Designing cellulose crystallization inhibitors: side-chain structure, molecular dynamics and surface adsorption

Background and motivation

Amorphous solid dispersions (ASDs) raise the oral bioavailability of poorly water-soluble drugs because they dissolve to concentrations above the crystalline solubility, that is, to a supersaturated solution. Such a solution is metastable: sooner or later the drug nucleates and crystallizes, and the advantage is lost. Polymers are added to the formulation partly to delay this crystallization, by slowing nucleation and crystal growth. The abstract states plainly that there is "limited evidence" on how polymers achieve this, and the introduction notes that variations in polymer chemistry "drastically influence" how long they hold supersaturation (refs 7, 8). C1 The paper frames two questions: which chemical groups make a polymer an effective inhibitor, and what those groups actually do in solution. C1

The Edgar and Taylor groups had been building new cellulose derivatives to answer the first question. Olefin cross-metathesis lets carboxylic acids, esters and other groups be attached to cellulose under mild conditions, and a following thiol-Michael addition can add a second branch to the side chain (refs 13, 14, 32–36). P1·P2 In earlier work, polymers carrying carboxylic acids outperformed polymers with amides, amines or alcohols (Mosquera-Giraldo 2016, ref 7), and polymers with a higher degree of substitution of carboxylic acid appeared to be better inhibitors, although they differed in other features too (Ilevbare 2013, ref 8). C26 One result from ref 7 is the starting point of this paper: two cellulose acetate derivatives with the same carboxylated side chain, A5a and A5b, differed only in the length of the hydrocarbon tether that joins the side chain to the cellulose ring, yet the short-tether A5a was much better for telaprevir, a change the authors described as a "drastic impact". R13 Their reading then was that a polymer that associates with itself has fewer chances to meet the drug. C27 Whether this held for drugs other than telaprevir was open.

For the second question, computation had mostly meant gas-phase quantum-chemical binding energies between one polymer monomer and one drug molecule (refs 17, 28, 29), or molecular dynamics (MD) used to estimate an enthalpy of mixing and from it a Flory–Huggins interaction parameter (refs 21, 23). C3·C4 The authors list four limits of the monomer approach: the chosen conformations favour interactions such as hydrogen bonds; once an ASD meets water, drug–polymer hydrogen bonds are "probably largely disrupted"; quantum methods only handle monomers and dimers; and gas-phase conformations may differ from those in solution. C3 Atomistic MD of whole chains in water, they argue, can still estimate how favourable an interaction is, and combined experimental and computational work had "shown promise" (ref 7). C4

The paper therefore tests a hypothesis it calls multifaceted: that synthesizing new cellulose polymers, measuring their crystallization inhibition against a deliberately diverse set of nine drugs, imaging how they adsorb on an amorphous drug surface and simulating a whole polymer chain next to a drug cluster in water, together, can reveal the chemical features that make a good polymeric crystallization inhibitor. C5

Background concepts

Supersaturation, induction time and the classes of crystallizer. A drug solution above its crystalline solubility can stay clear for a while before crystals appear. The time until crystals are first detected in an unseeded solution is the nucleation induction time; a longer induction time with a polymer present means the polymer inhibits crystallization. Here it was followed with a UV/Vis fibre-optic probe: the time at which absorbance at a non-absorbing wavelength rises (light scattered by particles) while absorbance at the drug's absorbing wavelength falls (drug leaving solution) is taken as the induction time. P4 The authors group the nine drugs into fast crystallizers (griseofulvin, celecoxib, ezetimibe, danazol), intermediate ones (telaprevir, nifedipine, nevirapine) and slow ones (ritonavir, atazanavir), without stating the basis of the classes. M1–M9

LLPS and GLPS concentrations. If a supersaturated solution is pushed above a certain concentration, it separates into a drug-rich phase, which is a liquid (liquid–liquid phase separation, LLPS) or, for drugs whose drug-rich phase is glassy at the test temperature, a glass (glass–liquid phase separation, GLPS). That concentration, the amorphous solubility, is the highest supersaturation a single solution phase can hold. Table S2 lists the crystalline solubility and the LLPS or GLPS concentration of each drug, all taken from earlier papers of the group. M1·M3 The experiments were meant to stay below the LLPS or GLPS concentration so that crystals nucleate from a homogeneous solution; the exception is telaprevir, tested at 150 µg/mL against a GLPS concentration of 90 µg/mL. P4·M3

