How readily drugs crystallize after fast solvent evaporation, and why the drug matters more than the route

Background and motivation

Many new drug candidates are poorly soluble, and making the drug amorphous, often as a solid dispersion in a polymer, is one of the standard ways to raise its apparent solubility (Introduction). C1 The gain can be large: calculations from thermodynamic data suggest up to 1600-fold, although the figure is highly compound specific (Hancock and Parks 2000, ref 8). C2 The weak point is crystallization. It can happen while the solid is stored or while it dissolves, and it can remove most of the advantage (refs 9–14, among them Alonzo et al. 2010, ref 14). C2

Crystallization from the amorphous state is the sum of nucleation and crystal growth, and many factors act on one or both. The review by Bhugra and Pikal (ref 15), which the authors lean on, sorts them into thermodynamic factors (the free energy difference between the crystal and the amorphous form, $\Delta G_v$), kinetic factors (molecular mobility, for which the glass transition temperature $T_g$ is the usual indicator) and molecular factors (hydrogen bonding), and adds moisture and the way the amorphous form was made. C3 Amorphous solids can be made in many ways, but in industry two routes dominate: cooling a melt, as in melt extrusion, and removing a solvent, as in spray drying or freeze-drying (Willart and Descamps 2008, ref 17). C4

Most stability studies look at one or a few drugs. The authors argue that a larger set of compounds handled under standardized conditions "would potentially yield" a more general understanding. C5 They point to Greaser et al. (2009, ref 21), who related the amorphous stability of 12 APIs, made by cooling the melt in DSC pans, to kinetic and thermodynamic parameters; that study found a reasonable correlation between stability above $T_g$ and configurational entropy, concluded that no simple approach would suffice, and suggested multivariate analysis. C5 The same group as the present paper had then classified the glass-forming ability and glass stability of 51 structurally diverse compounds after cooling from the melt in DSC pans, again with multivariate analysis (Baird et al. 2010, ref 1). C6

Solvent evaporation is at least as common a route to amorphous material, but fewer studies have measured how readily drugs crystallize after it. C7 The open question the paper takes up is therefore simple: if the same compounds are made amorphous from solution instead of from the melt, do they rank the same way? Put differently, how much of a drug's crystallization tendency belongs to the drug, and how much to the preparation method? To answer it, the authors took the 51 compounds of the melt study, dissolved each in three solvents of different polarity, dried the solutions within seconds by spin coating, followed the films for a week, sorted the results into three classes, compared them with the melt classes and asked, with principal component analysis (PCA), which compound properties go with each class.

Background concepts

Spin coating as a model of fast drying. In spin coating a small volume of solution is placed on a flat substrate, which is then spun fast. Following Bornside et al. (ref 23), the authors describe four stages: deposition, spin-up (wetting the surface), spin-off (excess liquid thrown off, leaving a film of nearly uniform thickness) and evaporation. C10 For a dissolved drug the evaporation stage matters most: as solvent leaves, the solution becomes supersaturated, and nuclei can form and grow; where supersaturation is reached and how fast the viscosity rises depend on the solvent. C10 Because the whole process takes seconds and the film is thin with a large surface open to air, the authors expect that it "should minimize" nucleation and growth while the drug is still in solution. C11 They compare this with spray drying ("some similarity") and contrast it with slow processes such as evaporation at room temperature or rotary evaporation. C12 Here every film was made from 200 µL of solution on an 18 × 18 mm cover slip spun for 20 s at 8000 rpm, which gave films about 1 µm thick. P2

Scoring crystallinity by polarized light. Crystals are birefringent and light up between crossed polarizers; an amorphous film stays dark. Because the films are so thin, the authors assume that a crystal, once formed, spans the whole film height, so the share of the film area that is birefringent is a semi-quantitative measure of crystallinity ("it is reasonable to assume"). C8 Each film was scored on five steps: AAAA (completely amorphous), AAAC (slightly crystalline, amorphous regions dominate), AACC (crystalline and amorphous areas comparable), ACCC (crystalline material dominates) and CCCC (completely crystalline). P8 The Methods call AACC and ACCC "semi-crystalline" and "predominantly crystalline"; the footnote of Table 3 calls them "intermediately crystalline" and "highly crystalline"; the codes mean the same. X11

