Will a drug melt crystallize or form a lasting glass? A three-class DSC screen of 51 organic molecules

Background and motivation

The drug in a solid oral dosage form may be one of several polymorphs, a solvate or an amorphous solid (Zhang 2004, ref 1). Because many new drug candidates have low aqueous solubility (Lipinski 1997, ref 2; Fahr 2007, ref 3), amorphous formulations, and process steps that make a drug amorphous, attract growing interest. In melt extrusion the crystalline drug melts and is cooled again; on the way down it can crystallize, to the original form or another, or be trapped as an amorphous solid. Physical stability during processing and storage decides whether the route is viable, and the authors call current understanding of crystallization from the undercooled melt (the liquid below its melting point), and of its link to the long-term stability of the glass, "marginal at best".

The textbook measure of how easily a liquid forms a glass, its glass-forming ability (GFA), is the critical cooling rate $R_{\mathrm{c}\mathrm{r}\mathrm{i}\mathrm{t}}$, the slowest cooling that still gives a glass (Nascimento 2005, ref 5). Any heat-stable material vitrifies if it is cooled below its glass transition temperature $T_g$ fast enough that nuclei cannot form or cannot grow to a detectable size. Measuring $R_{\mathrm{c}\mathrm{r}\mathrm{i}\mathrm{t}}$ needs laborious time–temperature–transformation (TTT) or continuous-cooling curves (Uhlmann 1972, ref 6; Onorato 1976, ref 7). Barandiaran and Colmenero (1981, ref 8) proposed a differential scanning calorimetry (DSC) shortcut, extrapolating the crystallization peak measured at several cooling rates, and Cabral et al. (2003, ref 9) refined it. Both worked reasonably for inorganic glasses (refs 9–11) but had not been tried on organic molecules, and $R_{\mathrm{c}\mathrm{r}\mathrm{i}\mathrm{t}}$ had rarely been determined for drugs (Kaushal 2008, ref 12; Damian 2001, ref 13).

Forming a glass is only half the problem. A glass has a higher free energy than its crystal and tends to crystallize (devitrify) over time, which usually removes the dissolution advantage; resistance to this is glass stability (GS). GS is usually judged by heating the glass, often above $T_g$ (refs 14–20), so the authors call the term "somewhat of a misnomer": what is measured is often the reheated undercooled liquid, not the glass. Despite work in metallurgy and pharmaceutics (refs 21–31, including Marsac 2006, ref 25), mechanistic understanding and predictive methods are lacking.

The field therefore uses simpler numbers from DSC thermal events. The reduced glass transition temperature $T_{rg}$ (Kauzmann 1948, ref 32; Turnbull 1969, ref 33) compares $T_g$ with the melting point. Glass-stability parameters compare the onset of crystallization of a heated glass with $T_g$ and the melting point (refs 5, 14, 16–19); for many inorganic glasses higher values went with lower measured $R_{\mathrm{c}\mathrm{r}\mathrm{i}\mathrm{t}}$ (Nascimento 2005, ref 5). Zhou et al. (2002, ref 20) applied a reduced temperature $T_{\mathrm{r}\mathrm{e}\mathrm{d}}$ to five pharmaceutical glass formers. Liquid fragility has also been linked to GFA: good glass formers tend to be "strong" liquids and marginal ones "fragile" (refs 36–38). Most of this evidence comes from inorganic glasses.

Links between GFA and GS have been pursued in metallurgy "with some success" (refs 4, 14, 15, 50), but for organic molecules such a correlation "has not been established or even attempted"; if it held, one screen could forecast both. The paper therefore sets out to screen by DSC how 51 organic molecules, described as "randomly selected", crystallize on cooling from the melt and on reheating from the glass, to build a classification from the result, and to look for links between GFA and GS. Along the way it tests the Barandiaran–Colmenero and Cabral estimates of $R_{\mathrm{c}\mathrm{r}\mathrm{i}\mathrm{t}}$, the standard thermal parameters and fragility, and asks by principal component analysis (PCA) which descriptors go with each class.