Cellulose derivatives and degree of substitution. Each anhydroglucose unit (AGU) of cellulose has three hydroxyl groups (positions C2, C3 and C6) that can be substituted. The degree of substitution (DS) is the average number of a given substituent per AGU, so the DS values of all substituents on a ring add up to 3. The esters here start from cellulose acetate (acetyl DS 1.82 in Table 1), the ethers from an ethyl cellulose derivative (ethyl DS 2.3). M10·M12 On top of these, a carboxylated side chain is attached at DS 0.67–0.79. M10·M11·M12 In the esters it is linked through an ester bond, in the ethers through an ether bond. M10·M12

What the simulations measure. The radius of gyration ($R_g$) of a polymer chain is the root-mean-square distance of its atoms from the chain's centre of mass: a larger $R_g$ means a more extended main chain. The solvent accessible surface area (SASA) per atom measures how much of each atom is exposed to water; in the authors' earlier work it tracked the conformation of the side chains rather than the backbone. C18 The radial distribution function (RDF) between polymer and drug gives the relative likelihood of finding polymer atoms at a distance $r$ from the drug molecules; a high peak at short distance means the two are in contact. P11 The free energy of solvation estimates how favourably a chain interacts with water, computed in GROMACS from the exposed area of each atom. P7

The free energy of interaction. To put one number on the polymer–drug interaction, the authors extracted the Lennard-Jones and Coulomb energies between the polymer and the drug energy groups along the 10 ns production run and averaged them exponentially (Eq. 1): P10·E1

In plain terms: $U$ is the potential energy of the polymer–drug interaction at each recorded step, $k$ is the Boltzmann constant, $T$ is the simulation temperature (310 K) and the angle brackets denote the average over all recorded steps; the result is a free energy relative to a reference state in which polymer and drug do not interact (the ideal gas). E1 A more negative value means a more favourable interaction. C25 The authors present it as a "general indication of the likelihood of a polymer-drug interaction", borrowed from protein–ligand practice, and they stress that it carries no kinetic information. C42·C45 The same section names the Fedors group-contribution method, used to compute a solubility parameter of each polymer from its structure; it is named, not printed, and works only for neutral structures. E2·P12

Methods in detail

The study has four parts that run side by side: a set of polymers designed in pairs, induction-time measurements with nine drugs, AFM imaging of polymer adsorption on an amorphous drug film, and MD of single polymer chains alone and next to a cluster of drug molecules. The flow below shows how they fit together.

Contribution 1: polymers that differ in one feature at a time

The polymers are best read as two series (Figure 1, Table 1). The two cellulose esters, CA-A5a-079 (A5a) and CA-A5b-067 (A5b), were made from cellulose acetate CA-320S (DS 1.8) by esterification with pent-4-enoyl chloride or undec-10-enoyl chloride, then cross-metathesis with hydroxyethyl acrylate and thiol-Michael addition of 3-mercaptopropionic acid. P1·M16 Both end up with a side chain that carries one carboxylic acid (from the thiol) and one hydroxyethyl ester branch ending in an alcohol; what differs is the tether, two methylene groups in A5a and eight in A5b. M10·M11 Their carboxylated-substituent DS is 0.79 and 0.67, and their residual hydroxyl DS 0.39 and 0.51. M10·M11

The four cellulose ethers start from EC2.30C5, an ethyl cellulose with pentenyl ether groups made from microcrystalline cellulose in one pot with ethyl iodide and 5-bromo-pent-1-ene (ethyl DS 2.3, side-chain DS 0.69). M17·P2 Cross-metathesis with acrylic acid followed by hydrogenation gives ECA, a single unbranched chain ending in one carboxylic acid (a 5-carboxypentyl ether). M12 The same chemistry as for the esters gives ECB, so ECB carries the A5a side chain on an ethyl cellulose backbone. M13 Cross-metathesis with methyl acrylate or benzyl acrylate, thiol-Michael addition of 2-mercaptoethanol or 3-mercaptopropionic acid and saponification with tetrabutylammonium hydroxide give ECC (one acid and one alcohol in other positions) and ECD (two acids per side chain). M14·M15·P2 ECC could not be used: its DMSO solution precipitated on contact with buffer. R3

Table A. The six polymers (from Table 1, Figure 1 and Table S3). Table 1 prints A5b's second substituent as "A5a (0.67)", which the equivalent abbreviation and Figure 1 show to be a misprint. M10·M11·M12·M13·M14·M15·X1

