Professional Documents
Culture Documents
Can Large Language Models Empower Molecular Property Prediction
Can Large Language Models Empower Molecular Property Prediction
Can Large Language Models Empower Molecular Property Prediction
∗
Chen Qian1 , Huayi Tang1 , Zhirui Yang1 , Hong Liang2 , Yong Liu1
1
Renmin University of China 2 Peking University
{qianchen2022,huayitang,yangzhirui,liuyonggsai}@ruc.edu.cn, lho@stu.pku.edu.cn
Abstract
Graph a molecule SMILES c1cc(c(cc1N)[N+](=O)[O-])N
Molecular property prediction has gained sig- …the presence of the nitro groups gives
Caption this molecule… this molecule is known as
nificant attention due to its transformative po- 2-amino-4-nitrophenol and is a common
intermediate used in the synthesis of other
tential in multiple scientific disciplines. Con- organic compounds…
arXiv:2307.07443v1 [cs.LG] 14 Jul 2023
in Figure 1, given the SMILES line of a molecule, rameter updates (Liu et al., 2022a; Lu et al., 2022;
ChatGPT can accurately describe the functional Wu et al., 2022; Wei et al., 2022b). Therefore,
groups, chemical properties, and potential pharma- we attempt to leverage the ICL capability of Chat-
ceutical applications w.r.t. the given molecule. We GPT to assist in molecular classification task by
believe that such textual descriptions are meaning- well-designed prompts, as shown in Figure 2. This
ful for assisting in molecular-related tasks. paradigm makes it much easier to incorporate hu-
However, the application of LLMs in molecu- man knowledge into LLMs by changing the demon-
lar property prediction tasks is still in its primary stration and templates.
stages. In this paper, we move towards this goal
from two perspectives: zero/few-shot molecular Captions as New Representations. With vivid
classification task, and generating new explana- world knowledge and amazing reasoning ability,
tions for molecules with original SMILES. Con- LLMs have been widely applied in various AI do-
cretely, inspired by the astonishing in-context learn- mains (He et al., 2023; Liu et al., 2023). Also, we
ing capabilities (Brown et al., 2020) of LLMs, reckon that LLMs can empower LLMs can greatly
we first prompt ChatGPT to perform in-context contribute to the understanding of molecular prop-
molecular classification. Then, we propose a novel erties. Taking a commonly used dataset in the field
molecular representation called Captions as new of molecular prediction for a toy example, PTC
Representation (CaR), which leverages ChatGPT (Helma et al., 2001) is a collection of chemical
to generate informative and professional textual molecules that reports their carcinogenicity in ro-
analyses for SMILES. Then the textual explanation dents. We conduct a keyword search using terms
can serving as new representation for molecules, such as ‘toxicity’ ‘cancer’, and ‘harmful’ to re-
as illustrated in Figure 1. Comprehensive experi- trieve all explanations generated by ChatGPT for
mental results highlight the remarkable capabilities the originally SMILES-format PTC dataset. Inter-
and tremendous potential of LLMs in molecular estingly, we observed that the majority of these key-
property prediction tasks. We hope this work could words predominantly appeared in entries labeled as
shed new insights in model design of molecular -1. This demonstrates that ChatGPT is capable of
property prediction tasks enpowered by LLMs. providing meaningful and distinctive professional
explanations for the raw SMILES strings, thereby
2 Method benefiting downstream tasks.
In this section, we will elaborate on our prelimi-
nary exploration of how LLMs can serve molecular Towards this end, we propose to leverage Chat-
property prediction tasks. GPT to understand the raw SMILES strings and
Zero/Few-shot Classification. With the continu- generate textual descriptions that encompass var-
ous advancement of LLMs, In-Context Learning ious aspects such as functional groups, chemical
(ICL) (Brown et al., 2020) has emerged as a new properties, pharmaceutical applications, and be-
paradigm for NLP. Using a demonstration context yond. Then, we fine-tune a pre-trained small-scale
that includes several examples written in natural LM (e.g., RoBERTa (Liu et al., 2020)) on various
language templates as input, LLMs can make pre- downstream tasks, such as molecular classification
dictions for unseen input without additional pa- and properties prediction.
Table 1: Testing evaluation results on several benchmark datasets with Random Splitting. For classification task
reporting ACC and ROC-AUC (%, mean ± std), for regression tasks reporting RMSE (mean ± std). ↑ for higher is
better, ↓ contrarily. ‡ denotes the results cited from origin paper. CoR with superior result is highlighted.
GIN 89.47 ± 4.71 58.29 ± 5.88 78.01 ± 1.77 66.19 ± 5.10 92.08 ± 1.11 0.67 ± 0.04 0.79 ± 0.03
ChebyNet 64.21 ± 5.16 61.43 ± 4.29 79.74 ± 1.78 80.68 ± 5.10 91.48 ± 1.50 0.75 ± 0.04 0.85 ± 0.04
D-MPNN‡ - - - 66.40 ± 2.10 90.60 ± 4.30 0.58 ± 0.05 0.55 ± 0.07
ECFP4-MLP 96.84 ± 3.49 85.71 ± 7.67 94.64 ± 3.14 90.19 ± 4.88 95.81 ± 2.09 0.60 ± 0.11 0.60 ± 0.16
SMILEs
Accuracy
molecules
ACC
0.5
ACC
0.5
0.4 0.4
90 deberta
0.3 0.3
deberta(de novo)
0.2 0.2 85
0.1 0.1
0.0 0.0
2 3 5 2 3 5 80
Shots Shots
MUTAG PTC CLINTOX
GIN GCN ECFP ChatGPT
ROC-AUC ↑ RMSE ↓
Method
Sider ClinTox Bace BBBP Esol Lipo
GCN 55.81 ± 2.92 50.32 ± 2.46 76.78 ± 4.74 71.90 ± 5.35 1.09 ± 0.11 0.88 ± 0.03
GIN 58.86 ± 2.57 51.79 ± 5.18 77.05 ± 5.68 75.30 ± 4.66 1.26 ± 0.49 0.88 ± 0.02
GNNs
ChebyNet 60.87 ± 1.68 52.92 ± 9.36 77.31 ± 3.55 73.89 ± 4.95 1.09 ± 0.08 0.89 ± 0.04
InfoGraph‡ 59.20 ± 0.20 75.10 ± 5.00 73.90 ± 2.50 69.20 ± 0.80 - -
G-Motif‡ 60.60 ± 1.10 77.80 ± 2.00 73.40 ± 4.00 66.40 ± 3.40 - -
GraphMVP-C‡ 63.90 ± 1.20 77.50 ± 4.20 81.20 ± 0.90 72.40 ± 1.60 1.03 0.68
Mole-BERT‡ 62.80 ± 1.10 78.90 ± 3.00 80.80 ± 1.40 71.90 ± 1.60 1.02 ± 0.03 0.68 ± 0.02
ECFP4-MLP 64.86 ± 3.45 52.93 ± 5.92 81.58 ± 4.02 73.37 ± 6.05 1.77 ± 0.25 1.03 ± 0.04
SMILEs
$ &