Target Audience

  • Researchers and practitioners interested in AI-driven biological simulation and Virtual Cell research.
  • Readers interested in AI applications for drug discovery, molecular design, and the prediction of protein and antibody functions.
  • Anyone seeking to understand recent developments in biological foundation models and multimodal AI.

Summary

  • MAMMAL is a foundation model designed to represent diverse types of biological data—including proteins, antibodies, small molecules, gene expression, and numerical information—within a unified language.
  • Historically, MAMMAL represents the transition from models specialized for individual biological entities toward generalist models capable of integrating multiple entities and tasks.
  • More recent studies, such as BioMatrix, which integrates 3D structures and natural language, and MAP, which focuses on predicting cellular responses to previously unseen drugs, have extended this research in different directions.
  • MAMMAL's principal strength is its ability to handle a wide range of biological entities and tasks within a single model. However, limitations remain in representing 3D structures and generalizing to previously unseen conditions.
  • Looking ahead, important research directions include integrating 3D structures, physical and geometric information, biological mechanisms, and generalization capabilities into unified foundation models.

1. Introduction: What Is MAMMAL?

MAMMAL (Molecular Aligned Multi-Modal Architecture and Language) is a foundation model designed to handle heterogeneous biological data—including proteins, antibodies, small molecules, and gene expression—within a single model.

The preprint was released on October 28, 2024, and the peer-reviewed version was published in npj Drug Discovery on May 4, 2026. [1][2]

Therefore, when assessing MAMMAL's novelty, it is more appropriate to compare it with the state of research in 2024 than to treat it as a newly introduced model in 2026.

MAMMAL's historical significance can be summarized as follows:

MAMMAL represents an important stage in the transition from foundation models specialized for individual molecules or proteins toward generalist models that represent heterogeneous biological entities and drug discovery tasks using a common language.

However, MAMMAL is not necessarily the most comprehensive multimodal model by the standards of 2026.

Its multimodal capabilities primarily involve representing protein sequences, antibody sequences, SMILES, gene expression, and numerical values within a shared Transformer architecture.

In contrast, explicit 3D structures of proteins and small molecules, as well as natural-language information from scientific literature, are not fundamental components of its representation.

Recent models such as ESM3 and BioMatrix have extended biological foundation models toward the direct integration of these additional modalities. [3][4]

This article reviews the research developments leading to MAMMAL, examines its technical characteristics, compares it with recent models, and discusses implications for drug discovery and biological simulation.

2. Stage One: Language Models for Individual Biological Domains

A major research trend between approximately 2019 and 2022 was the development of foundation models specialized for individual biological domains.

During this period, separate language models were developed for different biological entities, including proteins, small molecules, and cells.

2.1. Protein Language Models

In protein research, models such as ESM and ProtTrans demonstrated the potential of language modeling approaches.

These studies showed that training language models on large collections of amino acid sequences could produce internal representations containing information about protein evolution, structure, and function.

For example, a representative ESM study used approximately 250 million protein sequences for training. [5]

ProGen further explored protein generation using language models and experimentally validated the functionality of generated proteins.

The key insight was that amino acid sequences could be treated similarly to sentences in natural language.

By learning statistical regularities within these sequences, models could acquire representations that capture biologically meaningful information.

2.2. Small-Molecule Language Models

In small-molecule research, models trained on molecular representations such as SMILES became increasingly prominent.

One representative example is MoLFormer.

MoLFormer learned molecular representations from approximately one billion SMILES strings. [6]

SMILES is a notation system that represents molecular structures as character sequences.

This representation makes it possible to apply language modeling techniques to molecular data.

The research landscape at this stage can be summarized as follows:

  • Proteins → Protein language models
  • Small molecules → Chemical language models

In other words, individual biological and chemical entities were increasingly treated as languages, but their models remained largely separate.

2.3. Cellular Foundation Models

Around the same period, foundation models for cellular data also began to emerge.

In particular, researchers developed models capable of learning cellular states from large-scale gene expression datasets, including single-cell RNA sequencing data.

Representative examples include Geneformer and scGPT.

