DMPKformer: An Interpretable Multimodal Deep Learning Framework for Reliable ADMET Property Prediction

publication
Published on 27 July 2026

bioRxiv preprint doi: https://doi.org/10.64898/2026.05.28.728612


1. Introduction

ADMET property prediction is an indispensable component of modern drug discovery

pipelines, enabling early identification of liabilities related to pharmacokinetics and

toxicity before experimental validation. Despite substantial progress in computational

chemistry, predictive modeling of ADMET endpoints remains challenging due to the

inherent heterogeneity and complexity of molecular data. Molecules exhibit multi-scale

structure–property relationships, ranging from atom-level connectivity and local

substructures to global physicochemical characteristics, that are not fully captured by any

single data modality.

Recent advances in deep learning for molecular representation have introduced

architectures tailored to specific input formats. Graph neural networks (GNNs) efficiently

encode topological and relational information from molecular graphs, while Transformer[1]based models capture global dependencies across fingerprint features, including large-scale

molecular language models such as MolFormer [1-6]. Similarly, descriptor-based neural

networks can leverage curated physicochemical features such as hydrophobicity, polar

surface area, and molecular weight [7-8]. However, single-modality models are inherently

limited in their representational capacity, as they often overlook cross-modal correlations

and complementary information across feature spaces [9-12].

To address this limitation, we propose a multimodal ADMET prediction framework that

integrates three complementary molecular representations: (1) MACCS structural

fingerprints, which encode fragment-level structural patterns; (2) graph-based molecular

representations, which capture atomic connectivity and local neighborhood interactions;

and (3) physicochemical descriptors derived from RDKit, PaDEL, and Mordred, which

summarize global molecular properties. The modality-specific embeddings are

subsequently integrated through a fusion network that learns complementary structure–

property relationships from these heterogeneous molecular representations.

Beyond predictive performance, an important objective of this work is to improve the

interpretability and reliability of ADMET prediction models. To this end, DMPKformer

incorporates modality-specific attention mechanisms for substructure-level interpretation

and a latent-space out-of-distribution (OOD) confidence estimation framework to identify

regions of reliable model operation. This combination of multimodal representation

learning, interpretability, and reliability-aware prediction distinguishes DMPKformer from

purely performance-driven ADMET benchmarking models.

available under a CC-BY-NC-ND 4.0 International license.

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

bioRxiv preprint doi: https://doi.org/10.64898/2026.05.28.728612; this version posted May 29, 2026. The copyright holder for this preprint

2. Methodology

2.1 Data Sources

All datasets used in this study were obtained from the Therapeutics Data Commons (TDC),

a standardized benchmark suite for machine learning in drug discovery. TDC provides a

diverse collection of datasets spanning multiple ADMET (Absorption, Distribution,

Metabolism, Excretion, and Toxicity) endpoints, along with predefined data splits to

facilitate consistent and reproducible evaluation [13].

The use of these standardized splits ensures consistency with prior benchmarks and enables

fair comparison across different modelling approaches and the distribution of the number

of datapoints across various ADMET endpoints is graphically depicted as below in Fig. 1

Fig. 1 Data distribution of TDC across various ADMET endpoints

2.2 Model Architecture

The proposed architecture consists of three independent modality-specific subnetworks

and a final multimodal fusion network which is graphically depicted as below

available under a CC-BY-NC-ND 4.0 International license.

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

bioRxiv preprint doi: https://doi.org/10.64898/2026.05.28.728612; this version posted May 29, 2026. The copyright holder for this preprint

Fig. 2 Overall architectural design of DMPKformer

2.2.1 MACCS Fingerprint Track Architecture

The MACCS track is designed to extract expressive, task-adaptive representations from

fixed-length binary molecular fingerprints, specifically the MACCS structural keys. Each

molecule is represented as a 166-bit binary vector, where each bit encodes the presence or

absence of a predefined chemical substructure or functional motif. While MACCS

fingerprints are traditionally used with linear or shallow models, their fixed semantics and

compact dimensionality make them well suited for contextual modelling via attention

mechanisms. The architectural details of the MACCS self-attention transformer within the

DMPKformer framework is given in Fig.3

Input Representation and Embedding

The raw MACCS fingerprint, a binary vector of length 166, is first reshaped into a sequence

of length 166 with a single feature channel per position. Each bit position is treated

analogously to a token in a sequence, enabling the model to learn interactions between

structural keys. A learnable linear embedding layer maps each scalar bit value into a

continuous embedding space of dimension ��model. This projection allows the network to

represent both the presence of a structural feature and its contextual role relative to other

features.

