Dr Jiayun Pang

Dihydrofolate reductase (DHFR) maintains the intracellular pool of tetrahydrofolate through catalysis of hydrogen transfer from reduced nicotinamide adenine dinucleotide to 7,8-dihydrofolate. We report results for pre-steady-state kinetic studies of the temperature dependence of the rates and the hydrogen/deuterium-kinetic isotope effects for the reactions catalysed by the enzymes from the mesophilic Escherichia coli and the hyperthermophilic Thermatoga maritima. We propose an evolutionary pattern in which catalysis progressed from a relatively rigid active site structure in the ancient thermophilic DHFR to a more flexible and kinetically more efficient structure in E. coli that actively promotes hydrogen transfer at physiological pH by modulating the tunnelling distance. The E. coli enzyme appeared relatively robust, in that kinetically severely compromised mutants still actively propagated the reaction. The reduced hydrogen transfer rates of the extensively studied Gly121Val mutant of DHFR from E. coli were most likely due to sterically unfavourable long-range effects from the introduction of the bulky isopropyl group.

The temperature dependence of the primary kinetic isotope effect (KIE), combined temperature-pressure studies of the primary KIE, and studies of the cc-secondary KIE previously led us to infer that hydride transfer from nicotinamide adenine dinucleotide to flavin mononucleotide in morphinone reductase proceeds via environmentally coupled hydride tunneling. We present here a computational analysis of this hydride transfer reaction using QM/MM molecular dynamics simulations and variational transition-state theory calculations. Our calculated primary and secondary KIEs are in good agreement with the corresponding experimental values. Although the experimentally observed KIE lies below the semiclassical limit, our calculations suggest that similar to 99% of the reaction proceeds via tunneling: this is the first "deep tunneling" reaction observed for hydride transfer. We also show that the dominant tunneling mechanism is controlled by the isotope at the primary rather than the secondary position: with protium in the primary position, large-curvature tunneling dominates, whereas with deuterium in this position, small-curvature tunneling dominates. Also, our study is consistent with tunneling being preceded by reorganization: in the reactant, the rings of the nicotinamide and isoalloxazine moieties are stacked roughly parallel to each other, and as the system moves toward a "tunneling-ready" configuration, the nicotinamide ring rotates to become almost perpendicular to the isoalloxazine ring.

The physical basis of the catalytic power of enzymes remains contentious despite sustained and intensive research efforts. Knowledge of enzyme catalysis is predominantly descriptive, gained from traditional protein crystallography and solution studies. Our goal is to understand catalysis by developing a complete and quantitative picture of catalytic processes, incorporating dynamic aspects and the role of quantum tunnelling. Embracing ideas that we have spearheaded from our work on quantum mechanical tunnelling effects linked to protein dynamics for H-transfer reactions, we review our recent progress in mapping macroscopic kinetic descriptors to an atomistic understanding of dynamics linked to biological H-tunnelling reactions.

We have identified multiple reactive configurations (MRCs) of an enzyme-coenzyme complex that have measurably different kinetic properties. In the complex formed between morphinone reductase (MR) and the NADH analogue 1,4,5,6-tetrahydro-NADH (NADH4) the nicotinamide moiety is restrained close to the FMN isoalloxazine ring by hydrogen bonds from Asn-189 and His-186 as determined from the X-ray crystal structure. Molecular dynamic simulations indicate that removal of one of these hydrogen bonds in the N189A MR mutant allows the nicotinamide moiety to occupy a region of configurational space not accessible in wild-type enzyme. Using stopped-flow spectroscopy, we show that reduction of the FMN cofactor by NADH in N189A MR is multiphasic, identifying at least four different reactive configurations of the MR-NADH complex. This contrasts with wild-type MR in which hydride transfer occurs by environmentally coupled tunneling in a single kinetic phase [Pudney et al. J. Am. Chem. Soc. 2006, 128, 14053-14058]. Values for primary and alpha-secondary kinetic isotope effects, and their temperature dependence, for three of the kinetic phases in the N189A MR are consistent with hydride transfer by tunneling. Our analysis enables derivation of mechanistic information concerning different reactive configurations of the same enzyme-coenzyme complex using ensemble stopped-flow methods. Implications for the interpretation from kinetic data of tunneling mechanisms in enzymes are discussed.