Geneformer was trained on approximately 30 million single-cell profiles, while scGPT used more than 33 million profiles. [7][8]

These studies demonstrated that large-scale representation learning based on Transformers and language modeling principles could also be applied to cellular states.

By this point, foundation modeling had become established in three major domains:

  • Proteins
  • Small molecules
  • Cells

However, a unified framework for integrating information across these domains had not yet been fully developed.

3. Stage Two: Aligning Different Biological Modalities

Between 2022 and 2024, research increasingly focused on connecting biological data types that had previously been modeled separately.

Major research directions included:

  • Aligning molecules with natural language
  • Aligning molecules with proteins
  • Integrating molecular representations with 3D structures

3.1. Connecting Molecules and Natural Language

MolT5 is a model designed to translate between molecular strings and natural-language descriptions. [9]

MoleculeSTM uses contrastive learning to align molecular structures with textual descriptions.

These studies moved beyond learning molecular sequences in isolation.

Instead, they sought to connect molecular representations with descriptions of molecular properties and behavior.

3.2. Learning Molecular 3D Structures

Uni-Mol explored the direct learning of molecular 3D structures.

This represented an important development because it moved beyond one-dimensional representations such as SMILES toward modeling the spatial arrangement of molecules.

Understanding molecular function and interactions often requires information about how atoms are positioned in three-dimensional space.

As a result, integrating molecular language representations with 3D structural information became an important research direction.

3.3. BioT5 and BioT5+: Toward a Unified Biological Language

BioT5 and BioT5+ are particularly important predecessors of MAMMAL.

BioT5 integrated proteins, small molecules, and natural language within a T5-based architecture.

BioT5+ further strengthened capabilities such as instruction tuning and numerical representation. [10]

Therefore, the fundamental idea of treating heterogeneous biological data as a language within a Transformer architecture was not first introduced by MAMMAL.

To understand MAMMAL's contribution, it is necessary to examine not only its use of a common language, but also how that language is designed and which tasks it supports.

3.4. Tx-LLM: Generalist Modeling for Drug Discovery

Tx-LLM, introduced in 2024, was another important development.

It combined information about small molecules, proteins, nucleic acids, cell lines, and diseases with natural language.

The model addressed 66 types of therapeutic development tasks using 709 datasets within a single generalist LLM. [11]

Tx-LLM is therefore an important contemporary comparison for MAMMAL.

At this stage, the field was moving beyond models that represented individual molecules or cells toward unified architectures capable of supporting diverse tasks across drug discovery.

4. Stage Three: MAMMAL and a Unified Biological Language

What, then, did MAMMAL contribute to this evolving research landscape?

Its distinctive feature is not simply the addition of biological knowledge to a large general-purpose language model.

Instead, MAMMAL can be understood as an approach that designs a domain-specific language (DSL) for biology and drug discovery.

4.1. Representing Biological Entities in a Common Format

MAMMAL represents different biological entities using the following formats:

Biological entity Representation in MAMMAL
Small molecules SMILES
Proteins Amino acid sequences
Antibodies Amino acid sequences
Cells and gene expression Gene-name sequences ranked by expression level
Continuous numerical values Direct projection into the embedding space

Special tokens are used to separate these representations, allowing multiple biological entities to be described within a single structured prompt. [2]

For example, a task involving the prediction of an interaction between a protein and a small molecule can include both the protein sequence and the molecular SMILES representation within the same input.

The important point is not merely that different data types can be concatenated.

Rather, the types of biological entities and their relationships can be expressed through a common syntax.

4.2. Direct Representation of Continuous Numerical Values

One distinctive feature of MAMMAL is its treatment of continuous numerical values.

In conventional language models, numerical values are often converted into text and then tokenized.

MAMMAL instead allows values such as IC50 and binding affinity to be directly projected into the embedding space through learnable projection layers.

This design is particularly relevant to drug discovery, where quantitative measurements play a central role in many prediction tasks.

4.3. Classification, Regression, and Generation Within One Model

MAMMAL uses a Transformer architecture based on T5.

The common model supports different classes of tasks, including:

  • Classification
  • Regression
  • Sequence generation

Its pretraining involved approximately two billion samples.