The authors argue that molecular weight can be set aside: the syntheses are mild and modular, so similar molecular-weight distributions "would be expected", and at 5 µg/mL the polymers do not change the viscosity of the buffer. C6 Gel permeation chromatography failed for A5a, A5b and ECB because the chains aggregate; the two measured values, ECA 20.0 and ECD 60.5 kg/mol, differ three-fold. C6·M12·M15·X6

Contribution 2: induction times for nine chemically diverse drugs

Each polymer was dissolved first in a little DMSO (sonicated 90 min at 50 °C), added in aliquots to 20 mL of 100 mM sodium phosphate buffer at pH 6.8 and 50 °C, cooled and diluted to 5 µg/mL, with less than 1% DMSO in the final solution. P3·M18 Supersaturation was created by adding drug stock solution (methanol for seven drugs, DMSO for nevirapine, DMF for griseofulvin) to 47 mL of buffer at 37 °C stirred at 300 rpm; the drug concentrations ranged from 10 µg/mL (ezetimibe, danazol) to 500 µg/mL (nevirapine), and every condition was run three times. P4 The paper does not say how the 20 mL polymer solution relates to the 47 mL used for the test. X17 The drugs are un-ionized at pH 6.8, so the authors rule out ionic interactions with the anionic polymers. C7

Table B. Test conditions (Table S1) and cited solubilities (Table S2); the last column is the AI writer's calculation, not printed in the paper. Celecoxib was tested exactly at its LLPS concentration, although the text says the eight drugs other than telaprevir were below it. P4·M1–M9·X4·X11

Contribution 3: imaging how the polymers sit on an amorphous drug surface

Earlier work by the group related polymer coverage of a crystal surface to the rate of crystal growth (Schram 2015, ref 53). C37 Here the surface was amorphous, on the reasoning that an amorphous surface better represents the disordered phase "thought to be" the first stage of nucleation (ref 61). C37 Atazanavir was melted onto stainless steel pucks, pressed flat under a glass coverslip and cooled, then imaged in buffer containing 5 µg/mL polymer after 1 h at room temperature, in tapping mode, giving height and phase images of 2 µm × 2 µm. P5 The authors treat changes in topography and phase as an indication of how strongly the polymer interacts with the surface, which is "indirectly related" to drug–polymer–water interactions. C11

Contribution 4: molecular dynamics of a whole chain next to a drug cluster

Four-monomer oligomers were drawn with substituents placed to resemble the probable locations and average quantities of the substituents of each polymer (Tables 2 and 3), parameterized with the CHARMM force field and charges fitted to a Hartree–Fock electrostatic potential, and joined into 28-monomer chains. P6 Each chain carries 14 carboxylate groups (28 for ECD), all ionized, balanced by sodium ions. P6·P7 Two carboxylated substituents per four rings means a modelled DS of 0.50, below the 0.67–0.79 of the real polymers; acetyl and ethyl DS come out at 1.75 and 2.25 (AI writer's count from Tables 2 and 3). X9 The authors note that the real polymers are randomly substituted and "may slightly differ" from the model. C12

Single chains were simulated in about 187,400 SPC/E water molecules at 310 K and 1 bar: 5 ns at constant volume, 1 ns at constant pressure and a 5 ns production run. P7 Clusters of 25 telaprevir (TPV) or atazanavir (ATZ) molecules, optimized by density functional theory, were first formed in the gas phase and then placed 1 to 2 nm from an equilibrated chain; after equilibration, a 10 ns production run followed. P8·P9 The cluster stands for the drug aggregates seen in experiments; chain and cluster are expected to approach if their interaction is favourable and to drift apart if not. C20 TPV and ATZ were chosen because the group had used them most to test new cellulose polymers. C22 The polymer–drug runs cover A5a, A5b, ECA, ECB and ECD; ECC was simulated only as a lone chain. P9·R7 From the runs come the solvation free energy, $R_g$ and SASA of the lone chains, and $R_g$, SASA, RDFs and the free energy of interaction (Eq. 1) of the chain next to the drug. P7·P10·P11 One trajectory per polymer–drug pair is reported, without uncertainties for the free energy. X10

