Digital

Discovery
View Article Online
PAPER View Journal

Leveraging large language models for enzymatic

This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.

reaction prediction and characterization

Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

Cite this: DOI: 10.1039/d5dd00187k

a *b
Lorenzo Di Fruscia and Jana M. Weber

Predicting enzymatic reactions is crucial for applications in biocatalysis, metabolic engineering, and drug
discovery, yet it remains a complex and resource-intensive task. Large Language Models (LLMs) have
recently demonstrated remarkable success in various scientific domains, e.g., through their ability to
generalize knowledge, reason over complex structures, and leverage in-context learning strategies. In this
study, we systematically evaluate the capability of LLMs, particularly the Llama-3.1 family (8B and 70B),
across three core biochemical tasks: enzyme commission number prediction, forward synthesis, and
retrosynthesis. We compare single-task and multitask learning strategies, employing parameter-efficient
fine-tuning via LoRA adapters. Additionally, we assess performance across different data regimes to explore
Received 8th May 2025
Accepted 21st September 2025
their adaptability in low-data settings. Our results demonstrate that fine-tuned LLMs capture biochemical
knowledge, with multitask learning enhancing forward- and retrosynthesis predictions by leveraging shared
DOI: 10.1039/d5dd00187k
enzymatic information. We also identify key limitations, for example challenges in hierarchical EC
[Link]/digitaldiscovery classification schemes, highlighting areas for further improvement in LLM-driven biochemical modeling.

these approaches learn reaction patterns directly from molec-

1 Introduction ular strings or graphs with limited to no reliance on craed
Biochemistry plays a fundamental role in nearly every aspect of rules. Early work applied motif- and prole-based statistical
daily life, from medicine development to food production, from learners such as PRIAM9 for enzyme detection and classica-
the creation of fuels to personal care items, signicantly tion, and EFICAz10 for enzyme function inference. These
contributing to improved quality of life. Developing novel bi- statistical methods have since been surpassed by deep archi-
ocatalysts and discovering and optimizing biochemical reac- tectures such as the CNN-based DEEPre11 and DeepEC12 for
tions hold immense promise for addressing global challenges. enzyme function prediction, and RNN-based ones such as the
However, these discoveries are inherently complex, requiring sequence-to-sequence model from ref. 13 for retrosynthetic
a deep understanding of enzyme-substrate relationships, and reaction prediction.
they remain experimentally expensive and time-intensive.1–3 More recently transformers,14 architectures suited for appli-
To mitigate these experimental costs, computational cations like language translation, sentiment analysis and text
approaches date back many decades: early Computer-Aided completion, have proven to be effective for tasks such as
Synthesis Planning (CASP) began with hand-craed expert chemical reaction product prediction with the molecular
systems such as LHASA4 and SECS,5 that encoded chemistry transformer15 and for molecule optimization overall.16,17 In
libraries of reaction transformation rules to propose synthetic biochemistry, several ML models have been tailored for
routes. These rules were applied to produce templates, specic prediction tasks, including approaches where enzymes are
atom-mapping patterns describing the molecualr trans- represented using natural language (e.g. the enzymatic trans-
formations. These approaches evolved into more recent and former18), numerical classication schemes,19,20 or amino acid
extensive methods such as SYNTHIA,6 ICSYNTH7 and RetroSim8 sequences.21 Recent work has also shown that transformer
while still being primarily template-based. However, templates models trained on protein sequences only (also known as
require manual rule curation and heuristic atom-mapping, Protein Language Models or PLMs) can be used on downstream
which limits their scalability as reaction databases grew. To tasks such as predicting EC numbers from the amino-acid
overcome these limitations, the scientic community started sequence,22,23 usually obtained by adapting large protein
developing template-free methods. Data-driven at their core, encoder-only models, e.g. ProteinBERT24 and ESM-2.25 While
these specialized models deliver impressive results, they are
typically constrained to specic tasks and require extensive
a
Department of Intelligent Systems Del University of Technology, Del 2629 HZ, The domain-specic data and expertise for their development and
Netherlands. E-mail: [Link]@tudel.nl for the incorporation of biochemical knowledge.
b
Department of Intelligent Systems Del University of Technology, Del 2629 HZ, The
Netherlands. E-mail: [Link]@tudel.nl

© 2025 The Author(s). Published by the Royal Society of Chemistry Digital Discovery
 View Article Online

Digital Discovery Paper

We are now witnessing the emergence of foundation models zero- and few-shot capabilities for in-context learning.44 One key
like Large Language Models (LLMs),26,27 that have found their problem of learning from context is the high variance in the
application in chemistry as well.28 These transformer-based outputs returned by the model: slight changes in prompts can
architectures consist of up to hundreds of billions of parameters greatly affect the model performance, ranging from barely above
and are trained on text corpora comprising trillions of tokens. chance, to near state-of-the-art (SOTA) level.45 Additionally, LLMs
Despite being trained for next token prediction, these models have may produce made-up or irrelevant content, a phenomenon
shown emergent abilities that were not foreseeable for smaller known as hallucinations. To address these instabilities, research
sized models:29 they are capable of more than just completing has explored advanced prompting strategies such as Chain-of-
phrases in natural language, as to some extent they are able to Thought (CoT), a technique that guides the model to break
This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.

answer questions, understand examples and reason over prob- down answers as a series of connected thoughts. By explicitly
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

lems. Foundation models can be capable of solving multiple tasks decomposing complex problems into step-by-step reasoning, CoT
at once. Building on top of existing LLMs is straightforward to reduces output variability and enhances accuracy, particularly for
implement and they require relatively little expertise to use, cir- tasks requiring logical progression or multi-step calculations. By
cumventing the need to train a multitude of specialized models. acting in a way that mimics human reasoning, CoT showed to
LLMs excel in low-data regimes, adapting on the y from context improve the reliability of responses and to therewith make LLMs
and examples such as in real-world lab scenarios. They can ground more robust.46
their outputs via Retrieval Augmented Generation (RAG),30 or Another key task adaptation strategy is ne-tuning, which
knowledge-graph databases access,31 support agentic behaviour modies the weights of the pretrained model. It offers the
through external tool integration,32 and can expose their step-by- advantage of not being constrained by a limited context window
step reasoning to guide experimental workows.33,34 for input data, but it typically results in a model specialized for
Efforts to integrate LLMs into chemistry generally fall into two a single task. However, prior research47 showed that ne-tuning
distinct categories. The rst focuses on building chemistry agents outperforms in-context learning strategies in both in-domain
that leverage the LLMs planning abilities to work with task- and out-of-distribution tasks for models of comparable size,
specic tools and improve reasoning.35 For instance, in Bran with performance gains increasing as more training data
et al.,36 researchers augmented LLMs by providing access to becomes available. Fine-tuning limitations in principle include
expert-designed tools for drug discovery, materials design and the need for signicant training expertise and computational
organic synthesis. The second category involves using LLMs resources, with a reduced reusability compared to in-context
directly for downstream tasks such as property prediction, learning strategies. These shortcomings are partially mitigated
reagent selection and molecule captioning.37–40 In Guo et al.,37 by Parameter-Efficient Fine-Tuning (PEFT).48,49 PEFT techniques
they benchmarked LLMs in zero- and few-shot settings, demon- selectively adjust only a small portion of parameters, leaving the
strating their capabilities in explaining, understanding and rest unchanged. This approach preserves the base model general-
reasoning over chemistry. In Jablonka et al.,38 they show how by purpose capabilities while adding task-specic expertise in
ne-tuning GPT family models from OpenAI,41 they easily adapt a modular way, enabling greater adaptability to new tasks.
them to solve various tasks involving classication, regression, We focus on enzymatic reactions represented using SMILES
inverse design of chemicals, and many more. Their model proved (Simplied Molecular Input Line Entry System) notation50 for
to be useful especially in the low-data regime, where the LLM chemicals and EC numbers for enzyme classication. Speci-
performed at least as good as the conventional ML models. cally, we design tasks that test the model's ability to predict EC
Additionally, comprehensive instruction datasets for the chem- numbers, reaction products (forward synthesis), and substrates
ical and biochemical domains have been introduced.42,43 These (retrosynthesis). By introducing a multitask learning setup, we
datasets, encompassing millions of examples across applications investigate whether training on multiple tasks simultaneously
like molecule generation, name conversion and reaction predic- makes use of shared biochemical knowledge compared to
tion, enable small ne-tuned LLMs to surpass prompted SOTA single-task ne-tuning. Finally, we perform ablation studies to
LLMs, demonstrating the role of high quality datasets in examine the impact of several data regimes and ne-tuning
enhancing performance in molecular domains. setups on different models' performance.
While previous studies mainly focused on the investigation
of LLMs for chemical and materials tasks, we are interested in 2 Methods
understanding LLMs potential for biochemical reaction char-
acterization, discovery, and optimization. Specically, we are 2.1 Tasks and dataset description
interested in whether a single general-purpose LLM can be 2.1.1 Task selection. We assemble a representative set of
adapted on multiple tasks with one interface, with little engi- biochemical prediction tasks. The selected tasks are designed to
neering and limited labeled data. With such a model, scientists evaluate the capabilities of Large Language Models (LLMs) in
would be able to query multiple aspects possibly under understanding and predicting enzymatic reactions, when the
changing conditions in natural language. This study investi- chemicals are presented in string format and the enzyme in the
gates to which extent one LLM is sufficient, and in which EC numerical classication scheme. Specically:
aspects the connection to specialized models is still required.
EC number prediction: we assess whether LLMs can
The scientic community is building upon recent discoveries accurately assign EC numbers given the substrates and the
that scaling up LLMs in size and training data leads to promising products of each reaction.

