MAMMAL and the Evolution of Biological Foundation Models: A Unified Language, Hidden Assumptions, and the Road to Virtual Cells

Intended Audience#

• Researchers and practitioners interested in AI-driven biological simulation and Virtual Cells.
• Those exploring AI for drug discovery, molecular design, and protein and antibody function prediction.
• Readers seeking a systematic understanding of the history, recent advances, and unresolved challenges of biological foundation models.

Summary#

• MAMMAL is a biological foundation model that represents heterogeneous data—including proteins, antibodies, small molecules, gene expression, and continuous numerical values—in a unified format, enabling classification, regression, and sequence generation within a shared architecture.
• Historically, MAMMAL represents an important transition from models specialized for individual biological entities toward Generalist Models that integrate multiple entities and tasks. However, it was not the first model to propose a shared language for biological information.
• MAMMAL reports state-of-the-art performance on 9 of 11 benchmarks. Nevertheless, solving diverse tasks through a common representation should not be confused with understanding the biological mechanisms underlying those tasks.
• By 2026, research has expanded in several directions, including BioMatrix's integration of 3D structures and natural language, MAP's emphasis on predicting cellular responses to unseen drugs, and ADiT's use of physical and geometric information.
• For Virtual Cells, integrating biological entities and tasks is only part of the challenge. The critical questions concern the sufficiency of representations, cross-domain knowledge transfer, generalization to unseen conditions, causal relationships, and temporal dynamics.

1. Introduction: Is MAMMAL a Unified Model of Biology?#

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

Its 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, MAMMAL's novelty should primarily be evaluated against the research landscape of 2024 rather than treating it as a model first introduced in 2026.

Its historical significance can be summarized as follows:

MAMMAL represents an important step in the transition from foundation models specialized for individual biological entities toward Generalist Models that process heterogeneous biological entities and drug discovery tasks through a shared representation.

MAMMAL was pretrained on approximately 2 billion samples and reportedly achieved state-of-the-art (SOTA) performance on 9 of 11 downstream tasks. [2]

Demonstrating strong performance across such diverse tasks is a significant achievement.

However, it raises a fundamental question:

Does the ability to process different biological entities using a common model imply that the model has developed a unified understanding of biological phenomena?

Suppose a model can predict protein binding, generate antibody sequences, and estimate cellular sensitivity to drugs.

These capabilities demonstrate that the model can solve multiple prediction problems.

They do not necessarily prove that the model understands protein conformational changes, molecular interactions, intracellular signaling, and the temporal evolution of cellular states as a coherent biological mechanism.

This distinction is especially important when considering AI-driven drug discovery, Virtual Cells, and biological simulation.

This article first reviews the research developments leading to MAMMAL and examines its technical contributions.

It then investigates the implicit assumptions underlying its design and identifies capabilities that have not yet been sufficiently validated.

Finally, it compares MAMMAL with related research available in 2026 and discusses what next-generation biological foundation models may require.

2. Stage One: Language Models for Individual Biological Domains#

2.1. Protein Language Models#

Around 2019–2022, biological foundation model research primarily advanced through models specialized for particular biological domains.

In protein research, representative examples included ESM and ProtTrans.

These studies demonstrated that language models trained on large collections of amino acid sequences could learn internal representations containing information relevant to protein evolution, structure, and function.

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

ProGen also explored protein generation using language models, including experimental validation of the functions of generated proteins.

The important insight was that amino acid sequences could be treated similarly to sentences in natural language, allowing models to learn biologically meaningful information from statistical regularities in sequences.

However, learning statistical regularities in protein sequences is not equivalent to directly simulating the physical behavior of proteins.

This distinction would remain important in subsequent research.

2.2. Small-Molecule Language Models#

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

A representative example is MoLFormer.

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

SMILES is a notation for representing molecular connectivity and related structural information as strings.

This representation makes it possible to apply Transformer-based techniques originally developed for natural language processing to molecular learning.

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

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

The emphasis was on learning each type of biological or chemical entity as a language.

However, these models were generally developed independently, rather than being designed primarily to represent relationships between heterogeneous biological entities within a single system.

2.3. Cellular Foundation Models#

During a similar period, foundation models for cellular data also advanced.

In particular, models emerged that learned cellular states from large-scale gene expression datasets, including single-cell RNA sequencing (single-cell RNA-seq).

Representative examples include Geneformer and scGPT.