Experimental analysis

The tether: A5a beats A5b for every drug

Without polymer, the nine drugs crystallized after about 1 min (griseofulvin, danazol) to about 66 min (ritonavir) (Figure 4, values read from the plot). R1 With 5 µg/mL A5a every induction time grew: ritonavir to about 237 min, atazanavir from 40 to more than 300 min (the values printed in the table-of-contents graphic), telaprevir from about 15 to 91 min, nifedipine from about 6 to 58 min, nevirapine from about 8 to 45 min, and the four fast crystallizers to about 6–14 min. R1 A5b did much less: it still extended ritonavir (about 178 min) and atazanavir (65 min), but telaprevir (about 13 min) and nifedipine (about 6 min) stayed at control level, and the fast crystallizers at about 1–4 min. R1 The authors state that the short-tether polymer "performs better than the polymer with the long one for all the compounds studied". C8 Extending their earlier telaprevir result, they attribute A5b's weakness to its preference for interacting with itself, a reading they mark with "probably". C27

Note by the AI writer (not discussed in the paper): the two esters differ in more than tether length; A5b also has a lower carboxylated-substituent DS (0.67 vs 0.79) and more free hydroxyl groups (0.51 vs 0.39). X12 The difference in performance is large enough that a DS difference of 0.12 seems unlikely to explain it alone, but the paper does not separate the two.

The acids: one per side chain is better than two

Among the ethers (Figure 5), ECA, the single unbranched chain with one acid, lengthened the induction time of every drug: ritonavir about 154 min, atazanavir 199, telaprevir 86, nifedipine 166, nevirapine 109, griseofulvin 167, celecoxib 64, ezetimibe 72 and danazol 26. R2 For griseofulvin (about 1 min without polymer), celecoxib (about 4) and ezetimibe (about 3) this is more than a tenfold increase, which is why the authors write that ECA "appears to be an effective crystallization inhibitor, especially for the fast crystallizing compounds". R2·C9 ECB, which carries the A5a side chain on ethyl cellulose, was the best of the three for ritonavir (about 266 min) and atazanavir (about 237 min). R2

Table C. Nucleation induction times in minutes at 5 µg/mL polymer, pH 6.8, 37 °C, n = 3, read from Figures 4 and 5 (approximate; the no-polymer bars appear in both figures and read slightly differently for telaprevir and ezetimibe); atazanavir values without polymer, with A5a and with A5b are printed in the table-of-contents graphic. Error bars are not stated as standard deviations and are left out here. R1·R2

Comparing the two branched ethers, ECB (one acid) gave longer induction times than ECD (two acids) for ritonavir (about 266 vs 96 min), atazanavir (237 vs 125), telaprevir (94 vs 26), nifedipine (63 vs 40), nevirapine (39 vs 13) and celecoxib (28 vs 11); for griseofulvin (25 vs 29), ezetimibe (both about 23) and danazol (8 vs 6) the two were close. R2 The discussion summarizes this as ECB performing better "for most drugs; except for griseofulvin and ezetimibe, where both polymers appeared equally effective". C10 The results section says instead that the advantage holds for the slowly crystallizing drugs with "little difference" for the fast ones, which fits the plot less well, since telaprevir, nifedipine, nevirapine and celecoxib also show clear gaps. C10·X15 The authors conclude that more carboxylic acid groups per AGU "may not improve" inhibition, a result that "initially surprised" them given the earlier success of carboxylated and higher-DS(COOH) polymers. C30·C26

Note by the AI writer: ECD is not ECB plus one acid. In ECB the second branch is a hydroxyethyl ester; in ECD that ester has been saponified to an acid, so ECD loses a small hydrophobic ethylene unit as well as gaining a charge (Figure 1), and it was made by a different route. X12 The only measured molecular weights also differ three-fold between ECA (20.0 kg/mol) and ECD (60.5 kg/mol). X6

Surface adsorption: even spreading versus globules

On the amorphous atazanavir film, A5a left the surface looking like the polymer-free reference, with no globular structures; the authors describe it as appearing "homogeneously dispersed across the surface" (Figure 6). R4·C38 A5b formed large, unevenly distributed globules. R4 ECA and ECB formed small globules of similar size spread uniformly over the surface; ECD formed larger globules; in the phase images some large globules on ECB and ECD turned out to be agglomerates of smaller ones, and some small ECD globules hidden in the height topography became visible. R4 The authors rank surface coverage as extensive for ECA and ECB, lower for ECD and lowest for A5b. C39