Molecular dynamics simulations of the temperature-induced unfolding reaction of native dimeric dihydrofolate reductase from the hyperthermophile Thermatoga maritima (TmDHFR) and the experimentally inaccessible TmDHFR monomer were carried out at 400 K, 450 K and 500 K. The results revealed that the unfolding of TmDHFR subunits followed a similar path to that of the monomeric DHFR from the mesophile E. coli (EcDHFR). An initial collapse of the adenosine-binding domain (ABD) was followed by the loss of the N-terminal and loop domains (NDLD). Interestingly, the elements of the secondary structure of the isolated TmDHFR monomer were maintained for significantly longer periods of time for the hyperthermophilic enzyme, suggesting that subunit stability contributes to the enhanced resistance of TmDHFR to temperature-induced unfolding. The interactions between the subunits of the TmDHFR dimer led to a stabilisation of the NDLD. The hydrogen bonds between residues 140–143 in βG of one subunit and residues 125–127 in βF of the other subunit were retained for significant parts of the simulations at all temperatures. These intermolecular hydrogen bonds were lost after the unfolding of the individual subunits. The high stability of the dimer mediated by strong intersubunit contacts together with an intrinsically enhanced stability of the subunits compared to EcDHFR provides a molecular rational for the higher stability of the thermophilic enzyme. The computed unfolding pathways suggest that the partly folded dimer may be a genuine folding intermediate.

The evolution and exploitation of dynamical processes — the coupling of specific protein motion to the reaction coordinate—remains a key question in the studies of enzyme mechanism.[1–4] It is now accepted that hydrogen (proton, hydrogen or hydride)-tunneling reactions often occur during enzyme-catalyzed reactions, [1, 5–10] but the role of non-equilibrium dynamical processes (or promoting motions) remains contentious.[3] This is despite compelling evidence from a number of kinetic and computational studies of such motions.[11–16] The experimental identification of promoting motions (or more generally environmental coupling) is challenging. Experimental evidence supporting a general role for protein motions in enzymatic H-tunnelling reactions has been gained primarily from observations of kinetic isotope effects (KIEs) with unusual temperature-dependences.[17–19] The most extreme cases of an unusual temperature-dependence of a primary (18) KIE are those with a ‘breakpoint’, where the KIE is weakly temperature-dependent at high temperatures and becomes strongly temperature-dependent below a critical breakpoint temperature. Such a trend has been demonstrated in two enzyme-catalyzed reactions: 1) the hydride transfer from benzyl alcohol to NAD+ catalyzed by the thermophilic alcohol dehydrogenase (ADH-ht) from Bacillus stearothermophilus, [18] and 2) the hydride transfer from NADPH to dihydrofolate catalyzed by the thermophilic dihydrofolate reductase (TmDHFR) from Thermotoga maritima. [19] Herein we explore the usefulness of a-secondary (28) KIEs (their magnitude and temperature dependence) as probes of environmentally-coupled H-transfer reactions in enzyme systems