Digital Discovery © 2025 The Author(s). Published by the Royal Society of Chemistry
 View Article Online

Paper Digital Discovery

Substrate prediction: we test the model's capabilities of

predicting substrates based on reaction products and the EC
number (retrosynthesis).
Given the inherent similarities among the three tasks, we
investigate whether the model can improve its performance when
trained on all tasks simultaneously, by leveraging shared infor-
mation in a synergistic manner. To test this, we introduce
a multitask (MT) setup, in which a single model is trained
concurrently on all three tasks, inspired by what has been done in
This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.

Yu et al.43 This setup allows us to evaluate whether a multitask-

Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

trained model can outperform individually ne-tuned models

for each task (single-task, ST) producing a more general model
eventually capable of handling diverse biochemistry tasks
involving enzymes. The following sections explain data selection
and the data split suitable for both ST and MT experiments. To
ensure that the selected tasks are supported by high-quality data,
we preprocess the data to minimize biases and data leakage.
2.1.2 Dataset preparation. We make use of the ECREACT
dataset curated by Probst et al.19 This dataset results from the
combination of data coming from four different databases:
Fig. 1 Distributions of samples across EC levels for the BRENDA MetaNetX, Rhea, PathBank and BRENDA.51–54 The authors scre-
dataset. The innermost layer represents the main class (EC1 digit), and ened the enzymatic reactions, and determined the corresponding
the middle and outer layers represent levels EC2 and EC3 respectively. Enzyme Commission (EC) number for each of them. Further
The label for enzyme class 7 (translocases) is not visible due to the processing simplied and generalized the dataset. They removed
limited data available (<20 samples).
products also occurring as reactants in the same reaction, co-
enzymes, common by-products, and reactions without reactants
or multiple or missing products. In each reaction, substrates and

Product prediction: here we explore the model's ability to products are represented in SMILES, whereas EC numbers are
predict reaction products given substrates and the EC number tags for the reaction in the form of a 4-digit tag ‘X.X.X.X’. The
associated to the reaction (forward synthesis). digits follow a hierarchy, with the rst digit (EC1) representing

Fig. 2Data preprocessing pipeline. Individual reactions sharing the same {product, EC} or {substrate, EC} pair are grouped together (here groups
are numbered from 1 to 9, first row). The dataset is split into training and test set, while keeping each group intact. Within training and test, each
group is assigned to one of the three tasks on a rotating basis to balance the splits. Groups are randomly shuffled at the beginning of the
procedure, here we keep indices in order for visual clarity.

© 2025 The Author(s). Published by the Royal Society of Chemistry Digital Discovery
 View Article Online

Digital Discovery Paper

the main class of the enzymatic reaction. From these databases, {substrate, EC} pair, but differ in the remaining molecule, and
we only focus on BRENDA, for a total of n = 8496 enzyme- avoidance of task-specic leakage, ensuring that if e.g. a reac-
catalyzed reactions covering all seven different EC classes. This tion appears in forward synthesis, it must not appear in retro-
is mainly due to computational constraints, as ne-tuning large- synthesis as well. More details about these steps are reported in
scale LLMs on the full dataset without parallelized infrastructure Appendix Section A.1.
would require several weeks. The distribution of reactions These points imply that to maintain dataset integrity, each
according to their respective EC numbers is shown in Fig. 1. We above-mentioned reaction group is assigned exclusively to one
include all four EC digits (thus up to EC4) in the dataset, but our task and one dataset split (either training or test). By addressing
subsequent analyses will focus on up to sublevel EC3, as many these issues preemptively, we also ensure a consistent random
This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.

subcategories for EC4 consist of only a single enzyme-substrate dataset split for both single-task and multitask setups, enabling
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

example. Class 7 will not be included as well due to the limited fair comparisons between the two methodologies. Fig. 2 better
sample size for the class (<20 samples). illustrates this approach.
2.1.3 Data splitting. We implement several preprocessing We perform a 70–30 train–test split, ensuring that the frac-
steps: canonicalization of SMILES representations using the tion of groups assigned to each of the two sets maintains
RDKit library parsing functions to remove redundant entries, a balanced ratio. Of the train set, 10% is used for validation. The
grouping reactions that share the same {product, EC} or test sets remain constant across all training regimes and varying

Fig. 3 Distribution of reaction groups with repeating substrates and/or products. Unique reactions are included as elements with group size
equal to 1. Group sizes with a number of counts >10 closely follow the required 70–30 split ratio between train and test set.

Digital Discovery © 2025 The Author(s). Published by the Royal Society of Chemistry
 View Article Online

Paper Digital Discovery

training set sizes. EC prediction, forward synthesis and retro- reasoning, story-writing, conversation, tool-access and so on.
synthesis tasks have xed test sets of size 855, 857, and 837 With ICL, the model is not retrained, so no parameters are
examples respectively. The above-mentioned preprocessing changed. Instead, the model uses its existing knowledge to
steps prevent basic information leakage that could articially generalize “on the y” within that single interaction, which we
inate performance metrics, yet our similarity analysis shows refer to as zero-shot. LLMs are powerful zero-shot learners and
that substantial analogue overlap persists, that we report in can easily adapt to examples to improve their understanding,
Appendix Fig. 12. It is also worth noting that by being public it is which is called few-shot prompting.44 This approach is exible
likely that the LLM may have had access to this data during its and immediate, but its performance can vary signicantly with
extensive pretraining. Fig. 3 illustrates the nal distribution of prompt phrasing, the context provided and even the order of
This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.

groups across the train and test sets. words.

Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

2.1.4 Instruction creation. We manually cra question– When interacting with the LLM, each data point is formatted
answer textual pairs by converting reaction SMILES strings and as a conversation between a user and an assistant (the model
EC labels into clear natural language queries. Each reaction itself), as follows:
data point is converted into a single question-answer pair, but
A general system prompt assigns the model the role of
we generate multiple templates per task to avoid overtting on a biochemically knowledgeable assistant.
a single prompt style. For EC number prediction, example
The user prompt species the task, phrasing it in a exible
templates include: way to ensure a certain degree of variability. As mentioned

above, diverse templates are used in order to prevent overtting

on specic question structures.

The assistant provides a direct answer, formatted with

tagging elements such as hECi . h/ECi (for the EC number
The model is instructed to return only the answer (the EC prediction task), to enhance consistency and ease of parsing.
number in this case), without additional text. We vary synonyms A visual example of this is shown in Fig. 4.
e.g. “EC number” vs. “Enzyme commission number”, and 2.2.2 Fine-tuning. Fine-tuning refers to the process of
qualiers such as “feasible” vs. “probable” across up to 14 adapting a pretrained model to perform specic tasks by
unique templates per task. updating its parameters on a new dataset. This approach allows
the model to specialize in a narrower domain while retaining its
general pretrained knowledge. As models sizes grow into tens or
2.2 LLM interaction and adaptation hundreds of billions of parameters, retraining every weight for
2.2.1 In-context learning. In In-Context Learning (ICL), we each new task becomes prohibitively expensive in memory and
interact with a LLM solely through prompting. Prompting compute. To ne-tune our models efficiently, we use Parameter-
means giving a set of instructions to the model in natural Efficient Fine-Tuning (PEFT) techniques that leave most of the
language in order to make it perform a task: answering, base model weights unchanged in the process. Specically, we

Fig. 4 Example of a zero-shot prompt for the EC number prediction task. The model first receives the system prompt: a general prompt with
instructions that inform it about the task to perform. The [TASK] here is EC number prediction, and the [OBJECTIVE] is to assign the 4 digits of the
EC number given only reactants and product in SMILES notation. After that, the model receives the reaction SMILES from the user as
a [REQUEST], and the model associates an EC number to it as the [RESPONSE].

© 2025 The Author(s). Published by the Royal Society of Chemistry Digital Discovery
 View Article Online

Digital Discovery Paper

use Low-Rank Adaptation (LoRA).55 LoRA allows ne-tuning by
Efficiency: the model's computational cost-effectiveness,
updating only a small subset of the model's parameters, particularly in resource-constrained environments.
signicantly reducing computational demands. Instead of We aimed to use a LLM that balances computational power
directly updating a weight matrix of the pretrained model with exibility, ensuring it can be customized for specialized
W ˛ℝnm , LoRA models the update as the product DW = AB, biochemical applications. We prioritize general-purpose LLMs to
where W ˛ℝnm and B˛ℝrm are matrices with a rank r  min(n, evaluate their adaptability and scalability across multiple
m). The rank determines the size of the two matrices, and biochemical tasks. Equally important was choosing an open-
during forward passes the effective weight matrix becomes source model to facilitate accessibility and enable further devel-
opment by other researchers. Given these requirements, we
This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.