Three crystallization classes. From the score at each time point a drug–solvent combination was put in one of three classes: Class I (rapid) if it was at least AACC at day 0, within 30 min of spin coating; Class II (intermediate) if it was at least AACC by day 7; Class III (slow) if it was still AAAA or AAAC at day 7. P9 The melt study used the same three labels with different criteria: Class I if the sample crystallized while cooling at 20 °C/min in the DSC, Class II if it crystallized only on reheating at 10 °C/min, and Class III if it did not crystallize at all. P10 The two tests therefore differ in medium (solution versus melt), in time scale (a week versus minutes) and in temperature history (room temperature versus a sweep above and below $T_g$), a point the authors make themselves. C30

Descriptors and the equations behind them. Seven properties of each compound entered the analysis: the melting point $T_m$ (onset, K), the glass transition temperature $T_g$ (midpoint, K), the heat of fusion $\Delta H_{fus}$, the entropy of fusion $\Delta S_{fus}$, the free energy difference $\Delta G_v$, the molecular weight (MW) and the number of rotatable bonds (#RB). P11 $T_m$, $T_g$ and $\Delta H_{fus}$ were measured by DSC in the melt study. P11 The entropy of fusion follows from them (Eq. 1):

In plain terms: the entropy gained on melting equals the heat absorbed on melting divided by the temperature at which it happens. E1 The free energy difference between crystal and amorphous form was estimated with the Hoffman equation (Eq. 2; printed as "Hoffmann") at $T$ = 293.15 K:

In plain terms: the further the temperature $T$ lies below the melting point and the larger the heat of fusion, the more negative $\Delta G_v$, that is, the larger the thermodynamic drive for the amorphous form to crystallize. E2 All three energies were put on a per-volume basis by multiplying by the crystal density, taken from crystal structures or from helium pycnometry. E5 For 16 molecules that crystallized too fast for a $T_g$ to be measured, and for which no literature value existed, $T_g$ was estimated from the $T_g/T_m$ ratios of the 35 compounds whose $T_g$ was known. E7·P11

Classical nucleation theory (Turnbull and Fisher 1949, ref 32; James 1985, ref 33) links $\Delta G_v$ to how hard it is to start a crystal. The free energy needed to form a nucleus of critical size is (Eq. 3):

In plain terms: $W^{*}$ is the barrier a cluster must climb before it grows on its own, and $\gamma$ is the free energy per unit area of the interface between the nucleus and the surrounding liquid (a highly supersaturated solution before the solvent is gone, or the supercooled liquid or glass after). A larger driving force lowers the barrier with its square; a larger interfacial energy raises it with its cube. E3

PCA, cross-validation and the Coomans' plot. PCA replaces many correlated variables by a few new axes, the principal components, that capture most of their variation; each compound gets a score on each axis. The data were centred and scaled to unit variance, and the number of meaningful components was judged by cross-validation, which predicts left-out parts of the data from the rest; the share predicted is a guard against fitting noise. P11 Crucially, the class of a compound was not used to build the model, so any separation of the classes on the score plot comes from the descriptors alone. P11 A class contribution is the difference between the class average of a descriptor and the overall average, in units of the overall standard deviation. R20 For prediction the authors built one PCA model per class and plotted each compound by its distance to two of the models (DModXPS+) in a Coomans' plot; a line at the critical distance (significance level 0.05) marks where a compound no longer fits a model. P12

Methods in detail

The study has three parts: a miniaturized fast-drying screen with a fixed protocol, a classification of every drug–solvent combination with a comparison against the melt classes, and a descriptor analysis. The flow below shows how they connect.

Contribution 1: a fixed, miniaturized fast-drying screen

The problem. To compare many drugs fairly, every film must be thick enough for polarized light to show crystals, but the solution must stay dilute enough that poorly soluble drugs can still be dissolved. The concentration also changes how the film forms.