The simulation of hydrogen transfer in enzymes provides a uniquely detailed atomistic insight into the reaction. Despite recent advances, such simulations remain challenging due to the importance of both electronic and nuclear quantum effects as well as the effects of protein motion. The incorporation of electronic quantum effects is required for the description of the breakage and formation of chemical bonds. Nuclear quantum effects, such as zero-point energy and hydrogen tunnelling, are also significant. However, over the past decade – with the enhancements in both computer power and theoretical approaches – computational simulations have become increasingly attractive in the study of hydrogen-transfer reactions in enzymes. They can complement experimental studies by providing insight that is difficult to access by experimental approaches, and make predictions that can be tested experimentally. The interplay between experiment and calculations can yield a remarkably complete description of an enzyme-catalysed hydrogen-transfer process – computational simulations have played a major role in furthering our understanding at the atomic level. Computational simulations vary greatly in the level of accuracy they provide. Broadly speaking, they can be divided into two categories molecular mechanical (MM) and quantum mechanical (QM) methods (see Chapter 2). This chapter discusses these methods, how they can be combined to allow larger systems to be studied quantum mechanically, and how the H-tunnelling event can be simulated and rates and kinetic isotope effects (KIEs) calculated. To illustrate these methods, the proton-tunnelling event catalysed by the enzyme aromatic amine dehydrogenase (AADH) with tryptamine as substrate is presented as a case study.

Proteinprotein interaction is a fundamental process in all major biological processes. The hexameric Tim9Tim10 (translocase of inner membrane) complex of the mitochondrial intermembrane space plays an essential chaperone-like role during import of mitochondrial membrane proteins. However, little is known about the functional mechanism of the complex because the interaction is weak and transient. This study investigates how electrostatic and hydrophobic interactions affect the conformation and function of the complex at physiological temperatures, using both experimental and computational methods. The results suggest that, first, different complex conformational states exist at equilibrium, and the major difference between these states is the degree of hydrophobic interactions. Second, the conformational change mimics the biological activity of the complex as measured by substrate binding at the same temperatures. Finally, molecular dynamics simulation and detailed energy decomposition analysis provided supporting evidence at the atomic level for the presence of an excited state of the complex, the formation of which is largely driven by the disruption of hydrophobic interactions. Taken together, this study indicates that the dynamics of the hydrophobic residues plays an important role in regulating the function of the Tim9Tim10 complex. Proteins 2012. (C) 2011 Wiley Periodicals, Inc.

The purpose of the study was to investigate and identify the interactions within solid dispersions of cationic drugs and anionic polymers processed by hot-melt extrusion (HME) technique. Propranolol HCl (PRP) and diphenhydramine HCl (DPD) were used as model cationic active substances while pH sensitive anionic methacrylic acid based methyl methacrylate copolymers Eudragit L100((R)) (L100) and ethyl acrylate copolymer Eudragit L100-55 (Acryl EZE) (L100-55) were used as polymeric carriers. The extrudates were further characterised using various physicochemical characterisation techniques to determine the morphology, the drug state within the polymer matrices and the type of drug-polymer interactions. Molecular modelling predicted the existence of two possible H-bonding types while the X-ray photon spectroscopy (XPS) advanced surface analysis of the extrudates revealed intermolecular ionic interactions between the API amino functional groups and the polymer carboxylic groups through the formation of hydrogen bonding. The magnitude of the intermolecular interactions varied according to the drug-polymer miscibility. (C) 2012 Elsevier B.V. All rights reserved.

A computational study was performed on the experimentally elusive cyclisation step in the cofactor pyridoxal 5'-phosphate (PLP)-dependent D-ornithine 4,5-aminomutase (OAM)-catalysed reaction. Calculations using both model systems and a combined quantum mechanics/molecular mechanics approach suggest that regulation of the cyclic radical intermediate is achieved through the synergy of the intrinsic catalytic power of cofactor PLP and the active site of the enzyme. The captodative effect of PLP is balanced by an enzyme active site that controls the deprotonation of both the pyridine nitrogen atom (N1) and the Schiff-base nitrogen atom (N2). Furthermore, electrostatic interactions between the terminal carboxylate and amino groups of the substrate and Arg297 and Glu81 impose substantial "strain" energy on the orientation of the cyclic intermediate to control its trajectory. In addition the "strain" energy, which appears to be sensitive to both the number of carbon atoms in the substrate/analogue and the position of the radical intermediates, may play a key role in controlling the transition of the enzyme from the closed to the open state. Our results provide new insights into several aspects of the radical mechanism in aminomutase catalysis and broaden our understanding of cofactor PLP-dependent reactions.