Background concepts

Glass formation as a race against crystallization. Crystallization needs stable nuclei to form and then to grow to a detectable size, so a melt becomes a glass only if it is cooled below $T_g$ faster than the maximum crystallization rate. In its Discussion (Eq. 21, Figure 6) the paper writes the cooling rate needed to avoid nucleation as

where ${\tau }_{N,\mathrm{m}\mathrm{i}\mathrm{n}}$ is the shortest nucleation time at any temperature between the melting point $T_m$ and $T_g$. In plain terms: the melt must cross the region of fastest nucleation in less time than a nucleus needs to appear there. The nucleation time follows from the nucleation rate (Eq. 22):

with $J_N$ the nucleation rate and $V$ the sample volume. In plain terms: the faster nuclei form, and the larger the sample, the sooner the first one appears. Nucleation is slow near $T_m$, where the driving force is small, and near $T_g$, where the viscous liquid "can restrict" diffusion; in between it is fastest, which gives the nose-shaped curve of a TTT diagram. Because nucleation is "an intrinsically faster process on a shorter length scale" than growth, the rate needed to avoid detectable crystals ($R_{\mathrm{c}\mathrm{r}\mathrm{i}\mathrm{t}}$) is slower than the rate needed to avoid nuclei ($R_{N,\mathrm{c}\mathrm{r}\mathrm{i}\mathrm{t}}$). A melt cooled between the two forms nuclei that can neither grow nor dissolve before mobility is lost: they are "quenched in", and a glassy-looking material "may in fact contain scores of nuclei" (refs 60, 61).

Reduced glass transition temperature. The oldest single-number predictor of GFA is (Eq. 4)

with both temperatures in kelvin. In plain terms: if all liquids have about the same viscosity at $T_g$, a liquid whose $T_g$ lies close to its melting point stays viscous over the whole range where it could crystallize, so the closer $T_{rg}$ is to 1, the better the glass former. A typical value of 0.66 had been proposed (ref 32).

Glass-stability parameters. A glass reheated in a DSC passes $T_g$ (midpoint $T_{g,\mathrm{m}\mathrm{i}\mathrm{d}}$), may crystallize (onset $T_{\mathrm{c}\mathrm{r}\mathrm{y}\mathrm{s}\mathrm{t},\mathrm{o}\mathrm{n}\mathrm{s}}$, peak $T_{\mathrm{c}\mathrm{r}\mathrm{y}\mathrm{s}\mathrm{t},\mathrm{m}\mathrm{a}\mathrm{x}}$) and finally melts ($T_{\mathrm{m}\mathrm{e}\mathrm{l}\mathrm{t}}$). "Generally", a larger gap between $T_g$ and the crystallization onset means a better glass former. Table A lists the normalised forms used (Eqs. 5–11), temperatures in kelvin.

Table A. The glass-stability parameters used in the paper (Eqs. 5–11). The authors describe $K_1$ and $K_2$ as identified by Nascimento et al. but "of unknown origin".

Fragility and strength. Near $T_g$ the relaxation time of a fragile liquid changes steeply and non-Arrheniusly with temperature, that of a strong liquid gently. In the Vogel–Tammann–Fulcher (VTF) equation (Eq. 12) this is captured by the strength parameter $D$: above 30 for strong liquids, below 10 for fragile ones, 7–15 for many pharmaceutical glass formers (refs 20, 44). The fragility index $m$, the slope of $\log \tau$ against $T_g/T$ at $T_g$ ($\tau$ the relaxation time; Eq. 13), is below 40 for strong and above 75 for fragile liquids (ref 45). It is obtained calorimetrically as (Eq. 14)

where $\Delta E_{T_g}$ is the activation enthalpy of structural relaxation near $T_g$, found from the shift of $T_g$ with scan rate (Moynihan 1974, ref 47; Eq. 15, heating rate equal to cooling rate), and $R$ is the gas constant. In plain terms: the higher the barrier to relaxation compared with the thermal energy at $T_g$, the more fragile the liquid. With $\tau$ at $T_g$ of about ${10}^2$ s and a limiting relaxation time of about ${10}^{-14}$ s (refs 44, 48), the floor is $m_{\mathrm{m}\mathrm{i}\mathrm{n}}=16$ (Eq. 17), and $m$ converts to $D$ by (Eq. 18)

In plain terms: the closer $m$ is to 16, the larger $D$ and the stronger the liquid. A second estimate of $D$ comes from the VTF equation with its zero-mobility temperature set to the Kauzmann temperature $T_K$, calculated from the melting point, heat of fusion and heat-capacity step at $T_g$ (Eqs. 19, 20). Table 9 lists both, as $D^a$ (from $m$) and $D^b$ (from $T_K$).

Detecting a glass. Crystals shine between crossed polarizers and give sharp Raman bands; a glass stays dark and gives broader bands (refs 55–57). Temperature also shifts and broadens bands, so a sample quenched in liquid nitrogen (LN2, −196 °C) must be compared with a crystalline reference at the same temperature.