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

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

However, learning gene expression patterns does not mean that a model can reconstruct every aspect of a cell's state.

Cellular behavior also depends on protein activity, post-translational modifications, metabolic conditions, neighboring cells, and the extracellular environment.

By this stage, foundation models had become established across three major areas: proteins, small molecules, and cells.

Nevertheless, a sufficiently general framework for integrating these different types of biological information had not yet emerged.

3. Stage Two: Connecting Different Biological Modalities#

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

Three major directions emerged:

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

3.1. Integrating Molecules and Natural Language#

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

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

These approaches moved beyond treating molecules solely as strings.

Instead, they attempted to model relationships between molecular structures, their properties, and natural-language descriptions.

For example, a model might connect the structure of a compound with text describing its biological activity or chemical characteristics.

3.2. Learning Molecular 3D Structures#

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

This was important because it extended molecular representation beyond one-dimensional strings such as SMILES to the spatial arrangement of atoms.

Molecular function and interactions are not determined solely by chemical connectivity.

The three-dimensional configuration of atoms, molecular conformations, and spatial relationships with interacting partners can also play crucial roles.

Consequently, connecting molecular language representations with explicit 3D structural information became an important research direction.

3.3. BioT5 and BioT5+: Toward a Common Biological Language#

BioT5 and BioT5+ are particularly important predecessors to MAMMAL.

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

BioT5+ further strengthened instruction tuning and numerical representations. [10]

This means that the basic idea of treating heterogeneous biological entities as a shared language within a Transformer architecture did not originate with MAMMAL.

To understand MAMMAL's contribution, it is necessary to examine not only the integration of heterogeneous data but also the design of its shared language, the tasks it supports, its treatment of numerical information, and its training strategy.

3.4. Tx-LLM: Generalist Models for Drug Discovery#

Tx-LLM also appeared in 2024.

It combines information about small molecules, proteins, nucleic acids, cell lines, diseases, and other therapeutic entities with natural language.

The model addresses 66 therapeutic development tasks using a single Generalist LLM.

The study used 709 datasets. [11]

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

By this stage, biological AI research was moving beyond learning individual molecular or cellular representations toward performing diverse drug discovery tasks within shared models.

4. Stage Three: MAMMAL and a Unified Biological Language#

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

Its distinctive approach was not simply to add biological knowledge to a large general-purpose language model.

Rather, it was closer to designing a common Domain-Specific Language (DSL) specifically suited to biological data and drug discovery tasks.

4.1. Converting Heterogeneous Biological Entities into a Common Format#

MAMMAL represents different biological entities in the following ways:

Biological entity	Primary representation in MAMMAL
Small molecules	SMILES
Proteins	Amino acid sequences
Antibodies	Amino acid sequences
Cells and gene expression	Gene names ordered by expression levels
Continuous numerical values	Direct projection into the embedding space

Special tokens delimit these representations, allowing multiple biological entities to be expressed in a single Structured Prompt. [2]

For example, when predicting interactions between a protein and a small molecule, the amino acid sequence and the SMILES representation can be placed in the same structured input.

The important point is not simply concatenating heterogeneous data.

MAMMAL provides a structured format for specifying the types of biological data and the relationships between them.

This enables multiple drug discovery tasks to share a common interface.

4.2. Directly Representing Continuous Numerical Values#

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

General-purpose language models often process numbers by converting them into text tokens.

In contrast, MAMMAL can directly incorporate continuous values, such as IC50 measurements and binding affinities, into its embedding space through learned projection layers.

This design aims to treat numerical measurements as quantitative information rather than merely as strings.

Drug discovery relies heavily on continuous measurements, including binding affinity, drug sensitivity, and molecular properties.

Consequently, this capability is practically significant.

4.3. Unifying Classification, Regression, and Sequence Generation#

MAMMAL uses a Transformer architecture based on T5.

The shared model supports multiple task types:

• Classification
• Regression
• Sequence generation

Its pretraining dataset contains approximately 2 billion samples.

Training includes not only masked language modeling but also tasks related to protein–protein interactions and antibody generation. [2]

MAMMAL therefore goes beyond standardizing the representation of biological inputs.

It attempts to integrate heterogeneous biological data and different prediction or generation tasks within a shared model architecture.

4.4. How MAMMAL Differs from BioT5 and Tx-LLM#

MAMMAL, BioT5, and Tx-LLM all seek to integrate heterogeneous biological information.

However, their design priorities differ.