We show that hydride transfer from NADH to FMN catalysed by the N189A mutant of morphinone reductase occurs along parallel "chemical" pathways in a conformationally rich free-energy landscape. We have developed experimental kinetic and spectroscopic tools by using hydrostatic pressure to explore this free-energy landscape. The crystal structure of the N189A mutant enzyme in complex with the unreactive coenzyme analogue NADH(4) indicates that the nicotinamide moiety of the analogue is conformationally less restrained than the corresponding structure of the wild-type NADH(4) complex. This increased degree of conformational freedom in the N189A enzyme gives rise to the concept of multiple reactive configurations (MRCs), and we show that the relative population of these states across the free-energy landscape can be perturbed experimentally as a function of pressure. Specifically, the amplitudes of individual kinetic phases that were observed in stopped-flow studies of the hydride transfer reaction are sensitive to pressure; this indicates that pressure drives an altered distribution across the energy landscape. We show by absorbance spectroscopy that the loss of charge-transfer character of the enzyme-coenzyme complex is attributed to the altered population of MRCs on the landscape. The existence of a conformationally rich landscape in the N189A mutant is supported by molecular dynamics simulations at low and high pressure. The work provides firm experimental and computational support for the existence of parallel pathways arising from multiple conformational states of the enzyme-coenzyme complex. Hydrostatic pressure is a powerful and general probe of multidimensional energy landscapes that can be used to analyse experimentally parallel pathways for enzyme-catalysed reactions. We suggest that this is especially the case following directed mutation of a protein, which can lead to increased population of reactant states that are essentially inaccessible in the free-energy landscape of wild-type enzyme.

In recent years there has been a shift away from transition state theory models for H-transfer reactions. Models that incorporate tunneling as the mechanism of H-transfer are now recognized as a better description of such reactions. Central to many models of H-tunneling is the notion that specific vibrational modes of the protein and/or substrate can increase the probability of a H-tunneling reaction, modes that are termed promoting vibrations. Thus far there has been limited evidence that promoting vibrations can increase the rate of H-transfer. In the present communication we examine the single hydride transfer from both NADPH and NADH to FMN in the reductive half-reaction of pentaerythritol tetranitrate reductase (PETNR). We find that there is a significant promoting vibration with NADPH but not with NADH and that the observed rate of hydride transfer is significantly (approximately 15x) faster with NADPH. We rule out differences in rate due to variation in driving force and the donor-acceptor distance, suggesting it is the promoting vibration with NADPH that is the origin of the increased observed rate. This study therefore provides direct evidence that promoting vibrations can lead to an increase in rate.

Fourier transform infrared (FTIR) spectroscopy can be used to provide a detailed time-resolved probe of reaction intermediates in enzyme-catalyzed systems. Accurate assignment of the respective chemical species being studied is key to the success of this approach. The plethora of signals from the protein environment, leading to complexity in the spectra, presents a particular challenge. Here we present a combined QM/MM-based approach that can be used to assign key resonances in the FTIR spectrum of tryptophan tryptophyl quinone (TTQ) in the TTQ-dependent quinoprotein aromatic amine dehydrogenase (AADH). We show that consideration of the cofactor alone is not sufficient to identify correctly the experimentally observed resonances-inclusion of the protein is required for this. However, to enable accurate peak assignment, a stepwise approach is needed that builds up increasing levels of complexity from a simple system. This study serves as a benchmark for future QM/MM-based Studies to predict the spectroscopic changes during the interconversion of intermediates in the reductive half-reaction catalyzed by AADH, and more generally for using a combined QM/MM approach to calculate spectroscopic data of protein cofactors and cofactor-based adducts.

Computational insight into the multi-step reaction cycle of aromatic amine dehydrogenase is presented, identifying the energy landscape and pathway for multiple proton transfers. This atomistic picture of the reaction sequence-including short-lived reaction intermediates and a stepwise reaction mechanism-bridges the gap between a small number of crystallographic snapshots.