Methods in detail

The work splits into four contributions: the screen itself, a fixed DSC heat–cool–heat cycle refined by an LN2 quench and checked at other scan rates; a test of whether the standard predictors of GFA and GS reproduce the classes; a PCA of which descriptors go with each class; and storage of quenched glasses below $T_g$, interpreted with a TTT diagram and quenched-in nuclei.

Contribution 1: A heat–cool–heat DSC screen and a three-class scheme

① Problem. A GFA screen must work on many molecules with little material, which TTT-based critical cooling rates cannot, and should also say something about the glass. No comparable data existed for a large organic set.

② What they did. A TA Q2000 DSC (hermetic pans, nitrogen purge 50 mL/min) heated each sample at 10 °C/min to about 10 °C above its melting point, held it 3 min, cooled it at 20 °C/min to −75 °C and reheated it at 10 °C/min to just above the melting point, in triplicate (absolute sample mass not printed). Class (I): crystallization on cooling before the $T_g$ event. Class (II): no crystallization on cooling, but crystallization on reheating above $T_g$. Class (III): no crystallization on cooling or on reheating to the melting point. Because some class (I) molecules had been reported amorphous after DSC cooling or quenching (antipyrin, carbamazepine, griseofulvin, indoprofen; refs 51–54), class (I) melts were plunged into LN2, inspected between crossed polarizers (1–2 mg on glass slides), and measured by Raman spectroscopy (785 nm) while still immersed, so the laser could not warm and crystallize them. Molecules still crystalline became class (I-A); those trapped amorphous, class (I-B).

The 20 and 10 °C/min rates were "chosen arbitrarily", so classes (II) and (III) were cooled at 1 °C/min; class (II) was reheated at 20 °C/min, and class (III) at 2 °C/min (triplicate) and 10 °C/min. Class (II) was also cooled at 1, 5 and 20 °C/min to read its crystallization temperature on reheating, $T_{c,\mathrm{h}\mathrm{e}\mathrm{a}\mathrm{t}}$, and $T_{\mathrm{r}\mathrm{e}\mathrm{d}}$ (Table 7; reheating rate not stated).

③ Why it works. One run asks two questions: does the melt outrun crystallization at a moderate rate (a coarse test of GFA), and does the glass resist crystallization once mobility returns above $T_g$ (the usual proxy for GS)? The answers define the class without any model. The LN2 quench adds a much faster, though unmeasured, cooling rate, and two independent read-outs guard the amorphous calls.

④ How it differs from earlier work. Earlier work used few compounds or one parameter (Zhou et al., ref 20, studied five). Here one protocol covers 51 molecules and gives a categorical observation, combining cooling and reheating behaviour, rather than a model-derived number.

Contribution 2: Testing the standard predictors against the classes

① Problem. If a simple number ranked GFA, no screen would be needed. Would numbers validated mainly on inorganic glasses reproduce what the screen observes?

② What they did. Critical cooling rate methods need a crystallization peak on cooling, so only class (I) qualified. These molecules were cooled at 5, 10, 15 and 20 °C/min (Table 5); where fast crystallization warmed the thermocouple, the peak temperature was estimated from the time to maximum heat flow multiplied by the cooling rate. The Barandiaran–Colmenero (BC) relation (Eq. 1) was fitted:

Here rate is the cooling rate (K/min), $\Delta T$ the undercooling (how far below the melting point the melt crystallizes), $T_{\mathrm{c}\mathrm{r}\mathrm{y}\mathrm{s},\mathrm{m}\mathrm{a}\mathrm{x}}$ the crystallization peak on cooling, and $A$ and $B$ constants from a linear regression of $\ln (\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{e})$ against $1/(\Delta T{)}^2$. In plain terms: faster cooling should push crystallization lower, and the relation extrapolates to the rate at which it would vanish. Letting $\Delta T$ go to infinity gives (Eq. 2)

In plain terms: the critical cooling rate is the intercept of the fitted line. Cabral et al. argued that $\Delta T$ cannot exceed $T_m-T_g$ and wrote (Eq. 3)

with $A^{*}$ and $B^{*}$ the same constants and $T_m$ in kelvin. In plain terms: the BC intercept is corrected downward by a term that depends on the melting point. Card check: the text motivates the correction by the limit $T_m-T_g$, but the printed equation divides by $T_m^2$, and Table 6 reproduces with $T_m^2$ (griseofulvin: 588 × exp(−3711/491.15²) = 579 K/min). Table 6 also gives $A$ the unit K/min although it is the logarithm of a rate (X6). The four molecules that crystallized only partly on cooling were not fitted, leaving 19.