The pretraining strategy included not only masked language modeling but also tasks involving protein–protein interactions and antibody generation. [2]

Thus, MAMMAL aims to unify both the representations of biological entities and the prediction or generation tasks performed using those representations.

4.4. Conceptual Differences from BioT5 and Tx-LLM

The difference between MAMMAL and related models can be summarized as follows.

BioT5 and Tx-LLM can be viewed as approaches that bring biological information into the existing framework of large language models.

MAMMAL, by contrast, emphasizes the design of a common language or protocol specifically for biological data.

These approaches are not mutually exclusive.

However, MAMMAL places greater emphasis on structured representations of biological entities, numerical values, and interactions than on describing biological information primarily through natural language.

This distinction helps explain why MAMMAL is significant not merely as a multitask model, but as an attempt to establish a unified biological language.

5. What Does "Multimodal" Mean in MAMMAL?

An important consideration when evaluating MAMMAL is the meaning of the term "multimodal."

MAMMAL is indeed multimodal, but much of its biological information is converted into one-dimensional sequences.

Specifically:

  • Proteins → Amino acid sequences
  • Small molecules → SMILES
  • Cells → Gene-name sequences ranked by expression level

This creates two distinct dimensions along which the model can be evaluated.

5.1. Breadth Across Biological Entities

MAMMAL handles many different categories of biological entities within a shared model.

This can be described as breadth across entities or domains.

Drug discovery requires understanding relationships between heterogeneous entities, including:

  • Interactions between small molecules and proteins
  • Binding between antibodies and antigens
  • Effects of drugs on cellular states

MAMMAL's strength lies in providing a common framework for handling such relationships across different biological domains.

5.2. Depth of Representation Within Each Entity

On the other hand, MAMMAL does not directly represent every type of information associated with each biological entity.

For example, proteins are primarily represented by their amino acid sequences rather than explicit 3D structures.

In this respect, MAMMAL differs substantially from models such as ESM3.

ESM3 focuses on proteins as a specific class of biological entities while jointly modeling sequence, structure, and function. [3]

The distinction can therefore be summarized as follows:

MAMMAL emphasizes breadth across different biological entities.

ESM3 emphasizes depth across different modalities within proteins.

Neither approach is universally superior.

For foundation models intended for drug discovery and biological simulation, both breadth across biological entities and depth of representation within each entity are important.

By 2026, research on unified biological foundation models had expanded in several different directions.

The following table summarizes representative models and their relationship to MAMMAL.

Model Publication period Primary research focus Position relative to MAMMAL
MAMMAL 2024 preprint; 2026 peer-reviewed publication Unified language for small molecules, proteins, antibodies, gene expression, and numerical values Strong in breadth across biological entities and tasks, but limited in explicit 3D and unrestricted natural-language representation.
ESM3 2024–2025 Joint modeling of protein sequence, structure, and function Narrower biological scope, but deeper multimodal modeling within proteins. [3]
TxGemma 2025 66 therapeutic development tasks and natural-language interaction Emphasizes LLM-based interaction, reasoning, and general-purpose therapeutic task support. [12]
BioMatrix 2026 preprint Unified representation of molecular and protein sequences, 3D structures, and natural language Extends the common-language direction toward native integration of structural and textual modalities. [4]
ADiT ICLR 2026 Binding affinity prediction using all-atom and 3D structural information More specialized, but strongly grounded in physical and geometric information. [13]
MAP August 2026 Integration of drugs, proteins, mechanistic knowledge, and cellular responses Emphasizes predicting cellular responses to unseen drugs and generalization based on biological mechanisms. [14]
UCE 2026 Universal cellular representations across species Approaches biological foundation modeling from the cellular side, emphasizing cross-species representations. [15]

Among these models, BioMatrix represents one of the closest subsequent research directions to MAMMAL's common-language philosophy.

TxGemma is particularly relevant when comparing generalist models for therapeutic development.

For biological simulation and drug-induced cellular response prediction, MAP provides an especially interesting comparison.

