 0.500 STARCH 1500 & DCP COMPRESSIBILITY Starch DCP 0 50 100 150 200 250 300 350 Compaction Pressure (MPa) Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 29
Ternary geometric mixing results Avicel DCP APAP Excipients varying 0 100%, APAP varying 0 50% Avicel Lactose APAP Mannitol Starch APAP Despite being somewhat simplistic, the geometric mixing model seems to work well enough to be useful! Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 30
Blend Uniformity
Example API particle size distribution Desired State: Design Space Paradigm The design space for a given set of process parameters/controls (e.g., API particle size distribution) is the intersection of all parameter spaces for the quality attributes they impact. QA: Dissolution QA: Content Uniformity Mechanism: random mixing QA: Content Uniformity Mechanism: Segregation Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 32
API particle size distribution In addition to flow and compaction, it directly impacts blend uniformity Quantitative understanding of this relationship can aid in defining particle size distribution ranges Early identification of workable range in particle size can reduce development effort and promote more collaboration between API development and drug product development Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 33
Example API particle size distribution USP 29 Content Uniformity Calculator USP 29 Content Uniformity Calculator Inputs Target potency 100 %LC Sample size 30 no. tablets Sample mean potency 98 %LC Sample RSD potency 3.6 % Confidence level 95 % Calculations Sample mean upper bound 99.93 %LC Sample mean lower bound 96.07 %LC Sample SD upper bound 4.74 %LC Upper Bound: Stage 1 Stage 2 I1: 63.19% 91.62% I2: 12.93% 3.52% I3: 14.73% 4.86% Probability of Passing 90.85% 99.99% Maximum probability 99.99% Lower Bound: Stage 1 Stage 2 I1: 4.81% 0.24% I2: 0.01% 0.00% I3: 64.67% 97.74% Probability of Passing 69.50% 97.90% Maximum probability 97.90% Statistical model based on USP <905> tells us what CU we need to achieve Excel spreadsheet statistical model Calculates the probability of passing at a given confidence level Based on Berghum s statistical analysis ` Lower bound probability 97.9034% Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 34
Example API particle size distribution USP 29 Content Uniformity Calculator Probability of passing (or % success rate) for a given sample mean and sample RSD Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 35
Example API particle size distribution USP 29 Content Uniformity Calculator Sample std dev as a function of sample mean for a given passing probability or success rate Six-sigma criteria requires NMT 3% RSD Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 36
Example API particle size distribution Blend uniformity vs Content uniformity A content uniformity of 3% RSD seems to offer sufficient robustness You might expect a ~1% difference between blend and content uniformity for a well controlled drug product process (process scientist should confirm experimentally) So what kind of API particle size distribution gives us a 2% blend RSD? Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 37
Example API particle size distribution Blend uniformity RSD vs. drug loading, diameter RSD 100% d 6M 3 1 p p Not realistic Useful general guideline BU not really a concern past 5-10% DL Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 38
Example API particle size distribution Calculating blend uniformity from API size For a measured distribution (Stange Poole equation): RSD 100% fimi xm i x For a lognormal distribution (Rohrs et al, 2006): RSD 100% 6xM 1 2 1 9 3 2 2 2 exp 4.5ln 10 g 1 2 g x M f i m i g g mass fraction of API in blend mass of drug product mass fraction of diameter i mass of a single particle in mass fraction f i geometric mean diameter geometric standard deviation Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 39
Example API particle size distribution Actual distributions vs. lognormal approximation Example from ABT-894 project Matched d(0.5) and d(0.9) Good fit to measured data with some discrepancy in the tailing g g 1 ( 1. 28155 d d d(0.5) 0.9) (0.5) Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 40
Example API particle size distribution BU vs d(0.5), d(0.9) Reference Charts Charts are a design range with dimensions of d(0.5) vs. d(0.9) Don t depend on excipient particle size at low drug loading Offers more insight regarding API PSD, useful at early stage formulation Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 41
Example API particle size distribution Can also function as API Design Space Predicted Blend Uniformity vs d(0.5), d(0.9) Predicted BU correlated to actual measured CU API Predicted BU (RSD %) Fine 0.58 2.1 Medium 1.18 2.1 Coarse 1.85 2.8 Measured CU (RSD %) Green known good Red - unknown? Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 42
Example API particle size distribution What about controlling de mixing? Design space is common overlap of individual design ranges! De mixing or segregation occurs with significant size/density differences between API and excipients Unfortunately no mechanistic models to predict it yet To control segregation, design space can be modified by putting lower limit on d(0.5). Empirical at this point. Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 43
FEA