$T_{rg}$ was calculated for every molecule. $T_g$ was measured for classes (II) and (III) and five class (I) molecules, taken from the literature for griseofulvin and indoprofen (refs 53, 54), and estimated for the other 16 class (I) molecules, which crystallized before any $T_g$ appeared. For that, $T_{rg}$ was fitted against molecular weight (MW) for the 35 compounds with known $T_g$ (Figure 1):

with MW in g/mol. In plain terms: in this set heavier molecules tend to have $T_g$ closer to their melting point, so MW gives a rough $T_{rg}$ and, with the measured melting point, a rough $T_g$. The fitted $T_g$ deviated from the line by 9.3 K on average, which the authors call acceptable. The parameters of Table A were calculated for class (II) and for three class (I-B) molecules that crystallized partly on cooling (chlorpropamide, flufenamic acid, tolbutamide); fragility and strength for classes (II) and (III), from the shift of $T_g$ with equal cooling and heating rates of 5–20 °C/min and from the heat-capacity step at $T_g$.

③ Why it works. The classes and the quench are direct observations, so the design is a falsification test: a GFA predictor should place class (I) below class (III), separate (I-A) from (I-B), and flag the class (II) molecules that crystallized at 1 °C/min. A molecule with a low $R_{\mathrm{c}\mathrm{r}\mathrm{i}\mathrm{t}}$ estimate should vitrify in LN2 at least as easily as one with a high estimate.

④ How it differs from earlier work. These tools had been validated on inorganic glasses (refs 5, 9–11); here the reference is what organic molecules did in the instrument.

Contribution 3: Molecular descriptors by principal component analysis

① Problem. If the usual parameters fail, do basic thermal and molecular properties differ between classes?

② What they did. A PCA model (SIMCA-P+) was built for all 51 molecules from seven descriptors: melting onset $T_{\mathrm{m}\mathrm{e}\mathrm{l}\mathrm{t},\mathrm{o}\mathrm{n}\mathrm{s}}$, $T_g$ midpoint, $\Delta H_{\mathrm{f}\mathrm{u}\mathrm{s}}$, entropy of fusion $\Delta S_{\mathrm{f}\mathrm{u}\mathrm{s}}=\Delta H_{\mathrm{f}\mathrm{u}\mathrm{s}}/T_{\mathrm{m}\mathrm{e}\mathrm{l}\mathrm{t},\mathrm{o}\mathrm{n}\mathrm{s}}$, MW, number of rotatable bonds (#RB) and the crystal–amorphous free-energy difference from the Hoffman approximation at $T$ = 293.15 K:

In plain terms: the further room temperature lies below the melting point, and the larger the heat of fusion, the more free energy is gained by crystallizing. The three energy terms were put on a volume basis with the crystal density, and the 16 estimated $T_g$ values entered the model. Variables were centred and scaled, components chosen by cross-validation, and "no information was provided on the crystallization behavior (classification)". A class contribution (Figure 4) is the class average minus the data-set average, divided by the data-set standard deviation.

③ Why it works. An unsupervised model cannot be fitted to the labels, so any clustering by class reflects the descriptors. The authors also argue why the descriptors should matter. A flexible molecule exists in the melt as several similar conformers and must find the "right" one (ref 72), so more rotatable bonds should raise GFA. Thermodynamically, the enthalpy and entropy differences between crystal and amorphous state below the melting point are (Eq. 23)

where $\Delta C_p$ is the heat-capacity difference between amorphous and crystal. In plain terms: for equal $\Delta C_p$, a lower heat and entropy of fusion mean smaller differences at lower temperature, and so a smaller driving force to crystallize. Classical nucleation theory gives the barrier to a critical nucleus (Eq. 24; refs 73, 74):

where $\gamma$ is the interfacial energy per area between nucleus and amorphous matrix. In plain terms: a larger free-energy gain lowers the barrier with its square, so for similar $\gamma$ a more negative $\Delta G_v$ "would result" in higher homogeneous nucleation rates. Equal $\Delta C_p(T)$ and similar $\gamma$ are assumptions, not measurements.

④ How it differs from earlier work. Earlier attempts used one parameter at a time; taken together, several show that no single property decides the class.

Contribution 4: Storing quenched glasses below $T_g$, and the quenched-in-nuclei explanation

① Problem. The stability parameters and the reheating step measure crystallization above $T_g$, where rates are "much higher"; shelf life depends on crystallization below $T_g$, for which "very little data exists".