7. BioMatrix: From a Common Language to Integrated 3D Structures

BioMatrix, released in June 2026, provides an important reference point for understanding MAMMAL's historical position. [4]

BioMatrix integrates three major modalities for small molecules and proteins:

  • Sequence
  • Structure
  • Natural language

Its objective is to represent these modalities within a shared discrete token space and model them through a unified next-token prediction framework.

7.1. Technical Characteristics of BioMatrix

BioMatrix seeks to minimize dependence on external encoders and modality-specific output heads.

Instead, it treats multiple modalities as components of a common token-based representation.

The research includes models with 1.7 billion and 4 billion parameters.

Continual pretraining was conducted using 304.4 billion tokens, and evaluation covered 80 downstream tasks. [4]

7.2. How Does BioMatrix Extend the Direction Established by MAMMAL?

MAMMAL demonstrated a framework for representing heterogeneous biological entities within a common language.

BioMatrix takes a further step by treating 3D structures and natural language as native elements of that shared language.

In other words, BioMatrix can be viewed as extending the common-language philosophy toward richer modalities.

7.3. Does BioMatrix Completely Supersede MAMMAL?

Not necessarily.

MAMMAL includes representations for cellular gene expression and a mechanism for directly incorporating continuous numerical values into its embedding space.

BioMatrix, by contrast, focuses primarily on sequences, structures, and natural language associated with small molecules and proteins.

The distinction can be summarized as follows:

MAMMAL: Breadth Across Biological Entities

MAMMAL emphasizes the ability to handle a wide variety of biological entities within a common framework.

BioMatrix: Depth Across Modalities of Molecular Entities

BioMatrix emphasizes richer multimodal representations of molecular entities, including explicit 3D structural information.

This comparison illustrates an important design consideration for future biological foundation models: the balance between the breadth of biological entities and the depth of their representations.

8. MAP: Predicting Cellular Responses to Unseen Drugs

MAP, published in Nature Machine Intelligence on August 26, 2026, represents a different direction in the integration of biological information. [14]

Its central objective is to predict cellular responses to drug treatments.

A particularly important goal is to predict how cells will respond to drugs whose effects have not been measured during training.

This is an important challenge for Virtual Cell research and biological simulation.

8.1. Biological Information Integrated by MAP

MAP constructs a knowledge graph by integrating 14 public data resources.

The graph contains:

  • More than 180,000 drugs
  • Approximately 23,000 genes
  • Approximately 690,000 mechanistic relationships

These resources are used to align molecular structures, protein sequences, and text describing biological mechanisms.

The resulting representations are then combined with a single-cell foundation model to predict cellular responses to drug treatments.

8.2. Zero-Shot Generalization to Unseen Drugs

One of the most important aspects of MAP is its emphasis on generalization beyond conventional random train/test splits.

In particular, the model is evaluated on drugs for which cellular response profiles were not available during training.

The study reports improvements over existing baselines under such zero-shot evaluation conditions. [14]

This illustrates an important shift in research priorities.

Around 2024, an important question was:

Can a single model solve many different biological tasks?

By 2026, another question had become increasingly important:

Can a model generalize to drugs, cells, or biological conditions that were not observed during training?

8.3. The Importance of Learning Biological Mechanisms

MAP does not rely solely on statistical associations between molecular structures and cellular responses.

It also emphasizes the integration of knowledge about relationships between drugs and genes, as well as biological mechanisms of action.

This reflects a broader research direction in which biological knowledge is incorporated to improve predictions under previously unseen conditions.

Whereas MAMMAL emphasizes task integration through a shared representation, MAP places greater emphasis on generalization and the use of mechanistic biological knowledge.

9. ESM3, ADiT, UCE, and TxGemma: Different Directions of Development

Not all research related to MAMMAL aims to build a single, increasingly large biological language model.

Instead, developments between 2024 and 2026 have followed several distinct paths.

9.1. ESM3: Multimodal Modeling Within Proteins

ESM3 jointly models protein sequence, structure, and function. [3]

Compared with MAMMAL, it covers a narrower range of biological entities.

However, it offers a deeper integration of multiple information types within the protein domain.

This represents a direction that prioritizes detailed modeling of a specific biological entity rather than broad integration across many types of entities.