What they did. Felodipine, which the group knew forms an amorphous film by spin coating, was used to set the concentration. C13 Solutions of 0.03, 0.07, 0.13, 0.20, 0.26, 0.52, 0.78, 1.04 and 1.30 M were spin coated from DCM, MIX (1:1 w/w dichloromethane/ethanol) and EtOH; in ethanol only up to 0.26 M could be dissolved. P3 The average film thickness was estimated from the weight gained by the cover slip, using 1.28 g/mL for the density of amorphous felodipine (from Marsac et al. 2006, ref 22); n = 3. E6·M26·P3 The films were then stored for 2 days at 94% relative humidity (over saturated potassium nitrate) at room temperature, deliberately to crystallize them, so that the crystal coverage of the substrate could be judged. P4 From these tests 0.26 M became the standard concentration; six combinations with solubilities between 0.20 and 0.26 M (anthranilic acid/DCM, celecoxib/DCM, felbinac/EtOH, miconazole/EtOH, nilutamide/MIX, probucol/EtOH) were run at 0.20 M. P1·R4

Solubility in each solvent was measured by UV spectroscopy after shaking excess drug in 1 mL of solvent for 24 h at room temperature and separating the solid by centrifugation (16,000g, 10 min) or, for DCM, whose density (1.325 g/mL) is close to or above that of some drugs, by filtration through 0.2 µm PTFE filters; each measurement was done in triplicate. P5·M56 Table 1 lists, for 54 compounds, the solvent used for the standards, the detection wavelength and the R² of the calibration (0.998 to 1.000). P5 The saturated solutions were then checked by reversed-phase HPLC (gradient from 10% to 90% methanol against 0.1% trifluoroacetic acid in water, detection at 230 nm) against fresh solutions; any detectable degradation excluded the combination, because degradation products resemble the parent molecule and structurally similar molecules can change nucleation and growth ("it is well known"). P6·C9 Compounds whose films sublimed when put under vacuum were also excluded, unless they had already crystallized completely. P8

The storage protocol. Each combination was spin coated in triplicate. Films were scored within 30 min (day 0), kept overnight under vacuum to remove residual solvent and scored again (day 1), then stored over phosphorus pentoxide, that is, dry, and scored after 2, 4 and 6 more days (days 3, 5 and 7). P2·P8 Storage was at room temperature. P8

Why it works. The thin film and the short process are meant to leave little time for crystallization in solution, so that what is seen reflects the compound's tendency in the near-solid state; and the milligram amounts make a 51-compound, three-solvent screen practical. C11·C12

Contribution 2: a class for every drug–solvent combination, compared with the melt

Each of the 93 combinations was assigned Class I, II or III by the rules above. P9 Where the class of a compound depended on the solvent, it was recorded as I/II or II/III. R8 For the comparison with the melt study, each of the 93 spin-coating results was set against the melt class of the same compound, the combinations were grouped by class pair (for example Class I by evaporation and Class II from the melt), and the size of each group was expressed as a percentage of 93:

percentage of a group = 100 × (number of combinations in the group) / 93. E4

This is a count per combination, not per compound: a drug tested in three solvents counts three times. X6 As a separate test of the role of drying rate, every Class III combination was also dried slowly: 400 µL of solution was left to evaporate at ambient conditions in a small glass vessel (12 mm wide, 6 mm high), in triplicate, and scored in the same way, with day 0 taken once the solvent had gone. P7

Contribution 3: which compound properties go with each class

For PCA, the 40 screened compounds were described by the seven descriptors defined above; compounds of mixed class were counted as Class I (if I/II) or Class III (if II/III). P11 One PCA model was built for all 40 compounds and, separately, one three-component model for each class. P11·P12 Three compounds that were not used to build the models, one per class, served as a check: PABA (spin coated from MIX and EtOH; Class I), piroxicam (DCM; Class II) and clarithromycin (DCM; Class III). P12 What is new compared with the melt study is not the statistics but the data set: the same descriptors and the same compounds, now judged by how they crystallize after fast solvent removal and a week of storage.

Experimental analysis

Choosing the concentration (Figures 1 and 2)

Film thickness rose with concentration, "as expected" (Figure 1). C14 Read from the plot, felodipine films from DCM were about 0.09 µm thick at 0.03 M, about 0.8 µm at 0.26 M and about 6.2 µm at 1.30 M; from MIX they reached about 4.2 µm at 1.30 M. R1 The authors state that above 0.26 M DCM gives "significantly" thicker films than MIX and attribute this, "most likely", to the higher volatility of DCM: the deposit becomes too viscous to spin off sooner. C14·C15 In the plot the two solvents still coincide at 0.52 M (about 1.1 µm each, with a wide error bar on DCM) and separate from 0.78 M (about 3.3 against 1.9 µm); no statistical test is given. R1·X4 At 0.26 M and below the authors describe the films as similar in thickness irrespective of solvent; at 0.26 M itself the readings are about 0.80 (DCM), 0.65 (EtOH) and 0.52 µm (MIX). C14·R1·X4 At 1.04 M and above the films were visibly uneven and held crystals in the bulk, "suggesting" nonuniform deposition and slower evaporation. R2·C16