BioT5 and Tx-LLM can be understood as approaches that use natural-language-model architectures to incorporate biological information and therapeutic tasks.

MAMMAL, by contrast, places greater emphasis on explicitly structured representations of biological entities, numerical values, interactions, and tasks.

In other words, BioT5 and Tx-LLM emphasize integration with natural language and general-purpose LLM capabilities, while MAMMAL emphasizes a common protocol designed for biological data.

These approaches are not fundamentally incompatible.

MAMMAL's contribution is better characterized not as inventing the idea of a shared biological language, but as providing a structured, drug-discovery-oriented framework that unifies diverse biological entities and tasks.

5. What Does "Multimodal" Mean in MAMMAL?#

An important question when evaluating MAMMAL is what the term multimodal actually means.

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

For example:

• Proteins → Amino acid sequences
• Small molecules → SMILES
• Cells → Gene names ranked by expression levels

Its multimodal capabilities should therefore be evaluated from at least two perspectives.

5.1. Breadth Across Biological Entities#

MAMMAL can process different biological entities within a shared model.

This represents breadth across entities or domains.

Drug discovery requires understanding relationships among multiple types of biological systems, such as interactions between compounds and proteins, antibody–antigen binding, and cellular responses to drugs.

MAMMAL's strength lies in its ability to operate across these heterogeneous entities.

5.2. Depth Across Modalities#

However, MAMMAL does not explicitly represent every type of information associated with each biological entity.

For proteins, for example, amino acid sequences are a primary representation, while explicit 3D coordinates are not a native input modality.

This differs from the direction pursued by ESM3.

ESM3 integrates protein sequence, structure, and function within a unified model. [3]

The distinction can be summarized as follows:

MAMMAL emphasizes breadth across biological entities and tasks.

ESM3 emphasizes depth across modalities within proteins.

Neither direction is universally superior.

However, this comparison raises a more fundamental question:

When biological information is converted into a shared sequential representation, how much of the information required for accurate prediction is preserved?

This is central to understanding MAMMAL's technical limitations.

6. The Implicit Assumptions Behind MAMMAL#

MAMMAL demonstrates strong performance across numerous benchmarks.

However, its architecture relies on several important assumptions.

Here, implicit assumptions do not necessarily refer to claims explicitly made by the authors.

Instead, they refer to hypotheses suggested by the model's design and the properties that would need to hold for its broader ambitions to be realized.

6.1. Assumption 1: Biological Information Can Be Adequately Represented as One-Dimensional Sequences#

The first assumption is that converting heterogeneous biological entities into sequences preserves enough information for many downstream prediction tasks.

However, one-dimensional representations do not explicitly include every biologically relevant property.

Entity	Main representation	Important information not explicitly represented
Proteins	Amino acid sequences	3D structures, conformational dynamics, post-translational modifications, environmental dependencies
Small molecules	SMILES	Conformers, detailed spatial arrangements, solvent-dependent states
Cells	Ranked gene expression	Detailed expression magnitudes, protein activity, spatial organization, cellular history
Molecular interactions	Combinations of sequences	Binding poses, thermodynamics, kinetics, environmental conditions

Of course, protein structure and function can often be inferred from amino acid sequences.

Similarly, SMILES can encode molecular connectivity and, in some cases, stereochemical information.

However, information that can be inferred from a sequence is not necessarily identical to information explicitly supplied to a model.

Even proteins with identical amino acid sequences may adopt different conformations or exhibit different functions depending on their environment and interaction partners.

Likewise, two cells with similar gene expression rankings may differ in absolute expression levels, protein phosphorylation states, or metabolic conditions.

Such differences may affect their responses to drugs.

Therefore, the fundamental issue is not merely whether adding 3D information improves accuracy.

It is whether the available input information is sufficient to determine the biological phenomenon being predicted.

If essential information is missing from the input, increasing model size cannot necessarily recover it.

6.2. Assumption 2: Different Biological Domains Share Transferable Regularities#

MAMMAL uses a shared Transformer to learn from proteins, compounds, gene expression, and other biological data.

This design reflects the expectation that joint learning across different domains can produce representations that are useful across tasks.

This assumption is closely related to multitask learning and transfer learning.

However, training multiple domains within a common model is not equivalent to demonstrating that biologically meaningful shared representations have emerged.