② What they did. Melts of indoprofen (class I), droperidol and nifedipine (class II), and clotrimazole and felodipine (class III) were quenched in LN2 between two cover glasses and stored over phosphorus pentoxide (0% relative humidity, RH) 14–20 °C below $T_g$ (Table F). Crystallization was followed by polarized light microscopy at 5 min, 1 day and 3 months (Figure 5). Card check: the paper says the class (II) and (III) samples were stored "at 15 °C below each molecule's respective Tg"; from the printed $T_g$, droperidol was 14 °C below. To interpret the outcome the authors annotated a TTT diagram adapted from Angell (2008, ref 61; Figure 6): class (II) cooled faster than $R_{\mathrm{c}\mathrm{r}\mathrm{i}\mathrm{t}}$ but slower than $R_{N,\mathrm{c}\mathrm{r}\mathrm{i}\mathrm{t}}$ gives "a glass ... with quenched in nuclei", and class (III), cooled faster than both, "a nuclei free glass". Card check: the Figure 6 caption places class (I) where cooling "was lower than RNcrit", also true of class (II); the drawing places it beyond the $R_{\mathrm{c}\mathrm{r}\mathrm{i}\mathrm{t}}$ line (X5).

③ Why it works. The design pairs a class (II) and a class (III) molecule at the same storage temperature and nearly the same $T_g$ (droperidol with clotrimazole, nifedipine with felodipine; Table F). Note by the AI writer (not discussed in the paper): the pairing removes distance below $T_g$ as an explanation within each pair. The TTT reading links all observations to one cause: nuclei formed where nucleation is fast but growth slow meet the maximum growth rate above $T_g$, and below it they seed crystals as the glass relaxes (refs 68, 69).

④ How it differs from earlier work. GS had usually been judged above $T_g$; here it is observed below $T_g$ on glasses of known class. Quenched-in nuclei were known from metallic and some organic glasses (refs 62–67); the new step is using them to explain why two easily vitrified classes differ in stability.

Experimental analysis

Setup in brief. The 51 neat molecules (melting points 59–272 °C, MW 121–721 g/mol) came from suppliers or as gifts; no polymer or water was involved, and starting crystal forms are not reported.

Table B. Classification of the 51 molecules (cool 20 °C/min, reheat 10 °C/min, triplicate; Tables 1–3), with the LN2 split of class (I) (Table 4) and the rate tests reported in the Results text.

Main result 1: three classes, and a split within class (I). Class (I) "can be defined as" having a high crystallization tendency from the undercooled melt and so a lower GFA than classes (II) and (III). That four class (I) molecules crystallized on both cooling and reheating is "an indication" that part of each sample was trapped amorphous. After the LN2 quench 15 class (I) molecules were birefringent and 8 were not, and Raman broadening marked the same 8 (Table 4). Figure 2 shows the three outcomes: benzocaine kept its starting form (spectra overlap at about 1604 cm⁻¹); anthranilic acid crystallized even in LN2, to the polymorph also formed on air cooling (about 1350 against about 1371–1376 cm⁻¹); quenched griseofulvin gave a broader carbonyl band (maximum about 1710 cm⁻¹, extending to about 1725) than the references (about 1706–1708), "clear" evidence for the authors of an amorphous solid. Card check: the Figure 2 caption calls panel (b) the anthranilic acid "CH3 symmetrical stretch region", but the molecule has no CH3 group (X4).

No class (III) molecule crystallized at 1 °C/min cooling or 10 °C/min reheating, and the authors conclude that the classification "appears to be robust and reasonably independent" of the rates used. Counted by compound, though, 7 of 28 class (II) and (III) molecules behaved unlike their class in at least one test, so for a quarter of them the class is specific to the 20/10 °C/min protocol (X12). Structure was a poor guide: benzocaine and lidocaine are class (I) but procaine class (III), and haloperidol class (I) but droperidol class (II). Several class (I) and (II) molecules recrystallized to a different polymorph, "not unexpected" because the undercooled melt "frequently" gives a metastable form. Card check: the DSC polymorph calls for indoprofen, benzamide, caffeine, carbamazepine and flufenamic acid differ between Table 4 and the Table 5 footnotes (X2).

Table C. Critical cooling rate fits for class (I) (Table 6, Eqs. 1–3), sorted by estimated $R_{\mathrm{c}\mathrm{r}\mathrm{i}\mathrm{t}}$, against the LN2 quench outcome (Table 4) and the mean undercooling at each cooling rate (Table 5).