9.2. ADiT: Incorporating Physical and Geometric Information

ADiT investigates binding affinity prediction using all-atom and 3D structural information. [13]

Rather than addressing a broad range of tasks like MAMMAL, it focuses on a more specialized problem.

Its strength lies in directly incorporating atomic-level structural and geometric information relevant to molecular interactions.

This highlights the importance of integrating physical and geometric information alongside language-based representations in drug discovery AI.

9.3. UCE: Cellular Representations Across Species

UCE focuses on learning universal cellular representations that can be applied across biological species. [15]

MAMMAL represents cellular gene expression through gene-name sequences ranked by expression level.

UCE takes a different architectural approach, emphasizing cross-species cellular representation learning.

This direction is important when considering whether representing cells simply as sequences of gene names is sufficient for all biological modeling tasks.

9.4. TxGemma: General-Purpose LLMs for Therapeutic Development

TxGemma addresses 66 types of therapeutic development tasks. [12]

MAMMAL's distinctive feature lies in representing structured biological data through a common language.

TxGemma, by contrast, emphasizes general-purpose LLM capabilities, including natural-language interaction and broad task support.

The difference can therefore be understood as one between biological representation design and integration with the broader capabilities of large language models.

These models are notable because they do not simply scale up MAMMAL.

Instead, they address different limitations and research objectives.

10. How Should We Interpret MAMMAL's Evaluation Results?

MAMMAL reports state-of-the-art (SOTA) performance on 9 of 11 evaluated downstream tasks. [2]

These tasks include:

  • Small-molecule property prediction
  • Drug–target interaction prediction
  • Protein–protein interaction prediction
  • Drug response prediction
  • Antibody design

Achieving strong performance across such diverse tasks is an important result.

However, the findings should be interpreted carefully.

10.1. What Does "9 Out of 11 SOTA" Mean?

The result means that MAMMAL achieved state-of-the-art performance on nine of the eleven evaluated tasks under the conditions and comparisons reported in the study.

It does not mean that MAMMAL outperforms all existing models on every drug discovery task.

Different tasks may use different baselines, data splits, and fine-tuning conditions.

The MAMMAL paper also discusses the scope of its comparisons with models for which published results were available. [2]

10.2. Comparison with AlphaFold 3

The MAMMAL paper also includes a comparison with AlphaFold 3.

However, this comparison requires careful interpretation.

On the MAMMAL side, the task is performed using a fine-tuned binding classifier.

On the AlphaFold 3 side, confidence scores derived from structure predictions are used as proxies for binding likelihood.

Thus, the comparison does not involve two models trained for exactly the same objective under identical conditions.

It would therefore be inappropriate to conclude that MAMMAL is generally superior to AlphaFold 3 in binding prediction based solely on this result. [2]

10.3. Evaluating Performance Under Distribution Shift

As of 2026, there has been growing recognition that conventional benchmark scores are insufficient for evaluating biological foundation models.

Important evaluation conditions include:

  • Distribution shift: Differences between training and deployment data distributions
  • Activity cliffs: Cases where structurally similar compounds exhibit substantially different biological activities
  • Unseen targets: Targets not observed during training
  • Unseen drugs: Drugs not observed during training
  • Uncertainty: The reliability and uncertainty of model predictions

Without examining performance under such conditions, it is difficult to determine how useful a model will be in practical drug discovery.

A Perspective published in Nature Methods on September 4, 2026, also emphasizes challenges in evaluating biomedical foundation models, including reproducibility, generalization, and replicability. [16]

A related 2026 review similarly discusses the importance of realistic data splitting, out-of-distribution evaluation, and uncertainty assessment. [17]

Consequently, MAMMAL should be evaluated not only by the number of tasks it supports or its average benchmark performance, but also by its ability to generalize to previously unseen biological conditions.

11. MAMMAL's Position in the History of Biological Foundation Models

The developments discussed so far can be organized into five broad stages.

11.1. Stage One: 2019–2022 — Language Modeling of Individual Domains

Representative models:

  • ESM
  • ProtTrans
  • MoLFormer

During this period, researchers developed language models specialized for individual biological entities, such as proteins and small molecules.