After 2 days at 94% RH, felodipine films from 0.26 M upward were fully covered by crystals; at 0.20 and 0.13 M only patches crystallized, and at 0.07 and 0.03 M no crystals were seen (Figure 2, DCM; the other solvents behaved alike, data not shown). R3 Below 0.26 M the drug simply did not cover the substrate, which limited crystallization. C17 The authors took 0.26–0.78 M as the workable range and 0.26 M as the standard, the lower end, to lose as few poorly soluble drugs as possible. C17

Note by the AI writer (not discussed in the paper). The six combinations run at 0.20 M lie below the concentration at which felodipine films crystallized over their whole surface. Each of the six has the same class as the same drug in its other solvents, so no compound's class depends on them. X5

Which drugs could be screened

Of the 153 drug–solvent combinations (51 compounds in three solvents), 93 combinations of 40 compounds passed the solubility, stability and processability criteria. R7 Griseofulvin, nifedipine, procaine and tolfenamic acid showed degradation in solution, and the films of antipyrin and ibuprofen sublimed under vacuum. R5·R6

Note by the AI writer (not discussed in the paper). The compound counts do not quite add up in print. Table 1 lists 54 compounds, that is the 51 plus the three validation compounds (PABA, piroxicam, clarithromycin); the Materials section names 55, including niclosamide, which appears nowhere else. Five of the 51 (D-(−)-salicin, droperidol, haloperidol, indoprofen, theophylline) are missing from the results without a named reason; presumably they failed the solubility criterion. X3

Three classes, and the solvent rarely matters (Table 2)

Of the 93 combinations, 39 were Class I, 23 Class II and 31 Class III. R9 By compound, 14 were Class I in every solvent tested, 9 Class II and 13 Class III; clofoctol and flurbiprofen were Class I or II and cinnarizine and loratadine Class II or III depending on the solvent (Table 2). R8·R9 Table A gives the compounds by class.

Table A. The 40 screened compounds by crystallization class after spin coating (from Table 2). R8·R9

The classes look different under the microscope. Class I combinations were extensively crystalline right after spin coating, despite drying in seconds; the authors read this as "suggesting" extremely fast nucleation and growth, either from the supersaturated solution or in the viscous film of the last drying stage. R8·C18 Class II and III combinations gave substantially amorphous films. C19 About half of the Class II compounds showed a little crystallinity right after spin coating and the others none until later; the authors take both routes, nuclei carried over from spin coating and relatively rapid nucleation in the amorphous film, to be "probably important" ("may be due"). R10·C20 Either way, growth at room temperature was fast enough that by day 7 the crystalline area at least equalled the amorphous one. C21 Class III compounds showed no crystallinity right after spin coating and little after a week, "implying" slow nucleation and/or growth. R10·C22

The solvent mattered little. Of the 33 compounds tested in more than one solvent, 10 differed between solvents in their score at one or more time points (bifonazole, carbamazepine, celecoxib, cinnarizine, clofoctol, clotrimazole, felodipine, flurbiprofen, loratadine, nilutamide), and only 4 changed class (cinnarizine, clofoctol, flurbiprofen, loratadine). R11 The authors conclude that crystallization "depends to a large extent on the compound": although solvents are known to steer which polymorph forms (Khoshkhoo and Anwar 1993, ref 30), solvent choice "seems to be less critical" for whether amorphous material forms and for its short-term stability, "at least" for these solvents and this method. C23·C24 Their explanation is the drying time: evaporation and film formation are so fast that any solvent-specific effect on nucleation or growth is suppressed ("can be explained"). C25

Slow drying changes the picture (Table 3)

When the 31 Class III combinations were dried slowly in a vial, 22 were completely crystalline (CCCC) at the first time point. R13 Only four compounds kept some amorphous material, and for them the outcome depended on the solvent (Table B). R12·R13