Main result 2: undercooling is erratic, and fitted critical cooling rates contradict the quench. At 20 °C/min the undercooling ranged from 3.0 °C (caffeine) to 80.8 °C (tolbutamide). Antipyrin undercooled by 80.1 °C and 4-biphenylmethanol by 10.1 °C, although the authors call them physicochemically similar (thermal properties and MW are close, logP is not: 0.26 against 3.13; X18). Undercooling often did not rise steadily with cooling rate, and replicate scatter could be large (antipyrin: standard deviation 35.6 °C at 10 °C/min), "most likely" because of stochastic and unavoidable heterogeneous nucleation. Seven fits had r² of 0.74–0.99, twelve below 0.55. The usable estimates (20 to 3.26 × 10⁴ K/min) clashed with the quench: haloperidol (20 K/min) and phenacetin (64 K/min) crystallized in LN2, while carbamazepine (118 or 117 K/min) and griseofulvin (588 or 579 K/min) vitrified. The authors conclude that the approach "may be of limited utility for organic molecules"; for inorganic glasses, accurate predictions need rates at least two orders of magnitude below $R_{\mathrm{c}\mathrm{r}\mathrm{i}\mathrm{t}}$, so rates spanning several orders of magnitude "may be needed", against the 5–20 °C/min used.

Note by the AI writer (not discussed in the paper): sorted as in Table C, the four lowest estimates belong to molecules that crystallized in LN2 and the next two to molecules that vitrified. For six of seven molecules a higher estimate went with easier vitrification, the reverse of what $R_{\mathrm{c}\mathrm{r}\mathrm{i}\mathrm{t}}$ means. This argument does not need the unknown LN2 cooling rate: whatever it was, haloperidol at 20 K/min should have vitrified at least as easily as griseofulvin at 579 K/min. And because $B/T_m^2$ is small, the Cabral values differ from BC by at most about 2%, so the two columns are not independent checks.

Main result 3: $T_{rg}$ does not separate the classes. Mean $T_{rg}$ was 0.69 (±0.01) for class (I-A) and 0.7 (±0.03) for class (I-B) molecules with measured $T_g$, "not statistically different" (test not named), so $T_{rg}$ did not predict which could be quenched amorphous. Class (I) averaged 0.69 ± 0.02 and class (III) 0.73 ± 0.04, the expected order but not a significant difference, so $T_{rg}$ is "clearly not an infallible predictor". Within class (III), the two molecules that crystallized on slow reheating did not have lower values (nilutamide 0.72, nimesulide 0.70). Card check: all 15 class (I-A) $T_g$ values come from the MW regression, so their $T_{rg}$ is a function of MW alone (X17); the stated class (I-A) mean of 0.69 compares with 0.68 from the Table 1 values (X8).

Table D. Class (II): predictors against observed behaviour. $T_{\mathrm{r}\mathrm{e}\mathrm{d}}$ after 20 °C/min cooling (Table 7), $K_H$ and $K_{SP}$ (Table 8), $m$, $D^a$ and $D^b$ (Table 9), and the rate tests reported in the Results text.

Main result 4: within class (II), stability parameters and fragility pick the wrong molecules. Assuming the inverse relation to $R_{\mathrm{c}\mathrm{r}\mathrm{i}\mathrm{t}}$ seen in inorganic glasses, higher $T_{\mathrm{r}\mathrm{e}\mathrm{d}}$ and stability parameters pointed to bifonazole, celecoxib, cinnarizine and droperidol as the best class (II) glass formers, and lower values to acetaminophen, D(−)salicin, flurbiprofen and nifedipine as the worst. In fact celecoxib, dibucaine, flurbiprofen and nifedipine crystallized at 1 °C/min (Table D): one predicted best was among them, two predicted worst were not. Celecoxib had the highest $T_{\mathrm{r}\mathrm{e}\mathrm{d}}$, $K_{LL}$, $K_H$ and $K_1$ but the lowest $K_{SP}$ and $K_2$, which the authors attribute ("is believed to be") to its high $T_g$ and crystallization temperature relative to its melting point; the ranking thus depends on the parameter chosen (X15). Fragility did no better: $m$ ran from 18 to 731 (itraconazole) across classes (II) and (III) without correlating with the stability parameters, and the combined parameters favoured celecoxib, droperidol and nifedipine, but "again the experimental evidence does not support this conclusion". The authors' summary: none of the parameters "showed a strong correlation with the observed crystallization behavior", "suggesting these parameters may not be ideal" for organic molecules. Card check: $T_{\mathrm{r}\mathrm{e}\mathrm{d}}$ recomputed from the printed temperatures is 0.006–0.029 lower for all 11 molecules (acetaminophen 0.318 against 0.33; X8). Chlorpropamide's $K_{SP}$ and $K_2$ are both printed as 1.873 although their ratio must be 1.373 (X1). Tolazamide, nifedipine and aceclofenac are strong by $m$ and $D^a$ but fragile by $D^b$ (X14). The Table 7 20 °C/min column is a cooling rate, not the fast-reheating test (X9).