W0 = W + DW = W + AB (1) selected models from Meta AI's Llama 3.1 family,57 specically
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

the 8B and 70B parameter versions. The smaller 8B model offers

The small rank is what ensures that A and B contain far fewer a trade-off between efficiency and exibility for exploratory or
parameters than W and this drastically reduces memory foot- lower-resource settings, while the 70B model provides greater
print and ne-tuning time. An illustration of the algorithm is power. Further, we employed the instruct versions of these
shown in Fig. 5. While the model can be loaded in a quantized models, both for in-context learning and ne-tuning. These
format for efficient memory usage, ne-tuning occurs on variants are ne-tuned on instruction-response pairs, helping
a limited percentage of parameters that are stored in full/half them generate responses that align with the given instructions.
precision. This approach has shown to yield performance Lastly, we utilize both base models in the 4-bit quantized format
levels close to those of full model ne-tuning, while maintain- to reduce computational costs and inference time.
ing the model's general reasoning abilities and core capabil-
ities.55,56 Importantly, during ne-tuning, the model is provided
2.3 Evaluation metrics
with interrogation pairs in exactly the same way as in In-Context
Learning, using the same prompt template and tagging For the EC prediction task, a prediction is correct if the digits
conventions. Only now each question-answer pair is used to match exactly those of the ground truth. If only the rst digit is
adapt its internal parameters. correct, the model correctly predicted the EC class. If the rst
The use of LoRA adapters is particularly advantageous for two digits are correct, the prediction is correct up to digit EC2,
LLMs like ours. These adapters can be “plugged in” for domain- and so on. For the accuracy, we always compute the macro-
specic tasks and subsequently removed to revert to the base average to show performance across classes, treating each
model, which remains unaffected by ne-tuning, thereby class as equally important. Additionally, for the main class we
keeping computational costs under control. report F1 score, precision and recall, as they help provide
[Link] Model selection. In selecting a model for our a more complete picture especially with imbalanced datasets
biochemical prediction tasks, our primary selection criteria are: such as ours. To evaluate product and substrate predictions, we

Power: the model's ability to handle complex tasks and categorize predicted SMILES strings into ve distinct groups:
achieve high accuracy;
Invalid (I): if the RDKit parsing fails, the prediction is not

Flexibility: its ability to tackle diverse tasks both in in- a valid SMILES string, either because chemically implausible or
context learning and ne-tuning settings; incorrectly formatted. If the parsing succeeds, the pipeline
proceeds to the next steps;

Fig. 5 Illustration of LoRA framework. The input vector xi at layer i is passed through both the frozen i-th weight matrix of the pretrained model,
and the i-th LoRA head. After both blocks process the input, the two representations are summed together to obtain a new representation xi+1.
This procedure is repeated for all layers.

Digital Discovery © 2025 The Author(s). Published by the Royal Society of Chemistry
 View Article Online

Paper Digital Discovery

Non-canonical valid (NCV): the predicted SMILES does not
LoRA (41.9 M, 0.52% for the 8B, 207 M, 0.29% for the 70B):
correspond to the ground truth molecule, but it represents this is the basic setting and the one used throughout the paper.
a plausible molecule in a non-canonical SMILES representation; The adapter tunes all the attention modules and the feed-

Canonical valid (CV): the predicted SMILES does not forward networks (FFN).
correspond to the ground truth molecule, but it represents For training we use a single NVIDIA Tesla A100 80 GB GPU.
a plausible molecule in the canonical SMILES representation; To isolate the effect of model size, we report single-task training

Non-canonical match (NCM): the predicted SMILES does durations: the 8B model takes ∼30 min per epoch, while the 70B
represent the ground truth molecule, but in a non-canonical model takes ∼4 h per epoch with a batch size of 8. Inference
SMILES format; latency is ∼2.5 s per sample for the 8B model and ∼5.7 s per
This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.

Canonical match (CM): the predicted SMILES string does sample for the 70B model on a Intel XEON E5-6448Y 32C 2.1
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

represent the ground truth molecule in the canonical format: GHz CPU.
the predicted string exactly matches the ground truth one.
For valid chemicals, we additionally examine molecular
similarity to determine the potential relevance of the generated 3 Results and discussion
SMILES, using the Tanimoto similarity coefficient aer In this section, we present the results of ne-tuning the selected
computing daylight ngerprints58 for each molecule. A Tanimoto Llama models. The analysis encompasses ST and MT setups,
similarity >0.85 is oen considered indicative of structurally along with experiments designed to evaluate performance in
similar molecules, suggesting that even incorrect predictions low-data regimes and across different ne-tuning schemes. For
may still be chemically meaningful. High similarity scores could each task, the performance is compared against baselines.
for example suggest that the LLM-generated molecule might
serve as an alternative substrate in retrosynthetic applications,
potentially offering novel biochemical insights. It is important to 3.1 Single-task ne-tuning
note that the SOTA results that we mention are taken from Llama-3.1 models accurately predict the highest level of EC
existing studies and are based on models trained on the entire number classication, yet show a decline when tasked with the
ECREACT dataset, which comprises unique n = 62 222 enzymatic second and third digit. In a single-task settup, the Llama-3.1
reactions aggregated from four different databases. In contrast, model family exhibits some difficulties with exact product and
our experiments are conducted using only the reactions from the substrate prediction tasks. Interestingly, we nd that reason-
BRENDA database. While this difference in training data size ably large percentages of uncorrect predictions show a high
limits direct comparisons with SOTA models, our setup allows for Tanimoto similarity with the correct predictions, which can
a holistic experimental design within reasonable computational potentially still be useful in biochemical workows.
limits. While this limits the comparison to certain extent, it 3.1.1 EC prediction task. The 70B model accuracy for EC
allows us to focus on a single well-curated database, we can class prediction is consistent across most classes, with an
systematically evaluate different model sizes, ne-tuning strate- average accuracy of 91.7%. This indicates that it is fairly simple
gies, and data regimes, while still capturing a diverse range of for the ne-tuned model to correctly assign the highest EC
enzymatic reactions. All results are averaged over N = 3 experi- number given any reactant, product pair as request. However,
ments to provide robust performance metrics, with standard class 4 exhibits a noticeable performance dip, despite not being
deviations reported where applicable. the least-represented class in the dataset. To explore the
model's misclassication patterns, we present the confusion
matrix for EC class prediction in Fig. 6. The matrix reveals that
2.4 Fine-tuning setup classes 4 and 5 are sometimes wrongly assigned to each other.
All models are trained with a learning rate lr = 0.002 using In classes 1, 2 and 3 rare instances of misclassications either
a linear decay scheduler, and {a = 32, r = 16} for the LoRA happen between 1 and 2 or assign the reactions to class 4.
adapter. Minimal hyperparameter tuning was performed (lr ˛ For EC2 predictions, we see that the model frequently
[0.0005, 0.005], a ˛ {32, 64}, r ˛ {16, 32}). We explore two new misclassies subclasses within the same main class. This
LoRA setups in addition to the default one, to evaluate the trade- relates to the EC2 category distribution per main class. For
off between ne-tuning parameter count and model perfor- instance, class 1.X.X.X has numerous subclasses, whereas
mance. We adopted the same setups choice by ref. 43, to ensure classes 5.X.X.X and 6.X.X.X only have a few. Rare subclasses,
consistency with prior LoRA ne-tuning literature in the such as 2.2.X.X or 4.99.X.X, show clear exceptions with the
chemical domain. Here we list them, including in parenthesis model misclassifying outside the main class, likely due to their
the number of trainable parameters and their percentage with underrepresentation. Additionally, structural similarities
respect to the pretrained, base model: within main classes may further contribute to confusion,

LoRA light (6.8 M, 0.09% for the 8B, 32.8 M, 0.05% for the independent of dataset imbalance. The confusion matrix for the
70B): we only ne-tune the query and key matrices within the EC2 level, alongside the test set distribution for that depth, is
attention modules [qproj, kproj]. shown in Fig. 7.