Several possibilities must be distinguished:

• Genuine knowledge transfer occurs between biological domains.
• Different domains learn largely independent representations within the same architecture.
• Improvements arise from general regularization effects associated with large-scale pretraining.
• Some tasks interfere with others, resulting in negative transfer.

MAMMAL's benchmark performance supports the practical usefulness of its integrated architecture.

However, this alone does not establish that the model has discovered shared biological principles.

More controlled experiments would be needed, including comparisons between single-domain and multidomain training, removal of specific modalities, and restrictions on parameter sharing.

6.3. Assumption 3: Joint Learning Across Tasks Is Beneficial#

MAMMAL handles classification, regression, and sequence generation within a common framework.

However, the information required for these tasks is not necessarily identical.

For example, the features useful for classifying molecular binding may differ from those needed to generate novel antibody sequences.

Joint training may improve learning by reinforcing common features.

It may also cause interference between tasks.

Therefore, overall benchmark performance is not sufficient to establish the benefits of multitask learning.

It is also important to determine which training tasks contribute to improvements in which downstream tasks.

6.4. Assumption 4: Relationships Learned from Observational Data Generalize to Unseen Conditions#

This is particularly important in drug discovery.

Many MAMMAL tasks rely on existing experimental measurements as training labels.

However, drug discovery datasets contain multiple sources of selection bias.

For example, compounds selected for experimental testing are not random samples from the entire chemical space.

They are chosen based on synthetic accessibility, prior knowledge, research hypotheses, and laboratory capabilities.

Similar biases exist in the selection of proteins, cell lines, and experimental systems.

As a result, a model may achieve strong predictive performance by learning dataset-specific statistical patterns rather than broadly applicable biological relationships.

Evaluating this assumption requires more than random train/test splits.

It requires testing on unseen chemical scaffolds, targets, cells, experimental facilities, and other shifted conditions.

6.5. Assumption 5: Static Molecular and Cellular Information Is Sufficient to Predict Biological Function#

MAMMAL uses molecular sequences and gene expression information to predict particular properties or experimental outcomes.

However, biological phenomena are inherently dynamic.

Even when two cells begin in similar states, their responses may differ depending on drug dose, exposure duration, neighboring cells, and prior stimulation history.

Therefore, predicting drug sensitivity from a static cellular representation differs fundamentally from predicting how that cell's state evolves over time.

This distinction becomes especially important when considering Virtual Cells.

6.6. Assumption 6: Large-Scale Pretraining Produces Generalizable Biological Knowledge#

MAMMAL was pretrained on approximately 2 billion samples. [2]

This scale is substantial, but the number of samples does not necessarily correspond to the amount of independent biological information represented in the training set.

Large datasets may contain highly similar sequences, repeated experimental measurements, or data concentrated under particular conditions.

Thus, sample count and information diversity are not equivalent.

Even if larger datasets improve benchmark performance, those improvements do not necessarily translate into generalization to unseen conditions or understanding of causal relationships.

Evaluating the value of large-scale pretraining therefore requires examining not only dataset size but also diversity, independence, pretraining task composition, and downstream transfer performance.

6.7. Assumption 7: Strong Performance on Individual Tasks Translates into Practical Drug Discovery Value#

MAMMAL's major achievement is its performance across diverse downstream tasks.

However, successful drug discovery requires more than strong performance on individual prediction tasks.

A compound with high predicted binding affinity may still fail because of poor selectivity, toxicity, pharmacokinetics, or manufacturability.

Therefore, predictive accuracy on individual benchmarks and the effectiveness of decisions across an entire drug discovery workflow must be evaluated separately.

This is not a limitation unique to MAMMAL.

It is a broader challenge for AI-based drug discovery.

7. What Has MAMMAL Actually Demonstrated?#

MAMMAL reports SOTA performance on 9 of 11 downstream tasks and competitive results on the remaining two. [2]

These tasks include:

• Small-molecule property prediction
• Drug–Target Interaction (DTI)
• Protein–Protein Interaction (PPI)
• Prediction of cellular responses to drugs
• Antibody design

This is an important demonstration of MAMMAL's ability to address heterogeneous biological tasks using a common architecture.

Nevertheless, it is essential to interpret each result carefully.

7.1. What Does 9/11 SOTA Mean?#

Achieving SOTA performance on 9 of 11 tasks refers to the benchmarks and comparison models included in the evaluation.

It does not mean that MAMMAL outperforms every existing method across all drug discovery tasks.

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

Furthermore, the evaluation includes task-specific fine-tuning of a shared pretrained model.