Table E. Class contributions in the PCA, (class average − data-set average) / data-set standard deviation, read from Figure 4 (approximate, about ±0.02).

Main result 5: descriptors separate class (I) from class (III) on average. Three components explained 97.8% of the variance (56.3%, 30.4%, 11.0%), and cross-validation predicted 91.1% (33.9% with one component, 70.4% with two), "indicating acceptable predictivity". In the score plot (Figure 3) the class means sit at about (−1.1, −0.3) for class (I), (0.0, 0.0) for class (II) and (1.4, 0.4) for class (III), with outliers in every class (most notably lidocaine and tolbutamide in class I). Table E shows the direction: class (I) molecules are on average smaller and more rigid, with higher heats and entropies of fusion, higher melting onsets and a more negative $\Delta G_v$; class (III) the opposite; class (II) near the mean. Mean $T_g$ barely differs, "probably" because cooling from the melt takes every sample above $T_g$. Multiple properties "appear to play a role"; no single descriptor predicts the class. Note by the AI writer (not discussed in the paper): the 91.1% describes how well the model reproduces the descriptors, not how well it assigns classes, and no classification accuracy is given; the separation is a shift of means with overlapping clouds.

Table F. LN2-quenched glasses stored between two cover glasses over P₂O₅ (0% RH), judged by polarized light microscopy (Figure 5 and text). No replicate count or crystallinity measurement is given.

Main result 6: in storage, crystallization speed follows the class. Quenched indoprofen "began to crystallize almost immediately", 20 °C below its $T_g$, and was fully crystalline after 1 day. Droperidol and nifedipine crystallized at the edges within a day and "appeared to be completely crystalline" after 3 months. For clotrimazole and felodipine crystallization was "extremely slow", with only edge crystallization after 3 months. With the reported rapid crystallization of griseofulvin glass (ref 53), the authors conclude that "there may be a link" between GFA and GS, class (I) being the least and class (III) the most stable, and that class (II), though it vitrifies as easily as class (III), "appear[s] to have much lower GS". The Conclusions go further: "a potential link between GFA and GS has been established". Card check: indoprofen is labelled class (I-A) in this section, but Table 4 shows it amorphous after the LN2 quench, which defines class (I-B) (X3). In Figure 5, panel (c) (droperidol, "completely crystalline") shows bright crystals mainly along one edge, while panel (d) (clotrimazole, "only edge crystallization") shows a birefringent fan over most of the field (X16).

Controls and what they rule out. The label-free PCA rules out a separation forced by fitting; same-temperature Raman references rule out thermal band changes as the source of broadening; microscopy and Raman agreed on the same eight class (I-B) molecules; the rate tests clear 21 of 28 class (II)/(III) assignments of dependence on the arbitrary rates; and the matched storage pairs rule out distance below $T_g$. Not ruled out: surface crystallization (assumed minimized by the cover-glass sandwich), differences in starting form or purity, heterogeneous nucleation on the pan, and sample-mass effects on detecting small events.

Discussion

Where the screen applies. The class is an observation under one protocol (a few milligrams of neat drug, cooled at 20 °C/min and reheated at 10 °C/min) for molecules of MW 121–721 g/mol and melting points 59–272 °C. Degradation in the melt was not assessed (note by the AI writer). The (I-A)/(I-B) boundary belongs to an LN2 quench of unmeasured rate, and the stability link was shown only at 0% RH, between cover glasses, over 3 months.

What is observed and what is interpreted. The class counts, the LN2 split, the failure of the classical predictors, the average descriptor differences and the storage order of five glasses are observations. The quenched-in-nuclei mechanism is interpretation, worded as such ("It is believed", "it seems ... the most likely"). The authors concede that "it is possible that some nuclei are formed during cooling for class (III) molecules" and that "it may be that the growth of these nuclei is very slow". Note by the AI writer: that concession means the data cannot distinguish "fewer nuclei" from "slower growth" in class (III); both readings predict the same DSC traces and storage order.

Why the classical predictors may fail. The authors' view is that parameters developed for inorganic glasses "may not be ideal" for organic molecules, and that the BC and Cabral methods need far wider cooling-rate ranges. Note by the AI writer: the parameters rest on single crystallization temperatures, which scatter widely between replicates here, and on thermal history (cinnarizine's $T_{\mathrm{r}\mathrm{e}\mathrm{d}}$ is 0.36 after 1 °C/min cooling, 0.56 after 20 °C/min). The class, a yes-or-no outcome of a fixed protocol, is less exposed to both.