LoRA attention (13.6 M, 0.17% for the 8B, 65.5 M, 0.09% for The accuracy of the model declines at deeper EC levels,
the 70B): we extend ne-tuning to all matrices within the reecting the increasing challenge of capturing hierarchical
attention mechanism [qproj, kproj, vproj, oproj]. enzyme relationships. These difficulties also stem from

© 2025 The Author(s). Published by the Royal Society of Chemistry Digital Discovery
 View Article Online

Digital Discovery Paper

This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

Fig. 6 Confusion matrix representing Llama-3.1 70B accuracy in predicting the enzyme class given reactants and substrates, for one experiment.
The out-of-diagonal elements show how examples are misclassified. The histogram on the right shows the test set distribution stratified by main
class, roughly following how training data is distributed.

increased combinatorial complexity of sublevels and class so larger classes carry proportionally more weight, a limitation
imbalance. In fact, at level EC2 the model performs best for noted by the SOTA authors.20 Extended metrics (F1 score,
class 6 and worst for class 1, a result that aligns with the dataset precision, recall) for the EC class prediction task are reported in
distribution shown in Fig. 1: EC class 1 has a highly branched Appendix Fig. 14. Additionally, we compare our ne-tuned
EC2 structure, with 1.1.X.X accounting for almost half of the models with a zero-shot baseline with Llama-3.1 70B, in Table
samples, introducing class imbalance. Conversely, class 6 has 1. We see that the zero-shot prompting approach lacks far
a limited number of balanced subcategories (6.2.X.X and behind the ne-tuned models of this size and general capabil-
6.3.X.X), simplifying subclass predictions. Fig. 8 illustrates the ities. This indicates that at present it seems inevitable to ne-
model's performance in predicting EC numbers up to level EC3, tune the general purpose model for a complex and domain-
stratied by main class. specic task such as EC classication in biochemistry.
The ne-tuned 70B model comes on top of the ne-tuned 8B 3.1.2 Product and substrate prediction tasks. The 70B
model predicting EC digits at any depth. However, the model generates a high proportion of chemically valid mole-
compared SOTA retains a signicant edge across all levels (EC1 cules in canonical format, with canonical matches (the output
accuracy: 96.2%, EC2 accuracy: 93.4.6%, EC3 accuracy: string matches the ground truth string as it is) accounting for
91.6%).20 Please note that the authors have performed a micro- 24.9% and 13.0% for products and substrates respectively.
average, while we perform a macro-average that takes class While FS shows a higher percentage of canonical matches, RS
imbalance into account. Macro-averaged accuracy at any EC has a greater proportion of chemically valid but incorrect
level depth is computed by rst calculating the accuracy within predictions, indicating that retrosynthesis may involve more
each EC number class, and then taking the simple mean of complex structural reasoning. Chemically invalid predictions
those per-class accuracies. In this way, each class contributes are minimal (<5% of the total test set for both tasks), and wrong
equally regardless of its size. In contrast, micro-averaged accu- generations due to e.g. formatting errors are rare (<2%). This
racy is computed by averaging all test examples across classes, demonstrates that the LLMs can easily adhere to complex

Digital Discovery © 2025 The Author(s). Published by the Royal Society of Chemistry
 View Article Online

Paper Digital Discovery

This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

Fig. 7 Confusion matrix representing Llama-3.1 70B accuracy in predicting the EC number up to the second digit (EC2), given reactants and
substrates, for one experiment. The out-of-diagonal elements show how examples are misclassified. The misalignment in the diagonal elements
is due to the set of predicted classes having elements that are not present in the test set, like subclasses 3.3.X.X and 4.5.X.X, that the model
predicts in a few cases. The histogram on the right shows the test set distribution stratified by EC2 subclass.

domain specic grammar like SMILES and to requested output observed in the Appendix Fig. 11. Thus, for products, the model
formats, which is a useful property for the analysis of model either predicts a molecule very close to matching the ground
results. However, these results are not yet competitive with the truth, or it gets the wrong chemical. For substrates on the other
SOTA model19 (49.6% and 60.0% accuracy for exact matches for hand, having longer strings and more options in the RS task
FS and RS respectively). Note that the dataset used for our leads to generating many substrates that are not correct, but
models is not exactly the same as the one from SOTA, making show a relatively high Tanimoto score. Analyzing the highest
the results not directly comparable. Pie charts in Fig. 9 display Tanimoto values, we see that 7.1% of chemically valid products,
the distribution of predictions across the ve categories for FS and 6.7% of chemically valid substrates, report a score equal to 1.
and RS tasks, respectively, for Llama-3.1 70B. Examples of these chemicals are reported in Appendix Fig. 17 and
When the model fails to predict the exact molecule, it gener- 18. We summarize the results for both Llama-3.1 8B and Llama-
ates relevant alternatives that may hold biochemical utility in 3.1 70B on FS and RS tasks, including a baseline 0-shot perfor-
12% and 35% of the cases, for products and substrates respec- mance and comparison to SOTA in the Appendix Table 6.
tively. We classify such an output with biochemical utility if the 3.1.3 Generalization over unseen tasks. When ne-tuned
generated molecule shows a high Tanimoto similarity to the on a single biochemical task, the model not only retains its
correct output. Focusing on the set of valid chemicals, Tanimoto general capabilities on unseen, related tasks within the same
similarity scores are computed and shown in Fig. 10. In the sub-domain but also improves its performance compared to its
dataset, SMILES for products are shorter than substrates on zero-shot baseline. To evaluate this generalization effect, we test
average, and we also observe that for branching reactions, the set each of the three single-task (ST) ne-tuned models on the two
of products that are possible from certain substrates in a forward tasks they were not trained on, comparing their performance to
synthesis task, is generally smaller than the set of possible the respective zero-shot baseline. Results show that ne-tuning
substrates reachable from a product in a retrosynthesis task, as on either the FS or RS task signicantly improves EC class

© 2025 The Author(s). Published by the Royal Society of Chemistry Digital Discovery
 View Article Online

Digital Discovery Paper

This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

Fig. 8 Llama-3.1 70B accuracy in predicting the EC number up to level EC3, organized by main class. Accuracy measures if the model correctly
matches the ground truth EC number up to the EC level specified on the x-axis. Accuracies are computed considering each (sub)class as equally
weighted. These distribution patterns influence model performance, irrespective of reaction complexity or SMILES grammar.

Table 1 Performance comparison between Llama-3.1 70B and Llama-3.1 8B models fine-tuned for the EC prediction task, from predicting level
EC1 only, to all digits up to EC3 included. A baseline 0-shot prompting approach with the 70B model is reported as well. We also show our
models performance in micro-average next to the SOTA model20 in micro-average. Note that the dataset is not exactly the same (see Subsection
2.1) and thus results are still not entirely comparable

Metric Llama 8B Llama 70B Llama 70B 0-shot Llama 70B micro-avg SOTA

EC1 accuracy (%) 86.4  0.6 91.7 0.5 29.6  0.7 92.4  0.2 96.2
EC2 accuracy (%) 56.5  1.5 61.7  1.1 8.7  0.5 75.6  0.1 93.4
EC3 accuracy (%) 40.5  0.6 49.2 0.7 5.7  0.4 68.1  0.1 91.6
Validity (%) >99.9 100.0 89.4  0.3 100.0 —

Fig. 9 Pie charts showing the average distribution of predictions for forward synthesis (FS, left) and retrosynthesis (RS, right) for Llama-3.1 70B.
The outer layer indicates the proportion of correctly generated (blue/green), invalid chemicals (red), and wrongly generated predictions (grey),
while the inner layer differentiates correct outputs from structurally valid but incorrect outputs. Invalid and wrongly formatted predictions remain
<5% and <2% for both tasks, respectively. Results for each category are obtained averaging over N = 3 experiments, with standard deviations
below 5% of each category value. Percentages are shown for >2% slices only.

Digital Discovery © 2025 The Author(s). Published by the Royal Society of Chemistry
 View Article Online

Paper Digital Discovery

This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

Fig. 10 Histograms of Tanimoto similarities of ground truths against products (left) and substrates (right), that the model predicts as chemically
possible but not corresponding to the ground truth.

Table 2 Generalization of ST fine-tuned Llama-3.1 70B models when tested on the unseen related biochemical tasks. The zero-shot baseline is
reported for comparison. Performance on the original fine-tuned task is omitted to emphasize cross-task generalization. The reported match
values are here considered regardless of canonicity. The invalid category includes both incorrect SMILES notation as well as wrongly formatted
output from the LLM

EC FS RS

Fine-tuned on EC1 [ (%) Invalid Y (%) Match [ (%) Invalid Y (%) Match [ (%) Invalid Y (%)

EC — — 12.9 31.1 0.3 55.3

FS 54.3 0.3 — — 1.4 3.3
RS 42.1 6.3 0.6 5.5 — —
ICL 0-shot 29.6 10.6 <0.1 53.5 <0.1 82.4

prediction accuracy, nearly doubling the zero-shot baseline signicantly reduced training samples. Specically, we analyze
performance. Likewise, a model ne-tuned exclusively on EC performance degradation when the training set size is limited to
number prediction improves FS match accuracy from nearly 0% 600 and 200 compared to our default training (∼1800 samples
to 12.9% while also reducing invalid predictions by half. Table 2 per task). This analysis is conducted for both models, to provide
presents the generalization results, where each ne-tuned insights into their scalability when data availability becomes the
model is tested on the two unseen tasks. bottleneck. Both models show a steady performance increase
when training data is increased. The larger architecture holds an
3.2 Multitask ne-tuning edge over the smaller one regardless of data size across almost all
Using a multitask setup we show that we can improve perfor- tasks, conrming again its greater capabilities.
mance through the use of synergistic information from the For a fairer comparison, we include a simple XGBoost
related task, in particular for FS and RS tasks. For these the baseline. XGBoost59 is a gradient boosting model that performs
model performance for matches (regardless of canonicity) well with structured data and does not rely on large-scale pre-
increases by 7.9% and 5.3% respectively. The three ST datasets training, making it a suitable reference for evaluating whether
are merged together to provide the dataset used for the MT setup. LLM ne-tuning truly adds value in data-limited biochemical
The Llama 70B and 8B models are both ne-tuned, using the prediction tasks. We nd that across all tasks and for each data
best-performing conguration identied in the single-task scenario, our models outperform the XGBoost model. We report
experiments. Performance is compared against single-task our ndings in Table 4. More details on how XGBoost is trained
setups to assess multitask learning benets, with the main are reported in the Appendix in Subsection A.6.
results reported in Table 3.
3.4 Impact of different LoRA setups
3.3 Exploring low-data regimes We observe that adding more trainable parameters can lead to
Fine-tuned LLMs show promise in low data regimes: for Llama- performance improvement for most tasks. This indicates the
3.1 70B, we report almost double EC class accuracy when importance of parameter-efficient learning strategies in
comparing zero-shot prompting (29.6%) with the ne-tuned domains where ne-tuning is essential. We see the trend that
version with only N = 200 training samples (55.3%). We repli- LoRA default performs better than LoRA attention and LoRA
cate low-data scenarios to evaluate how the models perform with light in almost all settings. In most tested cases for the 8B