Two points of interpretation follow in the discussion. First, A5b's globules agree with the MD finding that this polymer associates with itself. C38 Second, ECA and ECB form globules yet are among the best inhibitors of atazanavir crystallization, so "the formation of globular structures onto the surface does not necessarily indicate that a polymer is a poor crystallization inhibitor"; coverage and globule size must also be known. C39·C41 They suggest, with "It is possible", that globules may weaken the ability to prevent crystal growth, or that globule-forming polymers may simply need a higher concentration. C41 The globule formation of the ethers is called "unsurprising", given their less favourable solvation in MD. C40 No globule sizes or coverage values are reported, so the ranking rests on visual inspection of one image per polymer. X16

Solvation and chain shape of lone chains

The simulated free energies of solvation separate the two families sharply (Figure 7a): about −11.6 kcal/mol for A5a and −11.1 for A5b, against about −0.1 (ECA), −0.9 (ECB), +1.8 (ECC) and −1.3 kcal/mol (ECD). R5 The Fedors solubility parameters of the neutral structures show the same split, about 25.5 and 24.3 MPa^1/2 for the esters against 19.9–22.2 for the ethers (Figure 7b); a higher value means a more hydrophilic material. R6 The authors draw three readings from this: the esters are solvated more favourably than the ethers; A5a more favourably than A5b, "as expected" for the shorter tether; and A5a and ECB, which share a side chain but differ in linkage, backbone substituent and DS, differ so much that "small variations in the chemical structure and DS may influence the solvation process". C13·C15·C16 ECC, the least favourable of all, is the polymer that precipitated when dissolved for the experiments. C17·R3 They also point out that the MD chains are ionized and the Fedors structures neutral. C14

The lone-chain radius of gyration (Figure 8a) was about 3.58 nm for A5a, 3.53 for ECB, 3.47 for ECA, 3.33 for ECD, 3.29 for ECC and 3.16 nm for A5b; the text gives the range as 2.9 to 3.7 nm, somewhat wider than the plotted means and whiskers (about 3.06–3.67 nm). R7·X19 ECC showed the widest spread of $R_g$ during the run, which the authors read as a more flexible chain with more conformations. C19 SASA per atom (Figure 8b) was highest for ECD, A5a and ECA (about 0.080–0.081 nm²), lower for ECB and ECC (about 0.072–0.074) and lowest for A5b (about 0.066), meaning, in the authors' reading, extended side chains for the first three and agglomerated side chains for A5b. R8·C19

The chain next to the drug cluster

With a 25-molecule drug cluster nearby, A5b became even more compact: $R_g$ about 2.86 nm with TPV and 2.64 nm with ATZ, SASA about 0.061 and 0.060 nm² (Figures 8c, 8d). R7·R8 The authors take this as self-interaction winning over interaction with the drug. C21 ECB's $R_g$ also fell with ATZ (about 2.71 nm); since a smaller $R_g$ could mean either self-attraction or a chain bending around the drug, the authors turn to the RDF to decide, and attribute ECB's drop to bending around the cluster. R7·C33·C34 A5a's $R_g$ also fell next to ATZ (about 2.99 nm), which the paper does not discuss. R7·X18 The text cites "Figure 8b and 8c" for these runs, whereas the caption assigns the polymer–drug runs to panels c and d. X2

The RDFs (Figure 9) show which chains actually reach the cluster. A5a peaks at about 15 near 1.4 nm with TPV and about 23 near 1.0 nm with ATZ; ECB at about 17 (1.4 nm) and 31 (1.4 nm); ECA at about 8.5 (1.9 nm) with TPV and 33 (1.2 nm) with ATZ. R9 A5b has only a small peak of about 3 far out, near 6.4 nm with TPV and 7.2 nm with ATZ, and ECD with TPV is near zero below about 2 nm with a small peak of about 3.3 at 5.5 nm, while ECD with ATZ reaches about 18 at 1.8 nm. R9 Snapshots (Figure 10) show the drug clusters lying on the extended A5a chain and sitting apart from the coiled A5b chain. R10 The authors summarize that A5a is much more likely than A5b to be found near the drug, that A5a attracts ATZ more than TPV, and that ECA and ECB come closer than ECD, especially for TPV. C23