11.2. Stage Two: 2022–2024 — Alignment Across Modalities

Representative models:

  • MolT5
  • MoleculeSTM
  • Uni-Mol
  • BioT5

During this period, research increasingly focused on connecting molecules, proteins, natural language, and 3D structures.

11.3. Stage Three: 2023–2024 — Generalist Biological Models

Representative models:

  • Geneformer
  • scGPT
  • Tx-LLM
  • MAMMAL

This stage emphasized the integration of multiple biological entities and tasks within common modeling frameworks.

MAMMAL particularly demonstrated the value of representing diverse drug discovery entities through a common structured language.

11.4. Stage Four: 2024–2026 — Toward Native Multimodality

Representative models:

  • ESM3
  • BioMatrix

During this stage, researchers increasingly integrated 3D structure, function, and natural language directly into biological foundation models.

In addition to representing diverse entities through a common language, models began to capture richer information within each biological entity.

11.5. Stage Five: 2026 Onward — Mechanisms, Generalization, and Virtual Cells

Representative models:

  • MAP
  • UCE

This direction emphasizes the ability to generalize to unseen drugs, cells, and biological species rather than merely achieving high benchmark scores.

It also places greater importance on incorporating biological mechanisms and predicting cellular responses to interventions.

Within this five-stage framework, MAMMAL can be viewed as a representative of Stage Three and a bridge toward Stage Four.

12. Implications for Biological Simulation and Drug Discovery

From the perspective of biological simulation and pharmaceutical development, MAMMAL offers several important insights.

Drug discovery involves heterogeneous information about proteins, small molecules, antibodies, and cells.

A unified framework for representing these entities is an important step toward integrating different stages of the drug discovery process.

MAMMAL demonstrates the potential of this approach.

At the same time, developments through 2026 indicate that unified representations alone may not be sufficient.

12.1. The Need for Richer Structural Information

BioMatrix and ESM3 integrate structural information alongside sequence representations.

ADiT and related approaches directly incorporate atomic-level geometric information.

These developments suggest that, beyond common sequence representations, biological foundation models may need to account for physical structures and molecular geometry.

12.2. Biological Mechanisms and Generalization to Unseen Conditions

MAP integrates mechanistic biological knowledge to predict cellular responses to previously unseen drugs.

For biological simulation, it is not enough to accurately predict outcomes under conditions similar to those in the training data.

A critical challenge is whether models can extrapolate to new experimental conditions.

Research on generalization to unseen drugs, targets, and cellular states is therefore likely to remain an important direction.

12.3. The Relationship Between Generalist and Specialist Models

MAMMAL demonstrates the value of generalist models capable of handling a broad range of biological tasks.

However, specialist models can offer important advantages when high-precision structural or physical modeling is required.

The current research landscape is not simply moving toward increasingly large versions of MAMMAL.

Instead, it is expanding in several directions:

  • BioMatrix: Native multimodality integrating 3D structures and natural language
  • MAP: Biological mechanisms and out-of-distribution generalization
  • ADiT: Physics- and geometry-informed modeling
  • TxGemma: Integration of therapeutic tasks with LLM reasoning, interaction, and agent-like capabilities

These developments suggest that future biological simulation and drug discovery systems may require different representations and architectures depending on their objectives.

13. Conclusion: MAMMAL at a Branching Point in Biological AI Research

As of 2026, MAMMAL is not the most advanced model for explicit 3D molecular structures.

Nor is it necessarily the most capable LLM for scientific reasoning.

Nevertheless, MAMMAL made an important contribution by demonstrating how proteins, antibodies, small molecules, cellular states, and continuous numerical values could be represented within a common structured biological language.

Rather than treating each biological entity through an entirely separate model, it proposed a framework for integrating diverse drug discovery tasks.

This gives MAMMAL an important place in the evolution of biological foundation models.

Particularly interesting is how subsequent research has developed in different directions rather than simply scaling up MAMMAL.

BioMatrix extends unified modeling toward native representations of 3D structures and natural language.

MAP incorporates mechanistic biological knowledge and emphasizes cellular response prediction for unseen drugs.