Table B. The four Class III compounds that were not fully crystalline at day 0 after slow evaporation (from Table 3). All other Class III combinations (celecoxib, felodipine, indomethacin, itraconazole, ketoconazole, pimozide, acetaminophen, atenolol, aceclofenac, cinnarizine, loratadine) were CCCC from day 0 in every solvent tested. R12

The authors call the loss of amorphous material "an obvious result" of the longer time the solution spends supersaturated before its viscosity rises, and the solvent dependence of the four partly amorphous compounds a sign that the solvent has "a much greater influence" during slow evaporation ("implies"). C26·C27·C28 Together with the spin-coating data this leads them to say that, under rapid solidification, the crystallization behaviour of an API "appears to be largely independent" of the solvent. C29

Note by the AI writer (not discussed in the paper). Table 3 also shows how strongly the drying rate itself matters. The same 31 combinations that stayed AAAA or AAAC for a week after spin coating were mostly fully crystalline at once after slow drying. The claim that the compound decides holds across solvents and across fast routes, not across drying rates. X7

Spin coating against the melt (Table 4)

Table 4 cross-tabulates the two classifications as percentages of the 93 combinations (Table C). R16

Table C. Spin-coating class against melt class, percent of 93 drug–solvent combinations (from Table 4). R16

The diagonal adds up to 68%: identical classes. R17 In 30% the classes differ by one step (Class II in one method, Class I or III in the other), and in 2% they are opposite: Class I from the melt but Class III after spin coating, all of it one compound, atenolol in MIX and EtOH. R17 Each cell can be recomputed from Table 2 and the melt classes listed in the text; for example, 33 combinations are Class I by both, and 33/93 is 35.5%, printed as 35. E4

Per compound (Table D), 13 of the 14 Class I compounds were also Class I from the melt; dibucaine was melt Class II. R14 The Class II compounds split three ways: 4-biphenylcarboxaldehyde, carbamazepine and chlorpropamide were melt Class I, bifonazole and tolazamide melt Class II, and fenofibrate, nilutamide, nimesulide and probucol melt Class III. R14 Of the 13 Class III compounds, 10 were melt Class III, acetaminophen and celecoxib melt Class II, and atenolol melt Class I. R14 The four solvent-dependent compounds were melt Class II (clofoctol, flurbiprofen, cinnarizine) or Class III (loratadine). R14

Table D. The same comparison per compound, from the melt classes listed in the text (melt data cited from ref 1). R14

The melt study adds a few details that sharpen the extreme classes. Of the 13 compounds that were Class I by both methods, only flufenamic acid and tolbutamide could be made amorphous even by quenching the melt in liquid nitrogen; for the other 11 no amorphous phase was found at all. R15 Clofoctol and flurbiprofen crystallized on cooling when the rate was lowered to 1 °C/min; the three spin-Class II compounds that were melt Class I could all be quenched to glasses; nilutamide and nimesulide were the only melt Class III compounds that recrystallized on slower reheating (2 °C/min). R15

The authors find the agreement "remarkable" given the different time scales and media, and draw a practical conclusion: a drug that crystallizes readily during solvent evaporation, as in spray drying, "would also be anticipated" to crystallize easily from a melt, as in melt extrusion. C32 More generally, the similarity "suggests" that under rapid solidification the crystallization tendency is "dominated by the physicochemical properties of the API", the method of solidification being less important. C33

Note by the AI writer (not discussed in the paper). Because Table 4 counts combinations, compounds tested in three solvents weigh three times as much as those tested in one. Counted per compound, 25 of the 40 agree in every solvent (13 of 14 Class I, 2 of 9 Class II, 10 of 13 Class III), and the four solvent-dependent compounds agree in some solvents only. Agreement is excellent at the extremes and poor for Class II. X6

What separates the classes (Figures 3 and 4)

The PCA model needed three components to explain 97.6% of the variation in the seven descriptors (55.9%, 30.7% and 11.0%); cross-validation predicted 89.8% (31.3% with one component, 69.5% with two), which the authors call "good predictivity". R18·C34 On the plot of the first two scores (Figure 3), Class I compounds tend to sit on the left and lower side and Class III on the upper right, "although some exceptions can be observed"; Class II is scattered. R19·C35 Read from the plot, the class averages lie at about (−0.9, −0.6) for Class I, (0.4, −0.2) for Class II and (0.7, 0.7) for Class III. R19 The third component did not separate the classes. R19 Since no class information entered the model, the authors call this "a certain extent of differentiation" from the descriptors alone. C35