Hansen solubility parameters (HSPs) have three components, delta d, delta p and delta h, accounting for dispersion forces, polar forces, and hydrogen bonding of a molecule, which were designed to better understand how molecular structure affects miscibility/solubility. HSP is widely used throughout the pipeline of pharmaceutical research and yet has not been as well studied computationally as the aqueous solubility. In the current study, we predicted HSPs using only the SMILES of molecules and utilise the molecular embedding approach inspired by Natural Language Processing (NLP). Two pre-trained deep learning models - Mol2Vec and ChemBERTa have been used to derive the embeddings. A dataset of similar to 1200 organic molecules with experimentally determined HSPs was used as the labelled dataset. Upon finetuning, the ChemBERTa model "learned" relevant molecular features and shifted attention to functional groups that give rise to the relevant HSPs. The finetuned ChemBERTa model outperforms both the Mol2Vec model and the baseline Morgan fingerprint method albeit not to a significant extent. Interestingly, the embedding models can predict delta d significantly better than delta h and delta p and overall, the accuracy of predicted HSPs is lower than the well-benchmarked ESOL aqueous solubility. Our study indicates that the extent of transfer learning leveraged from the pre-trained models is related to the labelled molecular properties. It also highlights how delta p and delta h may have large intrinsic errors in the way they are defined and therefore introduces inherent limitations to their accurate prediction using machine learning models. Our work reveals several interesting findings that will help explore the potential of BERT-based models for molecular property prediction. It may also guide the possible refinement of the Hansen solubility framework, which will generate a wide impact across the pharmaceutical industry and research.

Human vascular adhesion protein-1 (VAP-1) is an endothelial copper-dependent amine oxidase involved in the recruitment and extravasation of leukocytes at sites of inflammation. VAP-1 is an important therapeutic target for several pathological conditions. We expressed soluble VAP-1 in HEK293 EBNA1 cells at levels suitable for detailed mechanistic studies with model substrates. Using the model substrate benzylamine, we analyzed the steady-state kinetic parameters of VAP-1 as a function of solution pH. We found two macroscopic pKa values that defined a bell-shaped plot of turnover number kcat,app as a function of pH, representing ionizable groups in the enzyme-substrate complex. The dependence of (kcat/Km)app on pH revealed a single pKa value (∼9) that we assigned to ionization of the amine group in free benzylamine substrate. A kinetic isotope effect (KIE) of 6 to 7.6 on (kcat/Km)app over the pH range of 6 to 10 was observed with d2-benzylamine. Over the same pH range, the KIE on kcat was found to be close to unity. The unusual KIE values on (kcat/Km)app were rationalized using a mechanistic scheme that includes the possibility of multiple isotopically sensitive steps. We also report the analysis of quantitative structure-activity relationships (QSAR) using para-substituted protiated and deuterated phenylethylamines. With phenylethylamines we observed a large KIE on kcat,app (8.01 ± 0.28 with phenylethylamine), indicating that C–H bond breakage is limiting for 2,4,5-trihydroxyphenylalanine quinone reduction. Poor correlations were observed between steady-state rate constants and QSAR parameters. We show the importance of combining KIE, QSAR, and structural studies to gain insight into the complexity of the VAP-1 steady-state mechanism.

We present here an energetic and atomistic description of how D-ornithine 4,5-aminomutase (OAM), an adenosylcobalamin (AdoCbl; coenzyme B-12)-dependent isomerase, employs a large-scale protein domain conformational change to orchestrate the homolytic rupture of the Co-C bond. Our results suggest that in going from the open form (catalytically inactive) to the closed form (catalytically active), the Rossmann domain of OAM effectively approaches the active site as a rigid body. It undergoes a combination of a similar to 52 degrees rotation and a similar to 44 angstrom translation to bring AdoCbl-initially positioned similar to 25 angstrom away-into the active-site cavity. This process is coupled to repositioning of the Ado moiety of AdoCbl from the eastern conformation to the northern conformation. Combined quantum mechanics and molecular mechanics calculations further indicate that in the open form, the protein environment does not impact significantly on the Co-C bond homolytic rupture, rendering it unusually stable, and thus catalytically inactive. Upon formation of the closed form, the Co-C bond is activated through the synergy of steric and electrostatic effects arising from tighter interactions with the surrounding enzyme. The more pronounced effect of the protein in the closed form gives rise to an elongated Co-C bond (by 0.03 angstrom), puckering of the ribose and increased "strain" energy on the Ado group and to a lesser extent the corrin ring. Our computational studies reveal novel strategies employed by AdoCbl-dependent enzymes in the control of radical catalysis.