To preserve positional identity, since each MACCS bit corresponds to a specific chemical

key, a fixed sinusoidal positional encoding is added to the embedded sequence. This

ensures that the Transformer encoder can distinguish between different fingerprint indices

while remaining invariant to batch ordering.

available under a CC-BY-NC-ND 4.0 International license.

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

bioRxiv preprint doi: https://doi.org/10.64898/2026.05.28.728612; this version posted May 29, 2026. The copyright holder for this preprint

Transformer Encoder for Inter-Bit Dependency Modelling

The embedded fingerprint sequence is processed by a stack of Transformer encoder blocks,

each consisting of a multi-head self-attention module followed by a position-wise feed[1]forward network. Unlike conventional fingerprint models that treat each bit independently

or aggregate them linearly, self-attention enables the MACCS track to model higher-order

dependencies between structural keys. This is particularly important in cheminformatics,

where the predictive relevance of a substructure often depends on the presence or absence

of other motifs elsewhere in the molecule.

Multi-head attention allows the model to attend to multiple interaction patterns in parallel,

capturing diverse relationships such as co-occurrence of functional groups, mutually

exclusive fragments, or long-range structural complementarities. The feed-forward

sublayers further transform these context-enriched representations, while residual

connections and normalization stabilize training and preserve information flow across

layers.

Global Attention Pooling and Feature Attribution

Following the Transformer encoder, a global attention pooling mechanism aggregates the

per-bit representations into a fixed-length molecular vector. This pooling layer learns a

gating function that assigns an importance weight to each fingerprint position, effectively

performing a soft selection over MACCS keys. The gated representations are then summed

to produce a compact, molecule-level descriptor.

In addition to pooling, the attention weights are retained as per-bit importance scores,

providing interpretability by highlighting which structural keys contribute most strongly to

the model’s prediction. This aligns well with the interpretability expectations of

fingerprint-based models in medicinal chemistry.

Latent Embedding and Prediction Head

The pooled molecular representation is passed through a normalization and non-linear

activation stage to produce a latent embedding vector, denoted ��MACCS. This embedding

serves as a task-conditioned fingerprint representation that captures both local structural

information and global context learned through attention. ��MACCS can be used

independently for downstream analysis, similarity comparisons, or as an input to

multimodal fusion architectures.

For prediction, the latent embedding is processed by a dense feed-forward head composed

of one or more fully connected layers with non-linear activation and dropout regularization.

The final output layer produces a single logit, which is passed through a sigmoid activation

to yield a probability for binary ADMET classification tasks. Training is performed using

available under a CC-BY-NC-ND 4.0 International license.

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

bioRxiv preprint doi: https://doi.org/10.64898/2026.05.28.728612; this version posted May 29, 2026. The copyright holder for this preprint

a weighted binary cross-entropy objective to account for class imbalance commonly

observed in ADMET datasets.

Fig. 3 Architectural details of MACCS transformer

2.2.2 Molecular Graph Track Architecture

The Graph Track is designed to learn chemically grounded molecular representations

directly from molecular structure, encoded as graphs derived from SMILES strings. Unlike

fixed fingerprints, graph-based representations preserve explicit atomic connectivity and

local chemical environments, allowing the model to reason about molecular structure in a

manner that is closely aligned with physical chemistry. The architecture details of the

Graph Attention Transformer adapted for molecular property prediction task within the

DMPKformer framework is given in Fig. 4

Graph Construction and Feature Encoding

Each molecule is represented as a labelled graph �� = (��, ��), where nodes ��correspond to

atoms and edges ��correspond to chemical bonds. Atom-level features are constructed

using numerical descriptors that capture intrinsic chemical properties, such as atomic

number, formal charge, degree, hybridization state, valence, aromaticity, and ring

available under a CC-BY-NC-ND 4.0 International license.

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

bioRxiv preprint doi: https://doi.org/10.64898/2026.05.28.728612; this version posted May 29, 2026. The copyright holder for this preprint

membership. These features are encoded as continuous or categorical values rather than

one-hot vectors, enabling compact representations and smoother optimization.

Bond-level features encode the nature of chemical interactions between atoms, including

bond order (single, double, triple), aromaticity, conjugation, and ring status. Together,

these atom and bond descriptors provide a rich, physics-informed representation of

molecular structure while remaining invariant to atom ordering and SMILES

canonicalization.

Graph Attention Network for Local and Contextual Message Passing

The molecular graphs are processed using a Graph Attention Network (GAT), which

performs message passing between neighbouring atoms while learning to weight the

relative importance of different chemical interactions. In each GAT layer, node

embeddings are updated by aggregating features from neighbouring nodes, where the

contribution of each neighbour is modulated by a learned attention coefficient. These

coefficients are computed as a function of both the source and target node representations

(and, optionally, edge features), allowing the network to prioritize chemically relevant

interactions such as polar bonds, aromatic systems, or functional group connectivity.