Open questions and further work. The authors ask whether $R_{\mathrm{c}\mathrm{r}\mathrm{i}\mathrm{t}}$ models would work over a far wider range of cooling rates, how crystallization above $T_g$ relates to that below it, and how far the GFA–GS link reaches ("This area clearly warrants further investigation"). Suggestions by the AI writer, following the paper's own TTT picture: if quenched-in nuclei are the cause, holding a class (III) glass near the temperature of fastest nucleation before reheating should make it behave like class (II), and cooling a class (II) melt faster should do the reverse. The stability link would also need more molecules per class, several temperatures, a crystallinity measure, humidity, open surfaces and polymers.

What it suggests for formulation. The authors offer the screen as "a potentially useful tool" to assess crystallization tendency, GFA and potential GS of new chemical entities early, reducing unneeded experiments. The rest is the AI writer's reading, not the paper's. A class (I) drug will be hard to keep amorphous by cooling its neat melt, as in melt quenching or hot-melt extrusion: 15 of 23 could not be vitrified even in LN2, and the other 8 crystallized quickly on reheating. Class (II) is the subtle case: the cooling trace is clean, but the glass recrystallized on reheating and both stored examples appeared fully crystalline within 3 months, so a clean cooling curve is not enough. Class (III) was the most forgiving. A borderline drug should be retested at the rates of the real process, and for neat drugs at 0% RH the class is a first risk flag, not a stability prediction for a formulation.

Limitations

Limitations stated by the authors. The screen's rates were "chosen arbitrarily", hence the rate tests. For the critical cooling rate, "the range in cooling rates employed was relatively small", and crystallization temperatures scatter because of stochastic nucleation and "the impossibility of avoiding heterogeneous nucleation". They note that "it is not easy to detect the presence of nuclei, thus crystallization rather than nucleation is being measured", which bears on their own mechanism. $T_g$ had to be estimated for 16 class (I) molecules. Interpreting crystallization through molecular properties is "difficult given the complexity of the crystallization process", and in the Results the GFA–GS link is framed as "there may be a link".

The AI writer's own judgement (1): the classification and the predictor tests. The extractor rates the overall evidence as moderate. The classes (triplicate runs of 51 molecules) and the predictor failures (visible in the tables) are the paper's firmest results. Even so, 7 of 28 class (II) and (III) molecules behaved unlike their class in a rate test (X12), and the Discussion's statement that both classes form glasses "at cooling rates as slow as 1–2 °C min−1" conflicts with four class (II) molecules crystallizing at 1 °C/min, while no 2 °C/min cooling was run (X13). The $T_{rg}$ comparisons name no statistical test and use MW-derived values for class (I-A) (X17). $T_{\mathrm{r}\mathrm{e}\mathrm{d}}$ and the stability parameters cannot be reproduced from the printed temperatures (X8), one value is a misprint (X1), and both the class (II) ranking and the fragility category depend on which parameter is chosen (X14, X15). None of this reverses the negative verdict on the predictors, but the individual numbers are soft.

The AI writer's own judgement (2): the stability link and the mechanism. The GFA–GS link rests on one class (I) compound with an inconsistent label (X3) and two compounds each for classes (II) and (III), each at one temperature, judged by eye without a crystallinity measure or reported replicates; two micrograph panels do not visibly match the text (X16). The matched pairs are a strength, but with n = 2 per class the ranking is qualitative, and "there may be a link" fits the data better than "has been established". The mechanism was not tested: nuclei were not detected, nucleation and growth rates were not measured, and the LN2 cooling rate is unknown. Smaller issues: the "randomly selected" set is not described, niclosamide is listed but unreported (X11), and MW provenance is stated two ways (X10).

Conclusion

The paper shows that one DSC heat–cool–heat cycle sorts organic molecules into crystallization on cooling (class I, 23 of 51), crystallization only on reheating (class II, 11) and neither (class III, 17), and that an LN2 quench splits class (I) into 15 molecules that cannot be vitrified and 8 that can. The tools carried over from inorganic glasses did not reproduce these classes: fitted critical cooling rates contradicted the quench, and $T_{rg}$, the stability parameters and fragility neither separated the classes nor ranked class (II) correctly. Class (I) molecules were on average smaller and more rigid with higher heats of fusion, class (III) larger and more flexible, but no single descriptor decided the class.

The link to glass stability is the tentative half. It rests on five stored glasses, and the quenched-in nuclei that explain it fit the TTT picture but were not measured. Within those limits the screen gives formulators a cheap first flag for melt-based amorphous processing.