In this study molecular modeling is introduced as a novel approach for the development of pharmaceutical solid dispersions. A computational model based on quantum mechanical (QM) calculations was used to predict the miscibility of various drugs in various polymers by predicting the binding strength between the drug and dimeric form of the polymer. The drug/polymer miscibility was also estimated by using traditional approaches such as Van Krevelen/Hoftyzer and Bagley solubility parameters or FloryHuggins interaction parameter in comparison to the molecular modeling approach. The molecular modeling studies predicted successfully the drugpolymer binding energies and the preferable site of interaction between the functional groups. The drugpolymer miscibility and the physical state of bulk materials, physical mixtures, and solid dispersions were determined by thermal analysis (DSC/MTDSC) and X-ray diffraction. The produced solid dispersions were analyzed by X-ray photoelectron spectroscopy (XPS), which confirmed not only the exact type of the intermolecular interactions between the drugpolymer functional groups but also the binding strength by estimating the N coefficient values. The findings demonstrate that QM-based molecular modeling is a powerful tool to predict the strength and type of intermolecular interactions in a range of drug/polymeric systems for the development of solid dispersions.

Erv1 is an FAD-dependent thiol oxidase of the ERV (essential for respiration and viability)/ALR (augmenter of liver regeneration) sub-family and an essential component of the mitochondrial import and assembly pathway. Erv1 contains six tryptophan residues, which are all located in the highly conserved C-terminal FAD-binding domain. Though important structural roles were predicted for the invariable Trp(95), no experimental study has been reported. In the present study, we investigated the structural and functional roles of individual tryptophan residues of Erv1. Six single tryptophan-to-phenylalanine yeast mutant strains were generated and their effects on cell viability were tested at various temperatures. Then, the mutants were purified from Escherichia coli. Their effects on folding, FAD-binding and Erv1 activity were characterized. Our results showed that Erv1(W95F) has the strongest effect on the stability and function of Erv1 and followed by Erv1(W183F). Erv1(W95F) results in a decrease in the T-m of Erv1 by 23 degrees C, a significant loss of the oxidase activity and thus causing cell growth defects at both 30 degrees C and 37 degrees C. Erv1(W183F) induces changes in the oligomerization state of Erv1, along with a pronounced effect on the stability of Erv1 and its function at 37 degrees C, whereas the other mutants had no clear effect on the function of Erv1 including the highly conserved Trp(157) mutant. Finally, computational analysis indicates that Trp(95) plays a key role in stabilizing the isoalloxazine ring to interact with Cys(133). Taken together, the present study provided important insights into the molecular mechanism of how thiol oxidases use FAD in catalysing disulfide bond formation.