© 2025 The Author(s). Published by the Royal Society of Chemistry Digital Discovery
 View Article Online

Digital Discovery Paper

Table 3 Performance comparison between single-task and multitask that traded interpretability for coverage. Below, we contrast the
setups for Llama-3.1 8B and Llama-3.1 70B. Blue cells represent historical approaches, explain how LLMs complement special-
performance improvement, orange cells represent performance
ized tools, and how such generalist models currently t and will
reduction. The reported match values are here considered regardless
of canonicity. The categories “Match + (TS = 1)” and “Match + (TS > evolve in the current tool landscape.
0.95)” add to the previous one the share of valid chemicals with
Template-based models (LHASA,4 SYNTHIA6) are deter-
a Tanimoto score equal to 1 and greater than 0.95 respectively. ministic and transparent. The outputs are easy to inspect, but
Numbers are presented in bold if the best performance improvement every new reaction type demands manual rule curation, leading
does not fall within one standard deviation from the second-best
to rule explosion and limited ability to generalize.

Data-driven statistical learners (EFICAz10) can exploit larger
This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.

Llama-3.1 70B
databases, yet lose interpretability and can inherit the human
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

Task Metric (%) ST MT D biases embedded in handcraed descriptors.

Deep learning models (such as the molecular15 and enzy-
EC Accuracy EC1 [ 91.7 86.4
matic18 transformers) reach state-of-the-art accuracy by discov-
Accuracy EC2 [ 61.7 65.1 ering hidden patterns. They are, however, computationally
Accuracy EC3 [ 49.2 48.9 −0.3 intensive, data hungry and largely less interpretable.
FS Match [ 25.9 33.8
LLMs require large-scale pretraining, yet could cover
a broad task spectrum.38,39 They offer tentative interpretability
Match + (TS = 1) [ 33.0 44.4
via natural language interrogation, although hallucinations
Match + (TS > 0.95) [ 34.2 45.4 remain a risk.
Our benchmarks conrm that pretrained LLMs still require
Invalid Y 4.8 4.9 +0.1
task-specic tuning before tackling complex biochemical pipe-
RS Match [ 13.9 19.2
lines. Overall applicability depends on the goal: even aer ne-
Match + (TS = 1) [ 20.6 30.1 tuning them, for high-precision tasks, specialized models win.
However for breadth, adaptability and human-centric interac-
Match + (TS > 0.95) [ 36.1 45.4
tion, LLMs are compelling. Thanks to their fast repurposing and
Invalid Y 4.4 3.0 unied conversational interface, they can work as control layers
for the existing toolbox. Compact code logic, plug-in nature and
access to legacy tools raise the bar for automation. Researchers
already query LLMs for literature, accessing collected knowl-
edge more effectively. It is important to note that the LLM
model, LoRA attention performs slightly better than LoRA light, evaluation is not as reliable as human evaluation in chemical
while for the 70B model, LoRA light performs slightly better reasoning. Practitioners should choose LLMs for exploratory or
than LoRA attention in most tested cases. Performance across hypothesis-generation stages, rapidly changing tasks, or
all tasks with different LoRA setups are reported in Table 5. settings with sparse data, and stick to specialized models for
best accuracy.
Standard LLMs are not yet robust for complex zero- or few-
3.5 Context: where do LLMs t? shot biochemical tasks, so for now they serve as sparring part-
Computational tools for reaction prediction have progressed ners rather than oracles. But progress is rapid: during this study
from rule libraries to deep task-specic models, following waves we saw multimodal LLMs emerge, along with reasoning

Table 4 Performance of Llama-3.1 8B and Llama-3.1 70B across all tasks and for different training set sizes. The reported match values are here
considered regardless of canonicity. Each task is trained on a slightly different amount of samples (±20) because of how data has been split, thus
we report a reference number of 1800 samples in the corresponding rows. Numbers are presented in bold if the best performance does not fall
within one standard deviation from the second-best. A baseline XGBoost model is reported for comparison

EC FS RS

Model Train set size EC1 [ (%) EC2 [ (%) EC3 [ (%) Match [ (%) Invalid Y (%) Match [ (%) Invalid Y (%)

LLama-3.1 8B 200 43.5 15.5 8.5 2.6 4.6 0.2 11.2

600 65.6 30.1 17.4 8.3 6.4 2.8 10.2
∼1800 86.4 56.5 40.5 18.4 9.4 15.1 4.3
LLama-3.1 70B 200 55.3 28.5 17.7 7.7 7.7 2.9 7.3
600 73.5 45.8 33.1 11.0 4.4 4.1 7.2
∼1800 91.7 61.7 49.2 25.9 4.8 13.9 4.4
XGBoost 200 32.7 4.9 <0.1 <0.1 — <0.1 —
600 40.9 6.0 1.7 1.9 — 2.5 —
∼1800 54.0 23.7 15.9 5.1 — 3.6 —

Digital Discovery © 2025 The Author(s). Published by the Royal Society of Chemistry
 View Article Online

Paper Digital Discovery

Table 5 Performance of Llama-3.1 8B and Llama-3.1 70B across all tasks and for different fine-tuning setups. Perfomance for all tasks increases
with the number of fine-tuned parameters, the only exception being the attention fine-tuning for Llama-3.1 70B, where an increase in FS
performance comes with a degradation in RS and EC prediction tasks. The reported Match values are here considered regardless of canonicity.
Numbers are presented in bold if the best performance does not fall within one standard deviation from the second-best

EC FS RS

Model LoRA type EC1 [ (%) EC2 [ (%) EC3 [ (%) Match [ (%) Invalid Y (%) Match [ (%) Invalid Y (%)

LLama-3.1 8B light 72.0 44.1 30.3 10.2 9.1 4.7 12.8

attention 82.0 48.4 31.9 11.3 7.9 6.9 10.1
This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.

default 86.4 56.5 40.5 18.4 9.4 15.1 4.3

Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

LLama-3.1 70B light 85.8 58.5 45.2 21.4 6.0 13.7 3.9
attention 78.8 48.0 34.9 25.6 5.5 9.8 3.3
default 91.7 61.7 49.2 25.9 4.8 13.9 4.4

variants that spend extra inference time or call external tools to
Interpretability analysis: we focus on predictive metrics
verify answers and surface their chain-of-thought. We therefore only and do not analyze how the model assigns EC numbers
expect future workows to resemble small teams of models: one or predicts reaction outcomes, nor whether its intermediate
LLM engages the user, then hands off to specialist agents for reasoning aligns with biochemical knowledge. SMILES
planning or calculation. As these guard-railed systems mature, strings are not inherently human-readable, but can be con-
a single general-purpose LLM may absorb many routine tasks. verted into molecular graphs for deeper analysis of the
Until then, pairing a LLM with specialist tools remains the model's prediction. Inspecting LLM-generated rationales
safest and most productive path. step-by-step via chain-of-thought prompts is a promising
direction for future work, both for user's interpretability and
to feedback them back into the model for more robust
3.6 Limitations
responses.
While our study demonstrates the potential for researchers to
work with LLMs when studying biochemical reactions, several
limitations must be acknowledged. Addressing these will be key 4 Conclusions
to improving both model accuracy and applicability in real-
world biochemical workows. In this study, we systematically evaluated the potential of

Potential data leakage: although we ne-tune the LLM to Large Language Models (LLMs) for biochemical reaction
evaluate performance in low-data regimes, it is possible that the prediction, focusing on enzyme commission classication,
model has already been exposed to similar biochemical reaction forward synthesis, and retrosynthesis. By ne-tuning Llama-
data during pretraining, as such datasets are available online. 3.1 models, we demonstrated that LLMs can answer
For a fairer comparison, future evaluations should ensure that biochemical questions, although they are not yet fully
test sets are composed of truly held-out reactions that cannot be competitive with specialized models. Fine-tuning signi-
scraped or indirectly inferred from pretraining text on the cantly improves performance over in-context learning, with
internet. This would provide a clearer measure of the model's Llama-3.1 70B achieving 91.7% accuracy in EC class classi-
generalization ability beyond memorization. Moreover, our cation. Fine-tuning on a single task does not degrade the 70B
similarity analysis (Appendix A.1) shows that, even aer model capabilities on unseen related tasks, as we observe
grouping branching reactions, substantial analogue overlap performance improvement compared to zero-shot baselines
remains between train and test substrates and products, indi- that use the base, pretrained model. Multitask learning
cating that stricter split protocols are required. enhances forward synthesis and retrosynthesis predictions,