Multi-head attention is employed to enable the model to capture diverse interaction patterns

simultaneously. Each attention head can focus on different aspects of the local chemical

environment, such as steric effects, electronic properties, or substructure motifs. The

outputs of multiple heads are concatenated or averaged to form the updated node

embeddings, improving representational capacity and robustness.

By stacking multiple GAT layers, the model progressively expands the receptive field of

each atom, allowing information to propagate across larger molecular substructures. This

hierarchical message passing enables the Graph Track to capture both local atomic

environments and higher-order structural motifs such as rings, scaffolds, and functional

group arrangements.

Global Attention Pooling and Molecular Representation

After graph-level message passing, node embeddings are aggregated into a fixed-length

molecular representation using a global attention pooling mechanism. Rather than simply

averaging or summing node features, attention pooling learns a weighting over atoms that

reflects their relative importance to the prediction task. This allows the model to emphasize

chemically salient atoms—such as reactive centres, pharmacophores, or key substituents—

while down-weighting less informative regions of the molecule.

The pooling operation produces a single vector representation for each molecule that

integrates information across the entire graph while retaining interpretability via learned

available under a CC-BY-NC-ND 4.0 International license.

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

bioRxiv preprint doi: https://doi.org/10.64898/2026.05.28.728612; this version posted May 29, 2026. The copyright holder for this preprint

attention weights. These weights can be inspected post hoc to identify which atoms or

substructures most strongly influenced the model’s prediction, supporting mechanistic

interpretation and hypothesis generation.

Prediction Head and Latent Graph Embedding

The pooled molecular graph representation is passed through a dense feed-forward

prediction head composed of fully connected layers with non-linear activations and

regularization. The final layer outputs a single logit, which is transformed via a sigmoid

activation to produce a probability for binary classification tasks such as ADMET property

prediction.

The intermediate pooled graph vector, extracted prior to the final prediction layer, is

denoted as ��Graph . This latent embedding captures a task-conditioned, structure-aware

representation of the molecule that reflects both atomic-level chemistry and global

molecular context. ��Graphcan be used independently for downstream analyses, similarity

comparisons, or integration with other molecular modalities.

Fig. 4 Architectural details of Graph Attention Transformer for molecular property

prediction

2.2.3 Physicochemical Descriptor Track

Physicochemical and topological descriptors are computed from RDKit, PaDEL, and

Mordred, providing high-dimensional continuous features summarizing global molecular

properties. These features are fed into a tanh-activated multilayer perceptron (MLP)

network that captures nonlinear dependencies among descriptors. An embedding vector

available under a CC-BY-NC-ND 4.0 International license.

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

bioRxiv preprint doi: https://doi.org/10.64898/2026.05.28.728612; this version posted May 29, 2026. The copyright holder for this preprint

from an internal dense layer serves as E_Desc, representing the learned descriptor-level

representation. The architectural details of the descriptor track is provided in Fig. 5

Fig. 5 Architectural details of physicochemical descriptor track in the DMPKformer

framework

2.3 Phase-wise Training Strategy

Each subnetwork is independently trained for binary classification using a dataset of

molecules annotated with the target ADMET label. Binary cross-entropy loss is used for

optimization, and L2 regularization and dropout are applied to prevent overfitting. Once

individual models converge, intermediate embeddings E_MACCS, E_Graph, and E_Desc

are extracted for all molecules.

These embeddings form the input to the fusion network, where they are concatenated and

passed through a tanh-activated dense network that models inter-modality interactions. The

final classification layer produces the unified ADMET prediction output. During this

phase, the modality encoders remain frozen to preserve previously learned representations,

and only the fusion network is trained.

In the section that follows we present and discuss the results obtained for the independent

tracks and the combined track and compare them against state of art benchmarks and

demonstrate improvement over state of art.

available under a CC-BY-NC-ND 4.0 International license.

(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made

bioRxiv preprint doi: https://doi.org/10.64898/2026.05.28.728612; this version posted May 29, 2026. The copyright holder for this preprint

3. Results and Discussion

The following section presents the predictive performance of DMPKformer across

multiple TDC benchmarks together with an analysis of the interpretability provided by

the framework. In addition, we introduce a latent-space out-of-distribution (OOD) metric

for TDC tasks that enables reliability-aware confidence estimation and identification of

high-confidence regions of model operation.

3.1 Track-wise results of DMPKformer

The results from the MACCS self-attention transformer, Graph Attention transformer and

the physicochemical descriptor track are presented below for the classifications and

regression tasks in TDC in table 1a, 1b, 2a, 2b & 3a, 3b respectively.

Publication