ADiT highlights the importance of physical and geometric information.

TxGemma connects therapeutic development tasks with the general-purpose interaction and reasoning capabilities of large language models.

Viewed in this context, MAMMAL is best understood not as the final form of a biological foundation model, but as a generalist architecture that appeared near a branching point from which several next-generation research directions emerged.

For the future of biological simulation and AI-driven drug discovery, an important question remains:

How can foundation models combine breadth across diverse biological entities with depth in representing 3D structures, physical interactions, and biological mechanisms?

An equally important question is how accurately these models can predict outcomes for drugs, cells, and biological conditions that were not represented in their training data.

MAMMAL's lasting significance lies in demonstrating a common-language approach across diverse drug discovery tasks and helping establish the foundations for addressing these broader challenges.

References

[1] Shoshan, Y. et al., “MAMMAL - Molecular Aligned Multi-Modal Architecture and Language,” arXiv preprint, 2024. https://arxiv.org/abs/2410.22367

[2] Shoshan, Y. et al., “MAMMAL - Molecular Aligned Multi-Modal Architecture and Language for biomedical discovery,” npj Drug Discovery, 3, Article 14, 2026. https://www.nature.com/articles/s44386-026-00047-4

[3] Hayes, T. et al., “Simulating 500 million years of evolution with a language model,” Science, 387, pp. 850–858, 2025. https://pubmed.ncbi.nlm.nih.gov/39818825/

[4] Pei, Q. et al., “BioMatrix: Towards a Comprehensive Biological Foundation Model Spanning the Modality Matrix of Sequences, Structures, and Language,” arXiv preprint, 2026. https://arxiv.org/abs/2606.22138

[5] Rives, A. et al., “Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences,” Proceedings of the National Academy of Sciences, 118, e2016239118, 2021. https://pubmed.ncbi.nlm.nih.gov/33876751/

[6] Ross, J. et al., “Large-scale chemical language representations capture molecular structure and properties,” Nature Machine Intelligence, 4, pp. 1256–1264, 2022. https://www.nature.com/articles/s42256-022-00580-7

[7] Theodoris, C. V. et al., “Transfer learning enables predictions in network biology,” Nature, 618, pp. 616–624, 2023. https://www.nature.com/articles/s41586-023-06139-9

[8] Cui, H. et al., “scGPT: toward building a foundation model for single-cell multi-omics using generative AI,” Nature Methods, 2024. https://www.nature.com/articles/s41592-024-02201-0

[9] Edwards, C. et al., “Translation between Molecules and Natural Language,” Proceedings of EMNLP, 2022. https://aclanthology.org/2022.emnlp-main.26/

[10] Pei, Q. et al., “BioT5: Enriching Cross-modal Integration in Biology with Chemical Knowledge and Natural Language Associations,” Proceedings of EMNLP, 2023. https://aclanthology.org/2023.emnlp-main.70/

[11] “Tx-LLM: A Large Language Model for Therapeutics,” arXiv preprint, 2024. https://arxiv.org/abs/2406.06316

[12] “TxGemma: Efficient and Agentic LLMs for Therapeutics,” arXiv preprint, 2025. https://arxiv.org/abs/2504.06196

[13] “ADiT,” International Conference on Learning Representations (ICLR), 2026. https://proceedings.iclr.cc/paper_files/paper/2026/hash/5fe1b43c882d746c187456eb4c8cdf52-Abstract-Conference.html

[14] Feng, J. et al., “A knowledge-driven framework for predicting single-cell responses for unprofiled drugs,” Nature Machine Intelligence, 8, pp. 1478–1491, 2026. https://www.nature.com/articles/s42256-026-01286-w

[15] “Universal Cell Embeddings (UCE),” Nature, 2026. https://www.nature.com/articles/s41586-026-10689-z

[16] Saez-Rodriguez, J. et al., “Benchmarking biomedical foundation models,” Nature Methods, 23, pp. 1724–1733, 2026. https://www.nature.com/articles/s41592-026-03182-y

[17] “Review on foundation and multimodal models in drug discovery,” 2026. https://pmc.ncbi.nlm.nih.gov/articles/PMC13014059/