Figure 4 shows which descriptors drive the difference (Table E). R20

Table E. Class contributions: (class average − overall average)/overall standard deviation, read from Figure 4 (approximate). R20

Three patterns stand out. First, size and flexibility: molecular weight and the number of rotatable bonds tend to be higher for Class III and lower for Class I. R20·C36 The authors call this "intuitive" and in line with earlier work (ref 1; Yu et al. 2000, ref 31): to crystallize, a molecule must take the right orientation and conformation in the lattice, which is harder for large, flexible molecules. C37 Second, the driving force: heat and entropy of fusion are higher for Class I and lower for Classes II and III, so $\Delta G_v$ is more negative for Class I. R20·C38 By Eq. 3 a larger driving force lowers the nucleation barrier; the authors note that $W^{*}$ depends on ${\gamma }^3$ but only on $\Delta G_v^2$, and still find their correlation "consistent with theory", which "suggests" that differences in $\gamma$ between compounds are small compared with those in $\Delta G_v$. C39 Third, the glass transition: $T_g$ is low for solvent Class I (and II) and high for Class III, while $T_m$ differs little. R20·C40 In the melt study it was the other way round: some $T_m$ differences between classes, little $T_g$ discrimination. C40 The authors find this "readily understood": a drug that must stay amorphous for days at room temperature to earn Class III is favoured if its $T_g$ lies above room temperature, so that the film is glassy; the melt test is short, and every sample spends time above its $T_g$. C41

Note by the AI writer (not discussed in the paper). Two of these readings deserve care. The $\Delta G_v$ contribution is not monotonic across the classes: it is most negative for Class I but most positive for Class II, not Class III, so the driving-force link holds for Class I against the rest. X10 And $T_g$ was estimated from $T_m$ for 16 compounds of the melt-study set that crystallized too fast for a measurement; the 11 Class I compounds that gave no glass even on quenching are likely among them, so for Class I the $T_g$ contribution is partly a scaled $T_m$. X9

Predicting the class (Figure 5)

The class models explained 97.9% (Class I), 99.1% (Class II) and 96.1% (Class III) of the variability, with 87.9%, 88.1% and 76.7% predictable. R21 Every compound lay close to the Class III model, which therefore had "very limited discriminative potential" and was dropped. R21·C42 In the Coomans' plot of the Class I and Class II models (Figure 5), with critical distances of about 2.1 and 2.6 read from the plot, Class III compounds generally, "although not always", lie far from both models, Class I compounds often lie far from the Class II model, and Class II compounds cluster close to the Class I model in a small region; from this "some predictivity potential can be expected". R22·C43 Of the three validation compounds, PABA lay inside the Class I model only (about 0.9 and 7.9 from the two models) and clarithromycin far from both (about 8.6 and 7.3), and the authors call both "well predicted"; piroxicam lay inside both (about 1.95 and 0.95), which they read as "a potential Class II behavior, although other classifications might also be possible". R22·C44

Note by the AI writer (not discussed in the paper). The printed caption of Figure 5 cannot be read as it stands. It gives the key as white Class I, grey Class II, black Class III, the reverse of Figures 3 and 4, and it lists clarithromycin as Class II, whereas the text calls it the Class III validation compound. Only with the key of Figures 3 and 4 (black Class I, white Class III) do the plotted points match the text: the 15 white points match the 15 compounds counted as Class III, and the black triangle (PABA) lies inside the Class I model. The positions above are given with that reading. X1·X2

Discussion

The authors' picture. The paper's argument can be drawn as one chain that splits by class. Fast drying creates supersaturation but leaves almost no time for crystals to form in solution; whether the film then crystallizes depends on how fast this particular compound nucleates and grows, and that is set by its properties rather than by the solvent. C10·C11·C25·C33

Class II sits in between: the films start largely amorphous but crystallize within the week, from nuclei left by spin coating, from relatively fast nucleation in the film, or both ("may be due"). C20·C21 Size and flexibility, the driving force $\Delta G_v$ and $T_g$ relative to the storage temperature are the properties that, on average, separate the classes, but no single one does it alone: "consideration of a single descriptor cannot be used to understand crystallization behavior". R20·C45