The free energies of interaction (Figure 12, printed values) follow the same pattern: A5a −84 (TPV) and −113 kcal/mol (ATZ), A5b 0 and 0, ECA −8 and −99, ECB −51 and −93, ECD 0 and −40 kcal/mol. R12 The authors call A5b's interaction "insignificant" and ECD's interaction with TPV "marginal". C25

Table D. Simulation results. Solvation free energy, Rg and RDF peaks read from Figures 7a, 8a, 8c and 9 (approximate); free energies of interaction printed in Figure 12. One 10 ns production run per polymer–drug pair. R5·R7·R9·R12·X10

Which part of the chain touches the drug

Figure 11 splits the RDF into the methyl groups of the acetyl or ethyl substituents (CH3) and the carboxylates (COO−). For A5a, ECB and ECD with ATZ, the CH3 groups lie closest: TPV–A5a about 33 at 0.5 nm against a COO− peak of about 10 at 2.6 nm; ATZ–A5a 35 at 0.4 nm against 16.5 at 1.6 nm; TPV–ECB 27 at 0.5 nm against 14 at 2.2 nm; ATZ–ECB 36 at 0.4 nm against 23 at 1.9 nm; ATZ–ECD 25 at 0.4 nm against 15 at 2.2 nm. R11 For TPV–ECD the CH3, COO− and whole-polymer curves overlap at about 3 near 5–5.5 nm, with no close contact at all, and in every other pair the COO− curve rises above the CH3 curve beyond about 2.7–3.6 nm. R11·X13 The authors conclude that the interaction of the most effective polymers, A5a and ECB, with the drugs "is mainly driven by the hydrophobic interactions between the acetyl/ethyl group and the drug aggregate", not by specific interactions of the carboxylic acid. C35 They see this as support for using computed RDFs and interaction energies to understand solution behaviour, as had been done with experimental RDFs and interaction energies for solid dispersions (refs 17, 59, 60). C36

Discussion

The authors' picture: contact first, extension by charge

The discussion assembles the results into one causal chain. Interaction with the emerging nucleus is "clearly a prerequisite" for disrupting nucleation (Anwar 2009, ref 62), so a polysaccharide that hardly interacts with the drug "clearly will be ineffective"; for polymers that do interact, the extent of disruption "likely" also depends on how far drug self-association is broken up, where steric factors such as the bulk of side-chain groups "may contribute". C44 Ionized carboxylates on tethered side chains keep the chain extended and limit intramolecular association, and the hydrophobic acetyl or ethyl groups of the backbone make the contacts with the drug. C50·C35 The flow below shows how the three structural cases separate.

On the strongest form of this argument the authors are careful. The ECB–ECD comparison is read as "suggesting that by adding a second carboxylic acid group (ECD) we are actually weakening the polymer-drug interaction". C34 The free energies "agree" with the induction times and "may give valuable information" about a polymer's tendency to interact with the drug and prevent crystallization, but because crystallization depends on kinetic as well as thermodynamic factors, "it is not realistic to expect a direct correlation" between induction time and free energy. C43·C45 The conclusion states that the approaches "lend support to the hypothesis" that effective polysaccharide inhibitors have a balance of hydrophobic and hydrophilic groups. C50

How well the free energy tracks the experiments

With atazanavir the ranking works well: A5a has the most negative free energy (−113 kcal/mol) and the longest induction time (more than 300 min), A5b has 0 and the shortest polymer-assisted time (65 min), and ECA, ECB and ECD fall in between in nearly the same order. R12·R1·R2 With telaprevir the picture is mixed. A5a (−84) and ECB (−51) both strongly extend the induction time (about 91 and 94 min), but ECA, at only −8 kcal/mol and with the weakest close-range RDF of the three, extends it almost as much (about 86 min). R12·R9·R1·R2·X7 A5b and ECD both score 0 with telaprevir, yet ECD still raised the telaprevir induction time from about 16 to 26 min while A5b did not. R12·R1·R2 With atazanavir, A5b scores 0 but still raised the induction time from 40 to 65 min. R12·R1·X7 The AI writer's reading: the free energy separates polymers that never reach the drug from those that do, but it does not rank close candidates reliably, which matches the authors' own caveat. X7·C45

Where the approach applies and what it means for formulation

The authors see the method as a route to "rational design of new materials for crystallization inhibition" and to approaches "to design highly effective polymers to improve oral drug delivery". C51 They also caution that "careful planning is advised" before using it as a fast screening tool in industry, and suggest simulated annealing MD for wider conformational sampling and semiempirical or tight-binding methods for cheaper static properties. C47 The cellulose ethers are presented as having potential practical advantages over the esters: hydrolytic stability under extreme pH and a possibly shorter synthesis. C28