Data constraints: our study is based on the BRENDA subset with a match accuracy of 33.8% and 19.2% respectively,
of the ECREACT dataset, which, while extensive, does not fully indicating that leveraging shared biochemical knowledge
cover the diversity of enzymatic reactions and does not allow improves generalization. Additionally, LLMs have potential in
a direct comparison to current SOTA model. The limited low-data regimes, making them valuable for applications
representation of certain EC subclasses affects generalization. where labeled data is scarce. The choice of ne-tuning
Expanding training to the full ECREACT dataset or integrating strategy impacts the performance, with LoRA offering an
additional reaction databases could mitigate this issue and efficient and scalable adaptation method. Despite these
enhance model robustness, yet also here, ECREACT has been strengths, several challenges remain: LLMs struggle with
preprocessed and simplies complex biochemical reaction handling rare EC subclasses and ensuring reliable predic-
mechanisms to a certain degree. tions. As LLM architectures continue to evolve, their integra-

Computational constraints: ne-tuning LLMs is computa- tion into biochemical workows has the potential to
tionally expensive, even with PEFT strategies like LoRA, limiting accelerate discoveries in enzyme-substrate prediction and
accessibility for resource-constrained environments. biocatalysis design.

© 2025 The Author(s). Published by the Royal Society of Chemistry Digital Discovery
 View Article Online

Digital Discovery Paper

Grouping of related reactions: reactions that represent the

Author contributions
same underlying biochemical process but differ slightly due to
J. M. Weber and L. Di Fruscia conceptualized the study. L. Di variations in substrate or product representations are group-
Fruscia led data curation, formal analysis, investigation, ed: whenever a {substrate, EC} or {product, EC} pair maps to
methodology, soware development, validation, visualization, multiple valid counterparts, we treat all those reactions as
and draing of the manuscript. J. M. Weber supervised the a group that must stay together in any train/test split. A group
project, provided resources, and reviewed and edited the of size N contains N distinct reactions sharing the same
manuscript. {substrate, EC} or {product, EC}. With this denition, a group
size of N = 1 indicates a unique reaction with no branching
This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.

alternatives. We refer to this as substratebranching and pro-

Conflicts of interest
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

ductbranching respectively. All reactions within a group are

allocated to the same dataset split (training or test) to avoid
There are no conicts to declare.
leakage.

Avoidance of task-specic leakage: in forward synthesis
Data availability (FS) and retrosynthesis (RS), if a reaction appears in FS, then any
of its counterparts with the same product and EC number but
Data and code for this article are available at https:// different substrates, must not appear in RS. This prevents the
[Link]/Intelligent-molecular-systems/ model from gaining undue advantage by being exposed to
LLM_netuning_for_biochemistry. This study was carried out related information in the training phase.
using publicly available data at [Link] Branching groups distribution varies a lot on whether we
rxn4chemistry/biocatalysis-model. DOI: [Link] look at the products or the substrates, as most of the
10.5281/zenodo.17224080. branching substrates only lead to 2 or 3 possible products, but
the reverse task has a wider spread. We report this in Fig. 11.
We further analyze the substrates and products to assess
A Appendix whether overlapping reactions across groups are present. This
A.1 Data preprocessing and analysis is needed because e.g. a substrate, while branching into
The original dataset comprises 8496 reaction examples, multiple products, may also be part of a set of substrates
distributed across the seven EC classes as follows: Class 1 reachable from a specic product. We follow this by merging
(3361), Class 2 (1700), Class 3 (1596), Class 4 (964), Class 5 (504), those overlapping groups together and removing redundant
Class 6 (352), and Class 7 (19). We implemented a series of entries.
preprocessing steps to ensure a fair split across training and test A.1.1 Similarity check across splits. To determine whether
set and across tasks: our cold substrate and cold product splits still leak informa-

Canonicalization of SMILES representations: reactions tion through highly similar molecules, we compute the pair-
with substrates or products in different SMILES representations wise Tanimoto similarity between the training and test sets,
are unied by converting all SMILES strings to their canonical separately for substrates and products. For each training
forms using the RDKit library. This ensures that duplicate molecule, we compute the mean of its ten most similar test set
{substrate, product} pairs, differing only in molecular repre- neighbours. The resulting distributions reveal that 38% of
sentation, are identied and removed. In this step, 362 reac- training products and 67% of training substrates have a Tani-
tions (4.2% of the total) are reformatted, and no reactions are moto score >0.85. Such extensive overlap persists despite our
discarded. grouping of duplicate {substrate, EC} and {product, EC}

Fig. 11 Histograms of group size for duplicate {substrate, EC} (left) and duplicate {product, EC} (right). We can observe that while most duplicate
reactions branch into two possible products, substrates tend to branch into larger groups.

Digital Discovery © 2025 The Author(s). Published by the Royal Society of Chemistry
 View Article Online

Paper Digital Discovery

This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

Fig. 12 Histogram of mean top-10 Tanimoto similarities between training and test molecules. Substrates are shown in green and products in
blue. Each bar aggregates training molecules whose average similarity to their ten closest test set analogues falls in that interval. The dashed line
at 0.85 marks the high-similarity regime, that includes 67% of substrates and 38% of products.

Fig. 13 Scatterplots of SMILES-length correlations. Top: forward synthesis (Pearson r = 0.39). Bottom: retrosynthesis (Pearson r = 0.38). This
shows that true input and output lengths are weakly correlated. On the plots is further reported the correlation coefficient between predictions
and ground truths length, showing strong correlation for both tasks.

reactions, highlighting that exact-match splitting alone is SMILES lengths in our test set, we rst compute the Pearson
insufficient to eliminate analogue leakage in enzyme reaction correlation coefficient r between true substrate and true product
datasets. lengths, obtaining r ∼ 0.39, which indicates only a weak linear
A.1.2 Input-output SMILES length correlation. To quantify association. We then evaluate how well our model reproduces
the intrinsic relationship between substrate and product that trend by correlating predicted with true lengths: for

© 2025 The Author(s). Published by the Royal Society of Chemistry Digital Discovery
 View Article Online

Digital Discovery Paper

This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

Fig. 14 Left: radar plot showing accuracy, precision, recall and F1 score for Llama-3.1 70B and Llama-3.1 8B averaged over N = 3 experiments.
The plot demonstrates consistent outperformance of the larger model over the 8B variant across all metrics. For comparison, we also show the
70B model perforsmance when it is not fine-tuned, in a zero-shot format. Right: EC class accuracy for the fine-tuned Llama-3.1 70B stratified by
the class.

Table 6 Performance comparison between Llama-3.1 8B and Llama-3.1 70B models for forward- and retrosynthesis. All values for our fine-
tuned models are obtained averaging over N = 3 experiments, with standard deviations below 5% of each category value. A zero-shot baseline on
the pretrained 70B model is reported for comparison. We also report the SOTA model19 performance at the end. Note that the dataset is not
exactly the same (see Subsection 2.1) and thus results are still not entirely comparable. Numbers are presented in bold if the best performance
does not fall within one standard deviation from the second-best. NCM, CV, and NCV categories taken alone do not reflect model improvement,
thus we do not bold them. CM:canonical matching, NCM:non-canonical matching, CV:canonical calid, NCV:non-canonical valid

Model Task CM [ (%) NCM (%) CV (%) NCV (%) Invalid Y (%)

Llama-3.1 8B FS 17.6 0.8 53.8 14.0 9.4

RS 14.0 1.1 67.8 11.7 4.3
Llama-3.1 70B FS 24.9 1.0 58.8 10.5 4.8
RS 13.0 0.9 65.1 16.6 4.4
Llama-3.1 70B 0-shot FS <0.1 0 40.5 5.9 53.5
RS 0 <0.1 12.7 4.9 82.4
SOTA FS 49.6 — — — —
RS 60.0 — — — —

forward synthesis we observe r = 0.64, and for retrosynthesis r = A.3 Forward- and retrosyntesis comparison with ne-tuned
0.82. We report our results in Fig. 13. llama 8B
The 70B model performs better than the 8B one for forward
A.2 EC class prediction radar plots synthesis, and are both comparable when it comes to retro-
Computing precision, recall and F1 score alongside accuracy, we synthesis. We report the main results in Table 6, alongside the
observe that these four metrics are all consistent with each other SOTA model.
for both of our ne-tuned model sizes, with Llama-3.1 70B beating
Llama-3.1 8B in every metric. We compare them to a 0-shot
prompting setup with the pretrained Llama-3.1 70B as a baseline, A.4 Average Tanimoto scores in ground truth branching
observing the clear performance gap between in-context learning We observe that for the equally valid ground truths that the
with the larger model, against the ne-tuned 8B version. Focusing database stores for a given reaction, many examples show
on the ne-tuned 70B model, a stratication by main class shows a relatively low similarity score. Focusing on the product
us again that the values for the four metrics are consistent with prediction only, some of the reasons this happen can be due to
each other, per class, with class 4 being the most unbalanced. having a co-factor recorded in place of the main product, or
These ndings are reported in Fig. 14. some entries may report products that correspond to different

Digital Discovery © 2025 The Author(s). Published by the Royal Society of Chemistry
 View Article Online

Paper Digital Discovery

This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

Fig. 15 Average Tanimoto score computed across a ground truth product and each of its ground truth branching counterparts, for all groups
and stratified by group size. For branching groups of size 2, no standard deviation is shown as we only have one Tanimoto score computed
between the reference ground truth and its alternative option.