Finite Element Analysis: Drucker Prager Cap (DPC) Model Shear failure surface Cap surface (densification) The ability to perform DPC analysis on multi component blend estimates would add to the early modeling toolbox and provide additional early guidance on choice of tablet tooling design and choice of excipients/grades for investigation. Capable of predicting stress and density distribution in compacted tablet Applications in formulation development: Tooling design Identifying high stress risk zones for prediction of tablet defects Optimizing tablets for improved friability Garner, Ruiz, Strong, Zavaliangos (2014) Mechanism of crack formation in die compacted powders during unloading and ejection: An experimental and modeling comparison between standard straight and tapered dies, Powder Tech. 264:114 127 Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 45
Finite Element Analysis: Drucker Prager Cap (DPC) Model Die Compaction zz rr Describes the limit to elastic deformation One curve per relative density (denser=stronger) Described on a hydrostatic (p) and Mises (q) stress plane Calibrated easily by press simulator experiments Diametral and Simple Compression Strength Tests Two tests per density level Fully Instrumented Die Compaction Experiments Only one experiment required for all densities Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 46
DPC required inputs Defining the yield surface Cohesion d Diametral & Simple Friction angle β Compression Cap eccentricity R Die Compaction Hydrostatic yield stress p b Other inputs Elastic modulus Poisson ratio Can these inputs for a blend be estimated (well enough to be useful!) from individual Die Compaction components? Is this approach viable? Cunningham, Sinka, Zavaliangos (2004) Analysis of Tablet Compaction. I. Characterization of Mechanical Behavior of Powder and Powder/Tooling Friction, Powder Tech. 283:210 226 Garner, Strong, Zavaliangos (2015) The extrapolation of the Drucker Prager/Cap material parameters to low and high relative densities, Powder Tech. 283:210 226 Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 47
Tertiary Mixture: 33% MCC + 33% Lactose + 33% Mannitol Failure Surface Parameters 25 70 Cohesion [MPa] 20 15 10 5 Tertiary Mixture Mixture Model Friction Angle [degrees] 68 66 64 62 Tertiary Mixture Mixture Model 0 0.4 0.6 0.8 1 Relative Density 60 0.4 0.6 0.8 1 Relative Density Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 48
Tertiary Mixture: 33% MCC + 33% Lactose + 33% Mannitol Cap Surface Parameters Cap Eccentricity 2 1.5 1 0.5 0 Tertiary Mixture Mixture Model 0.4 0.6 0.8 1 Relative Density Hydrostatic Yield Stress [MPa] 400 350 300 250 200 150 100 50 0 Tertiary Mixture Mixture Model 0.4 0.6 0.8 1 Relative Density Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 49
Elastic parameters Extraction of Elastic Parameters Axial Stress [MPa] 300 250 200 150 100 50 Increasing RD 0 0.75 0.8 0.85 0.9 0.95 Volumetric Strain Linear estimates of unloading slopes for axial and radial stress versus volumetric strain are taken from the beginning stage of unloading Poisson s Ratio 1 1 Young s Modulus 1 1 σ zz = Axial stress σ rr = Radial stress ε = Strain 12 Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 50
Elastic Properties from Mixtures Young's Poisson's Modulus Ratio [GPa] 7 0.4 6 0.35 5 1:1 Mixture Model 0.3 4 0.25 3 1:1 Mixture 1:1 Mixture 1:1 MCC:Lactose 0.2 2 0.15 1 0.1 1:1 MCC:Lactose 0 0.05 1:1 Mixture Model 0.4 0.6 0.8 1 Relative Density 0 0.4 0.6 0.8 1 Relative Density Young's Modulus [GPa] Poisson's Ratio 8 Tertiary Mixture Tertiary Mixture 7 Tertiary Mixture 0.4 6 Mixture Model 0.35 5 0.3 4 0.25 3 0.2 2 0.15 1 0.1 Tertiary Mixture 0 0.05 Mixture Model 0.4 0.6 0.8 1 0 Relative Density 0.4 0.6 0.8 1 Relative Density Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 51
Finite element simulation setup Tooling was modeled as cylindrical flat face Axisymmetric geometry allowed for 2D simulation The mesh that represented powder material consisted of 3,000 4 node bilinear axisymmetric elements Material modeled as MCC and lactose mixtures Upper Punch Die Wa Equivalent Stress [MPa] 160 140 120 100 80 60 40 20 Axis of Symmetry Powder Material 0-10 10 30 50 70 90 110 130 150 170 190 Hydrostatic Stress [MPa] Lower Punch Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 52
Comparison of measured vs. modeled Relative Density Contours DPC Parameter Extraction from Measurement Comparison of Tertiary Mixture DPC Parameter Extraction from Mixture Model End of Unloading End of Unloading Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 53
Building on the blend model approach
Constructing an initial development space Acceptable tensile strength composition space Acceptable flowability composition space Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 55
Constructing an initial development space Acceptable tensile strength composition space Acceptable flowability composition space Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 56
Constructing an initial development space Can this multicomponent blend modeling be applied to more advanced modeling techniques? Acceptable flowability and tensile strength composition space Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 57
Final Notes & Conclusions Multi component blend modeling has the potential to help unload the early stages of development in regards to experimental effort, but material sparing measurements are a parallel approach. The first model it! approach is not easily or quickly implemented, requiring a comprehensive database of material properties to fully fuel the effort. Nor is it the complete solution. But once implemented, the approach would exemplify the QbD vision. Prediction of Blend Quality Parameters CPDG 09/22/2016 Copyright 2016 AbbVie 58