Therefore, two claims must be distinguished:

• A common architecture can support 11 different downstream tasks.
• A single model can solve all 11 tasks with high accuracy without additional training.

MAMMAL demonstrates the former, but the reported benchmark results should not automatically be interpreted as demonstrating the latter.

The comparisons are also necessarily limited to models for which relevant results were available. [2]

7.2. Comparison with AlphaFold 3#

MAMMAL's paper also includes a comparison with AlphaFold 3.

In an antibody–antigen binding benchmark, the authors report that fine-tuned MAMMAL prediction scores outperformed the AlphaFold 3 confidence-based approach for five of seven antigen targets. [2]

This is an interesting result, but the experimental conditions require careful interpretation.

MAMMAL was fine-tuned as a classifier for binding prediction.

AlphaFold 3, by contrast, was evaluated using confidence scores derived from structure prediction as proxies for binding likelihood.

The models were therefore not trained for exactly the same objective.

Consequently, the result should not be interpreted as evidence that MAMMAL is generally superior to AlphaFold 3 for binding prediction.

Nor does superior performance on a binary binding classification benchmark establish superiority in protein structure or complex structure prediction.

7.3. Generalization to Unseen Targets and Drugs Has Been Partially Evaluated#

MAMMAL's generalization capabilities should be assessed carefully and fairly.

It would be inaccurate to claim that the study did not evaluate unseen conditions at all.

The original paper includes DTI evaluations in which particular protein classes were withheld from training.

Specifically, the evaluation includes four held-out protein classes. [2]

This is a more challenging evaluation of target generalization than a conventional random split.

The study also conducted experimental validation of cellular drug response predictions using four drugs that were not included in the GDSC training data.

Cell viability was measured after 72 hours.

The reported ranking of drug potency in the experiments matched the model's predicted ranking. [2]

This is important because it extends beyond numerical comparisons on existing benchmarks.

However, these results do not establish broadly reliable quantitative prediction across all unseen chemical compounds or cellular conditions.

In particular, random splits of GDSC1 and GDSC2 may allow the same drugs or cell lines to appear in both training and test sets.

Predicting previously unmeasured combinations of known entities is not the same as predicting combinations of entirely unseen drugs and cells.

7.4. Antibody Sequence Recovery Is Not Equivalent to Experimental Antibody Design#

MAMMAL also evaluates sequence completion for antibody complementarity-determining regions (CDRs).

Improved Amino Acid Recovery (AAR), particularly for CDRH3, is reported. [2]

This demonstrates an ability to learn statistical characteristics of antibody sequences and reconstruct missing portions of known sequences.

However, high AAR does not guarantee that newly generated antibodies will exhibit high binding affinity, selectivity, or stability.

Practical antibody design requires additional evaluation of experimental binding, off-target interactions, expression yield, aggregation, thermal stability, and other developability properties.

7.5. Summary of Demonstrated and Unresolved Capabilities#

Evaluation dimension	What MAMMAL demonstrates	Remaining questions
Diverse downstream tasks	Strong performance across 11 tasks	Adaptation to untested tasks
Shared architecture	Integration of heterogeneous biological entities	Whether biologically meaningful shared knowledge transfers
Unseen-target generalization	Partial validation with held-out protein classes	Broader unseen-target and combined OOD conditions
Unseen-drug generalization	Experimental validation with four drugs	Extrapolation across broad chemical space
Antibody sequence generation	Improved sequence recovery	Experimental function and developability of novel antibodies
Explicit 3D representation	Not a primary native representation	Integration of atomic and structural information
Biological causality	Not sufficiently established	Intervention-based mechanistic validation
Cellular temporal dynamics	Not a central demonstrated capability	Prediction of dynamic cellular states
End-to-end drug discovery	Results on individual prediction tasks	Prospective validation of complete workflows

Taken together, these results make MAMMAL a promising generalist foundation for heterogeneous biological prediction.

However, further validation is needed before describing it as a unified model of biological phenomena.

8. Related Research in 2026: What Is Evolving?#

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

The following table summarizes representative approaches.