Fig. 16 Average Tanimoto score computed across a ground truth substrate and each of its ground truth branching counterparts, for all groups
and stratified by group size. For branching groups of size 2, no standard deviation is shown as we only have one Tanimoto score computed
between the reference ground truth and its alternative option.

reaction intermediates, a problem that strictly relates to the chemicals. Then, we compute the average Tanimoto score and
presence of branching reactions in the dataset. We compute the its standard deviation for that group. If the chemicals are all
average Tanimoto score across ground truth chemicals that similar to each other, we observe a high average with a relatively
belong to the same set of branching product/substrates, to get small standard deviation. On the other end, if the chemicals
insights over the chemical diversity of alternatives products/ present more variability, we expect to see a lower average with
substrates that are reported in the dataset. a wider spread in the standard deviation. We report the ndings
Given a group of size N, we compute the Tanimoto scores in Fig. 15 and 16.
between one element of the set and the remaining N − 1

© 2025 The Author(s). Published by the Royal Society of Chemistry Digital Discovery
 View Article Online

Digital Discovery Paper

A.5 Predictions with Tanimoto score equal to 1 for products

and substrates
This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

Fig. 17 Examples of predicted (left) vs. ground truth (right) products, when the prediction is not correct but produces a Tanimoto score equal to
1. We see that some predictions have an additional hydrogen (resulting in an OH group) while the ground truth recorded an oxygen ion (O–)
(rows 1, 3), while some others have a mismatch in chirality (rows 2 and 4).

Digital Discovery © 2025 The Author(s). Published by the Royal Society of Chemistry
 View Article Online

Paper Digital Discovery

This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

Fig. 18 Examples of predicted (left) vs. ground truth (right) substrates, when the prediction is not correct but produces a Tanimoto score equal to
1. We see that some predictions have a missing hydrogen (resulting in an oxygen ion O–) while the ground truth recorded an OH group (row 1),
while some others have a mismatch in chirality (e.g. rows 2 and 3).

© 2025 The Author(s). Published by the Royal Society of Chemistry Digital Discovery
 View Article Online

Digital Discovery Paper

A.6 XGBoost data preprocessing and training its effect on performance. Our tests show that the best perfor-
mance is achieved by including the EC number with the default
For each task, we encode the biochemical inputs into a struc-
scaling factor of 1, conrming that enzymatic information
tured format that XGBoost can process efficiently. Given its
contributes meaningfully to reaction prediction, even when it
reliance on tabular data, molecular and enzymatic information
constitutes a small fraction of the feature space.
is transformed into numerical feature vectors before being fed
into the model:

Molecular representation: for the product and substrate
prediction tasks, we represent molecules using Morgan nger-
Acknowledgements
prints to encode molecular structures into a xed-length binary
This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.

This work is supported by the [Link] program, a collaboration

vector. Each molecule is transformed in a 256-bit binary vector, between TU Del and dsm-rmenich, and is fully funded by
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

where each bit represents the presence or absence of a specic dsm-rmenich and the RVO (Rijksdienst voor Ondernemend
chemical substructure. Nederland).

Reaction representation: for the EC number prediction
task, the entire biochemical reaction (substrates + products) is
encoded as a 1024-bit reaction ngerprint. This representation References
captures reaction-specic features, such as changes in molec-
ular structures and functional groups. 1 B. Wiltschi, T. Cernava, A. Dennig, M. G. Casas, M. Geier,

EC number representation: we encode them in a way that S. Gruber, et al., Haberbauer, Marianne. Enzymes
preserves their hierarchical relationships. Instead of treating revolutionize the bioproduction of value-added
whole EC numbers as simple categorical labels, which would compounds: From enzyme discovery to special
ignore relationships between enzymes within the same cate- applications, Biotechnol. Adv., 2020, 40, 107520.
2 A. R. Alcántara, P. Domı́nguez de Marı́a, J. A. Littlechild,
gory, we encode them as four separate numerical features, one
M. Schürmann, R. A. Sheldon and R. Wohlgemuth,
for each EC digit. Each of these four digits is rst label-encoded,
then converted into a continuous representation via standard- Biocatalysis as key to sustainable industrial chemistry,
ization, approaching it as a regression task where similar EC ChemSusChem, 2022, 15(9), e202102709.
numbers remain closer in feature space. 3 R. A. Sheldon, Green chemistry and biocatalysis:
For all tasks, EC number label encoding is done on the full Engineering a sustainable future, Catal. Today, 2024, 431,
set of EC numbers, while standardization is performed using 114571.
only the training set statistics, preventing information leakage 4 W. T. Wipke and E. J. Corey, Computer-assisted design of
complex organic syntheses, Science, 1969, 166(3905), 178–
from the test set.
192.
A.6.1 Training and evaluation. XGBoost models are trained
separately for each task using the same training and test splits 5 W. T. Wipke, et al., Secs—simulation and evaluation of
as the LLM experiments. We run the model for 100 boosting chemical synthesis: Strategy and planning, Computer
rounds and include early stopping to avoid overtting. For the Representation and Manipulation of Chemical Information,
EC number prediction task, the problem is framed as a regres- 1977.
sion task with a squared loss, whereas for the other two tasks we 6 S. Szymkuć, et al., Computer-assisted synthetic planning:
use a logistic regression for the output bit-vector. The end of the beginning, Angew Chem. Int. Ed. Engl., 2016,
55(20), 5904–5937.

EC prediction task: the 1024-bit reaction ngerprint and
7 A. Bøgevig, et al., Route design in the 21st century: The
the standardized, 4D vector of the encoded EC number, repre-
sent input and output respectively. Evaluation is done by icsynth soware tool as an idea generator for synthesis
reverting the standardization process for the prediction and prediction, Org. Process Res. Dev., 2015, 19(2), 357–368.
checking whether the categorical encoding of the predicted EC 8 C. W. Coley, L. Rogers, W. H. Green and K. F. Jensen,
digits matches the true labels exactly. Computer-assisted retrosynthesis based on molecular

Product and substrate prediction: the input is represented similarity, ACS Cent. Sci., 2017, 3(12), 1237–1245.
by a concatenation of the 256-bit Morgan ngerprint with the 9 C. Claudel-Renard, C. Chevalet, T. Faraut and D. Kahn,
Enzyme-specic proles for genome annotation: Priam,
4D encoding of the EC number, and the output is a 256-bit
Nucleic Acids Res., 2003, 31(22), 6633–6639.
Morgan ngerprint. Since the ngerprints are binary, the
output is considered correct if the generated ngerprint exactly 10 W. Tian, A. K. Arakaki and J. S. Ecaz, a comprehensive
matches the ground truth ngerprint, as an upper bound proxy approach for accurate genome-scale enzyme function
of our “molecule matching” prediction task. inference, Nucleic Acids Res., 2004, 32(21), 6226–6239.
Since the EC number contributes with only four features to 11 Y. Li, et al., Deepre: sequence-based enzyme ec number
an input vector of hundreds of dimensions, we conducted prediction by deep learning, Bioinformatics, 2018, 34(5),
additional experiments to explore its impact. Specically we 760–769.
12 J. Y. Ryu, H. U. Kim and S. Y. Lee, Deep learning enables
inated the relative importance of the EC number by multi-
high-quality and high-throughput prediction of enzyme
plying its four components by factors ranging from 5 to 100. We
also completely removed the EC number from the input to test commission numbers, Proc. Natl. Acad. Sci. U. S. A., 2019,
116(28), 13996–14001.