Methanol intolerance of lipase is a major limitation in lipase-catalysed methanolysis reactions. In this study, to understand the molecular mechanism of methanol-induced inactivation of lipases, we performed molecular dynamics (MD) simulations of Thermomyces lanuginosus lipase (TLL) in water and methanol and compared the observed structural and dynamic properties. The solvent accessibility analysis showed that in methanol, polar residues tended to be buried away from the solvent while non-polar residues tended to be more solvent-exposed in comparison to those in water. Moreover, we observed that in methanol, the van der Waals packing of the core residues in two hydrophobic regions of TLL became weak. Additionally, the catalytically relevant hydrogen bond between Asp201 OD2 and His258 ND1 in the active site was broken when enzyme was solvated in methanol. This may affect the stability of the tetrahedral intermediates in the catalytic cycle of TLL. Furthermore, compared to in water, some enzyme surface residues displayed enhanced movement in methanol with higher Cα root-mean-square atomic positional fluctuation values. One of such methanol-affecting surface residues (Ile241) was chosen for mutation, and MD simulation of the I241E mutant in methanol was conducted. The structural analysis of the mutant showed that replacing a non-polar surface residue with an acidic one at position 241 contributed to the stabilisation of enzyme structure in methanol. Ultimately, these results, while providing molecular-level insights into the destabilising effect of methanol on TLL, highlight the importance of surface residue redesign to improve the stability of lipases in methanol environments.

Cooperativity is a feature many multimeric proteins use to control activity. Here we show that the bacterial heptose isomerase GmhA displays homotropic positive and negative cooperativity among its four protomers. Most similar proteins achieve this through conformational changes: GmhA instead employs a delicate network of hydrogen bonds, and couples pairs of active sites controlled by a unique water channel. This network apparently raises the Lewis acidity of the catalytic zinc, thus increasing the activity at one active site at the cost of preventing substrate from adopting a reactive conformation at the paired negatively cooperative site - a "half-site" behavior. Our study establishes the principle that multimeric enzymes can exploit this cooperativity without conformational changes to maximize their catalytic power and control. More broadly, this subtlety by which enzymes regulate functions could be used to explore new inhibitor design strategies.

Hexahydromethanocarbazole is a privileged scaffold in the discovery of new drugs and photoactive organic materials due to its good balance between structural complexity and minimized entropy penalty upon receptor binding. To address the difficulty of synthesizing this highly desirable bridged polycyclic scaffold, we designed a convenient multicomponent reaction cascade as intercepted Heck addition/C-H activation/C-palladacycle formation/electrophilic attack of ANP/N-palladacycle formation/Buchwald amination. A distinguishing feature of this sophisticated strategy is the successive generation of two key phenylnorbornyl palladium species to control the reaction flow towards desired products. DFT calculations further reveal the crucial roles of Cs CO and 5,6-diester substitutions on the norbornene reactant in preventing multiple side-reactions. This innovative method exhibits a broad scope with good yields, and therefore will enable the construction of natural-product-like compound libraries based on hexahydromethanocarbazole.

Transformer-based encoder-decoder models have demonstrated impressive results in chemical reaction prediction tasks. However, these models typically rely on pretraining using tens of millions of unlabelled molecules, which can be time-consuming and GPU-intensive. One of the central questions we aim to answer in this work is: can FlanT5 and ByT5, the encoder-decoder models pretrained solely on language data, be effectively specialised for organic reaction prediction through task-specific fine-tuning? We conduct a systematic empirical study on several key issues of the process, including tokenisation, the impact of (SMILES-oriented) pretraining, fine-tuning sample efficiency, and decoding algorithms at inference. Our key findings indicate that although being pretrained only on language tasks, FlanT5 and ByT5 provide a solid foundation to fine-tune for reaction prediction, and thus become 'chemistry domain compatible' in the process. This suggests that GPU-intensive and expensive pretraining on a large dataset of unlabelled molecules may be useful yet not essential, to leverage the power of language models for chemistry. All our models achieve comparable Top-1 and Top-5 accuracy although some variation across different models does exist. Notably, tokenisation and vocabulary trimming slightly affect final performance but can speed up training and inference; the most efficient greedy decoding strategy is very competitive while only marginal gains can be achieved from more sophisticated decoding algorithms. In summary, we evaluate FlanT5 and ByT5 across several dimensions and benchmark their impact on organic reaction prediction, which may guide more effective use of these state-of-the-art language models for chemistry-related tasks in the future. We evaluate FlanT5 and ByT5 across tokenisation, pretraining, finetuning and inference and benchmark their impact on organic reaction prediction tasks.