Where the approach applies. The results describe neat drugs, made amorphous from solution within seconds, stored dry at room temperature for a week. P2·P8 Within that frame the solvent hardly matters and the route (solution or melt) matters little, especially at the extremes. R11·R17 Outside it, conditions matter a great deal: slow drying from the same solutions crystallized most Class III combinations at once. R13

Implications the authors draw. They state three: a drug that crystallizes readily in spray drying "would also be anticipated" to crystallize from a melt, as in melt extrusion C32; Class I compounds "will be much more challenging" to keep amorphous than Class III compounds and "may not be appropriate candidates for solid dispersion technology" C48; and spin coating "proved" a useful, low-consumption way to screen crystallization tendency. C47 On prediction they are more careful: in the results, a PCA model of easily accessible descriptors allows "a first pass a priori evaluation"; in the conclusions this becomes PCA "may provide some a priori prediction" of crystallization tendency, "with a suitable data set". C46

Note by the AI writer (not discussed in the paper). For formulation work the useful message is a ranking that travels between processes: a drug that will not stay amorphous on a spin-coated slide is unlikely to behave better in a spray dryer or an extruder, and the screen needs only milligrams and a week. The ranking is for the neat drug, though. No polymer, humidity or elevated temperature was tested, so Class I is a warning about how much a polymer would have to do, not a verdict that a dispersion cannot work. And the slow-drying result is a reminder that any slow solvent removal step (film casting, slow tray drying, rotary evaporation) is a much harsher test than fast drying for the same drug. R13·X7

Limitations

Limits the authors state. They note that the melt test probes the response to temperature changes over a short time scale, whereas their test probes storage at fixed temperature after evaporation, and that "the way in which the classification criteria are defined will influence the resulting classification". C30 They bound the solvent result to "the solvent systems and preparation method employed in this study". C24 They call the Class II prediction uncertain ("other classifications might also be possible") and drop the Class III model as non-discriminating. C44·C42 They note that interfacial free energies are "difficult to either measure or estimate". C39 And they make the predictive use of PCA conditional on "a suitable data set". C46

The AI writer's own judgement.

• The classification itself is direct and internally consistent: the class counts in the text match Table 2, and every cell of Table 4 can be recomputed from Table 2 and the melt classes in the text. R9·R16·E4
• The comparison with the melt rests on classes cited from the companion study and defined by different criteria; the 68% counts combinations, and per compound 25 of 40 agree in every solvent, with poor agreement for Class II. X6·C30
• Crystallinity was judged by eye on a five-step scale; the per-time-point scores are only in the Supporting Information, which was not available here. P8·R10
• The descriptor values are not printed, the class contributions can only be read from a bar chart, some $T_g$ values were estimated from $T_m$, and the $\Delta G_v$ trend is not monotonic across classes. R20·X9·X10
• The predictive claim rests on three validation compounds, one per class, of which one was ambiguous; the printed caption of Figure 5 reverses the colour key and mislabels clarithromycin. X8·X1·X2
• Nucleation and growth rates, on which the reading of each class rests, were not measured; these readings are hypotheses consistent with the data. C18·C20·C22
• The statement that the compound rather than the processing decides holds for fast drying; slow drying changed the outcome for 22 of 31 combinations. X7·R13

Conclusion

The paper set out to learn whether a drug's crystallization tendency belongs to the drug or to the way it is made amorphous. With 93 drug–solvent combinations of 40 drugs, spin coated and stored dry for a week, it found three well-separated behaviours: 39 combinations crystallized at once, 23 within the week and 31 stayed mostly amorphous. R9 The solvent changed the class of only 4 of 33 drugs tested in several solvents, and 68% of combinations received the same class as from the melt, with only one drug, atenolol, at the opposite extreme. R11·R17 Slow crystallizers tend to be larger and more flexible and to have a higher $T_g$; rapid crystallizers have larger heats and entropies of fusion and so a larger driving force to crystallize. R20·C36·C38·C40 The authors conclude that under rapid solidification crystallization is "governed mainly by the inherent properties of the compound" ("suggests"), that rapid crystallizers "may not be appropriate candidates" for solid dispersions, and that PCA "may provide some a priori prediction" of the class. C33·C48·C46 The comparison holds for fast drying only, and the prediction has so far been tested on three compounds. X7·X8