Note by the AI writer (not stated in the paper): for a formulator choosing or designing a carboxylated cellulose inhibitor, the data point to three practical rules for these drugs and this concentration. Keep hydrophobic backbone groups such as acetyl or ethyl available to the drug; use carboxylic acids to keep the chain soluble and open, on a short tether; and do not assume that more acid per side chain is better. The best ether still depended on the drug, ECB for ritonavir and atazanavir and ECA for the fast crystallizers, so a design rule narrows the choice but does not replace testing with the drug in question. R1·R2

Limitations

Limitations stated by the authors

• The model chains place substituents at average positions; the real polymers are randomly substituted and "may slightly differ" in DS or substituent location. C12
• ECD was simulated with two charges per AGU, but "it is possible" that some of its acids are protonated under the experimental conditions. C32
• The MD gives thermodynamic information only, so no direct correlation with induction times is expected. C45
• Trajectory length and computing cost limit the method, and the absence of an experimental polysaccharide structure forces an extra step of simulating lone chains first. C46
• Polysaccharides are harder to simulate than proteins: no crystal structures, random substitution, a broad distribution of chain lengths, no binding pockets and greater flexibility. C42

The AI writer's own judgement

• Molecular weight was not controlled. The argument that molecular weight is irrelevant rests on an expectation; the two measured values differ three-fold (ECA 20.0, ECD 60.5 kg/mol) and three polymers could not be measured. C6·X6
• Each pair changes more than one feature. A5a and A5b also differ in DS; ECD differs from ECB by losing the hydroxyethyl ester as well as gaining an acid. X12
• The mechanism rests on two drugs and one image set. MD was run for telaprevir and atazanavir only, AFM for atazanavir only; the rules are then stated for polysaccharide inhibitors in general. X8
• Single, short simulations. One 10 ns run per pair, no replicate starts and no uncertainty on the free energies; the model chains carry fewer carboxylated groups (DS 0.50) than the real polymers and are fully ionized. X10·X9
• The CH3-contact pattern has exceptions. It holds at short range in five of six pairs, but not for TPV–ECD, and the carboxylates dominate beyond about 3 nm. X13
• AFM is qualitative. The coverage ranking has no numbers behind it, although the text says height images allow quantitation. X16
• Unequal supersaturation. The drugs were tested at about 4 to 58 times their crystalline solubility (AI calculation from Tables S1 and S2), telaprevir above its GLPS concentration and celecoxib exactly at its LLPS concentration, so induction times compare polymers within one drug but not drugs with one another. X11·X4
• Small source inconsistencies. Table 1 labels A5b's substituent "A5a"; the text cites the wrong panels of Figure 8; the captions of Figures 7 and 8 mention seven oligomers where six are shown and describe whiskers as quartiles although they extend beyond the standard-deviation boxes; the probe path length for nevirapine (0.2 cm) is not among those in the methods; references 13 and 34 are the same article. None of these changes the conclusions. X1·X2·X3·X5·X20

Conclusion

The paper makes three contributions. First, it shows with nine drugs that small side-chain changes in carboxylated cellulose derivatives change crystallization inhibition strongly: a long hydrocarbon tether (A5b) or a second acid per side chain (ECD) weakened inhibition compared with a short tether and one acid (A5a, ECA, ECB). C48·R1·R2 Second, it uses whole-chain MD in water, with a 25-molecule drug cluster, to connect inhibition to whether the chain reaches the drug: the effective polymers came close, contacted the drug mainly through acetyl or ethyl groups and had negative free energies of interaction, while A5b curled up and stayed away. R9·R11·R12·C35 Third, it adds AFM images of polymer adsorption on an amorphous drug surface, in which the effective polymers spread evenly. R4·C39 The authors' own wording of the outcome is that the approaches "lend support to the hypothesis" of a hydrophobic–hydrophilic balance. C50

The experimental rankings are direct and repeat across many drugs; the mechanism is supported by simulations and images for one or two drugs and by idealized model chains, and the polymer pairs are not perfectly matched. X8·X9·X12·X6 Testing the same rules with polymers matched in molecular weight and DS, with replicate and longer simulations, and with drugs beyond telaprevir and atazanavir would show how far they generalize.