Digital Discovery © 2025 The Author(s). Published by the Royal Society of Chemistry
 View Article Online

Paper Digital Discovery

13 B. Liu, et al., Retrosynthetic reaction prediction using neural 30 P. Lewis et al., Retrieval-augmented generation for
sequence-to-sequence models, ACS Cent. Sci., 2017, 3(10), knowledge-intensive nlp tasks, arXiv, 2021, preprint,
1103–1113. arXiv:2005.11401, DOI: 10.48550/arXiv.2005.11401.
14 A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, 31 S. Pan, et al., Unifying large language models and knowledge
A. N. Gomez, L. Kaiser, and I. Polosukhin, Attention is all graphs: A roadmap, IEEE Trans. Knowl. Data Eng., 2024,
you need, arXiv, 2017, preprint, arXiv:1706.03762, DOI: 36(7), 3580–3599.
10.48550/arXiv.1706.03762. 32 L. Wang, et al., A survey on large language model based
15 P. Schwaller, T. Laino, T. Gaudin, P. Bolgar, C. Bekas and autonomous agents, Front. Comput. Sci., 2024, 18(6), 186345.
A. A. Lee, Molecular transformer - a model for uncertainty- 33 M. C. Ramos, C. J. Collison, and A. D. White. A review of large
This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.

calibrated chemical reaction prediction, ACS Cent. Sci., language models and autonomous agents in chemistry,
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

2019, 5(9), 1572–1583. arXiv, 2024, preprint, arXiv:2407.01603, DOI: 10.48550/

16 G. Pesciullesi, P. Schwaller, T. Laino and J.-L. Reymond, arXiv.2407.01603.
Transfer learning enables the molecular transformer to 34 D. A. Boiko, R. MacKnight, B. Kline and G. Gomes,
predict regio- and stereoselective reactions on Autonomous chemical research with large language
carbohydrates, Nat. Commun., 2020, 11(1), 4874. models, Nature, 2023, 624(7992), 570–578.
17 R. Irwin, S. Dimitriadis, J. He and E. J. Bjerrum, 35 OpenAI et al., Gpt-4 technical report, arXiv, 2024, preprint,
Chemformer: a pre-trained transformer for computational arXiv:2303.08774, DOI: 10.48550/arXiv.2303.08774.
chemistry, Mach. Learn.: Sci. Technol., 2022, 3(1), 015022. 36 A. M. Bran, S. Cox, O. Schilter, C. Baldassari, A. D. White, and
18 D. Kreutter, P. Schwaller and J.-L. Reymond, Predicting P. S. Chemcrow, Augmenting large-language models with
enzymatic reactions with a molecular transformer, Chem. chemistry tools, arXiv, 2023, preprint, arXiv:2304.05376,
Sci., 2021, 12(25), 8648–8659. DOI: 10.48550/arXiv.2304.05376.
19 D. Probst, M. Manica, Y. G. N. Teukam, A. Castrogiovanni, 37 T. Guo et al., What can large language models do in
F. Paratore and T. Laino, Biocatalysed synthesis planning chemistry? a comprehensive benchmark on eight tasks,
using data-driven learning, Nat. Commun., 2022, 13(1), 964. arXiv, 2023, preprint, arXiv:2305.18365, DOI: 10.48550/
20 W. Qian, X. Wang, Y. Kang, P. Pan, T. Hou and C.-Y. Hsieh, A arXiv.2305.18365,.
general model for predicting enzyme functions based on 38 K. M. Jablonka, P. Schwaller, A. Ortega-Guerrero and B. Smit,
enzymatic reactions, J. Cheminf., 2024, 16(1), 38. Leveraging large language models for predictive chemistry,
21 Y. G. Nana Teukam, L. Kwate Dassi, M. Manica, D. Probst, Nat. Mach. Intell., 2024, 1–9.
P. Schwaller and T. Laino, Language models can identify 39 K. M. Jablonka, et al., 14 examples of how llms can transform
enzymatic binding sites in protein sequences, Comput. materials science and chemistry: a reection on a large
Struct. Biotechnol. J., 2024, 23, 1929–1937. language model hackathon, Digital Discovery, 2023, 2(5),
22 G. B. Kim, et al., Functional annotation of enzyme-encoding 1233–1250.
genes using deep learning with transformer layers, Nat. 40 Z. Zheng, et al., Shaping the water-harvesting behavior of
Commun., 2023, 14(1), 7370. metal–organic frameworks aided by ne-tuned gpt models,
23 J. Capela, et al., Comparative assessment of protein large J. Am. Chem. Soc., 2023, 145(51), 28284–28295.
language models for enzyme commission number 41 [Link]
prediction, BMC Bioinf., 2025, 26, 68. 42 Y. Fang et al., Mol-instructions: A large-scale biomolecular
24 N. Brandes, D. Ofer, Y. Peleg, N. Rappoport and instruction dataset for large language models, arXiv, 2024,
M. L. Proteinbert, a universal deep-learning model of preprint, arXiv:2306.08018, DOI: 10.48550/arXiv.2306.08018.
protein sequence and function, Bioinformatics, 2022, 38(8), 43 B. Yu, F. N. Baker, Z. Chen, X. Ning, and H. Sun, Llasmol:
2102–2110. Advancing large language models for chemistry with
25 A. Rives, et al., Biological structure and function emerge a large-scale, comprehensive, high-quality instruction
from scaling unsupervised learning to 250 million protein tuning dataset, arXiv, 2024, preprint, arXiv:2402.09391,
sequences, Proc. Natl. Acad. Sci. U. S. A., 2021, 118(15), DOI: 10.48550/arXiv.2402.09391.
e2016239118. 44 T. B. Brown et al., Language models are few-shot learners,
26 A. Radford and K. Narasimhan. Improving language arXiv, 2020, preprint, arXiv:2005.14165, DOI: 10.48550/
understanding by generative pre-training. Semantic Scholar, arXiv.2005.14165.
2018. 45 T. Z. Zhao, E. Wallace, S. Feng, D. Klein and S. Singh,
27 A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and Calibrate before use: Improving few-shot performance of
I. Sutskever, Improving language understanding by generative language models, Proceedings of the 38th International
pre-training, Semantic Scholar, 2019. Conference on Machine Learning, 2021, PMLR 139:12697-
28 A. D. White, et al., Assessment of chemistry knowledge in 12706.
large language models that generate code, Digital 46 J. Wei et al., Chain-of-thought prompting elicits reasoning in
Discovery, 2023, 2(2), 368–376. large language models, arXiv, 2023, preprint,
29 J. Wei et al., Emergent abilities of large language models, arXiv:2201.11903, DOI: 10.48550/arXiv.2201.11903.
arXiv, 2022, preprint, arXiv:2206.07682, DOI: 10.48550/ 47 M. Mosbach, T. Pimentel, S. Ravfogel, D. Klakow, and
arXiv.2206.07682. Y. Elazar, Few-shot ne-tuning vs. in-context learning: A

© 2025 The Author(s). Published by the Royal Society of Chemistry Digital Discovery
 View Article Online

Digital Discovery Paper

fair comparison and evaluation, arXiv, 2023, preprint, 53 D. S. Wishart, et al., Pathbank: a comprehensive pathway
arXiv:2305.16938, DOI: 10.48550/arXiv.2305.16938. database for model organisms, Nucleic Acids Res., 2020, 48,
48 Z. Han, C. Gao, J. Liu, J. Zhang, and S. Q. Zhang, Parameter- D470–D478.
efficient ne-tuning for large models: A comprehensive 54 S. Ida, C. Antje and S. Dietmar, Brenda, enzyme data and
survey, arXiv, 2024, preprint, arXiv:2403.14608, DOI: metabolic information—pubmed, Nucleic Acids Res., 2002,
10.48550/arXiv.2403.14608. 30, 47–49.
49 L. Xu, H. Xie, S.-Z. J. Qin, X. Tao, and F. L. Wang, Parameter- 55 E. J. Hu et al., Low-rank adaptation of large language models,
efficient ne-tuning methods for pretrained language arXiv, 2021, preprint, arXiv:2106.09685, DOI: 10.48550/
models: A critical review and assessment, arXiv, 2023, arXiv.2106.09685.
This article is licensed under a Creative Commons Attribution 3.0 Unported Licence.

preprint, arXiv:2312.12148, doi: DOI: 10.48550/ 56 T. Dettmers, A. Pagnoni, A. Holtzman, and L. Zettlemoyer,
Open Access Article. Published on 30 October 2025. Downloaded on 11/1/2025 7:00:37 AM.

arXiv.2312.12148. Qlora: Efficient netuning of quantized llms, arXiv, 2023,

50 D. Weininger, Smiles, a chemical language and information preprint, arXiv:2305.14314, DOI: 10.48550/arXiv.2305.14314.
system. 1. introduction to methodology and encoding rules, 57 A. Grattaori et al., The llama 3 herd of models, arXiv, 2024,
J. Chem. Inf. Comput. Sci., 1988, 28(1), 31–36. preprint, arXiv:2407.21783, DOI: 10.48550/arXiv.2407.21783.
51 M. Ganter, T. Bernard, S. Moretti, J. Stelling and M. Pagni, 58 Inc. Daylight Chemical Information Systems, Daylight
[Link]: a website and repository for accessing, theory manual: Fingerprints.
analysing and manipulating metabolic networks, 59 T. Chen and C. Guestrin, Xgboost: A scalable tree boosting
Bioinformatics, 2013, 29, 815–816. system, Proceedings of the 22nd ACM SIGKDD, 2016.
52 R. Alcántara, et al., Rhea – a manually curated resource of
biochemical reactions, Nucleic Acids Res., 2012, 40, D754–
D760.

Digital Discovery © 2025 The Author(s). Published by the Royal Society of Chemistry