Model	Release period	Main direction	Relationship to MAMMAL
MAMMAL	2024 preprint; 2026 peer-reviewed publication	Unified representations for small molecules, proteins, antibodies, gene expression, and numerical values	Emphasizes breadth across entities and tasks
ESM3	2024–2025	Integration of protein sequence, structure, and function	Deeper multimodal integration within proteins [3]
TxGemma	2025	66 therapeutic tasks and natural-language interaction	Emphasizes LLM capabilities, reasoning, interaction, and agentic workflows [12]
BioMatrix	2026 preprint	Integration of molecular and protein sequences, 3D structures, and natural language	Extends native multimodal representation [4]
ADiT	ICLR 2026	All-atom, 3D structure-based binding affinity prediction	Emphasizes physical and geometric representations [13]
MAP	August 2026	Integration of drugs, proteins, mechanistic knowledge, and cellular responses	Emphasizes generalization to unseen drugs [14]
UCE	2026	Universal cellular representations across species	Emphasizes cross-species cellular representation learning [15]

The important point is that these models are not evolving along a single trajectory.

A newer model does not necessarily subsume all capabilities of an earlier model.

The following sections examine several approaches that are particularly relevant to MAMMAL.

9. BioMatrix: From a Shared Language to Native 3D Multimodality#

BioMatrix, released in June 2026, is an important comparison when considering MAMMAL's position in the history of biological foundation models. [4]

9.1. Technical Characteristics of BioMatrix#

BioMatrix integrates three major modalities for small molecules and proteins:

• Sequence
• Structure
• Natural language

It maps these modalities into a shared discrete token space and aims to process them using a unified next-token prediction objective.

BioMatrix is built on Qwen3 and includes configurations with 1.7 billion and 4 billion parameters.

It was continually pretrained on 304.4 billion tokens and evaluated on 80 downstream tasks across six categories.

The study reports SOTA or competitive performance on 77 of these tasks. [4]

An important design goal is to minimize reliance on separate encoders, projection adapters, and modality-specific output heads.

Instead, BioMatrix seeks to read and generate different modalities through a common token-based representation.

9.2. The Fundamental Difference from MAMMAL#

MAMMAL converts different biological entities into a common structured language.

BioMatrix extends the shared-language approach by treating 3D structures and natural language as native tokenized modalities alongside molecular sequences.

This can be understood as an expansion of the common-language approach into richer forms of information.

However, BioMatrix does not necessarily encompass all of MAMMAL's capabilities.

MAMMAL includes representations for gene expression and direct projection of continuous numerical values into an embedding space.

BioMatrix focuses primarily on sequences, structures, and natural language for molecules and proteins.

Their priorities can therefore be summarized as follows:

MAMMAL: Breadth Across Biological Entities

It emphasizes integrating heterogeneous biological entities and tasks.

BioMatrix: Depth Across Modalities of Molecular Entities

It emphasizes richer representations of small molecules and proteins.

9.3. Does Integrating 3D Structures Lead to Mechanistic Understanding?#

An important distinction remains.

Incorporating 3D structures can improve a model's ability to represent molecular interactions.

However, handling 3D structure is not equivalent to accurately modeling biological causality or temporal dynamics.

For example, predicting a static binding structure does not necessarily enable prediction of intracellular signaling, metabolic changes, or long-term drug responses.

BioMatrix therefore represents an important direction in native multimodal modeling, but native multimodality alone does not satisfy every requirement for Virtual Cells.

10. 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 primary objective is to predict cellular responses to drugs.

In particular, it emphasizes drugs for which cellular response measurements were not available during training.

This is a central research challenge for Virtual Cells.

10.1. Integrating Biological Knowledge Through a Knowledge Graph#

MAP constructs a knowledge graph from 14 public information sources.

The resulting graph contains information at approximately the following scale:

• More than 180,000 drugs
• Approximately 23,000 genes
• Approximately 690,000 relationships related to biological mechanisms

These data are used to align molecular structures, protein sequences, and textual information about biological mechanisms.

The learned representations are then combined with a Single-Cell Foundation Model to predict cellular responses to drug treatment. [14]

10.2. Zero-Shot Generalization to Unseen Drugs#

A key feature of MAP is its explicit focus on generalization to unseen drugs.

The study evaluates whether the model can predict cellular responses for drugs whose response measurements were absent during training.

The authors report improvements over existing baselines under these zero-shot conditions. [14]

However, zero-shot does not necessarily mean that every type of information about a drug is unavailable.

Even if cellular response measurements are absent, molecular structures or known biological relationships may still be accessible.

Therefore, any claim of zero-shot generalization should clearly specify which information was withheld and which information remained available.

10.3. The Importance of Biological Mechanistic Knowledge#

MAP emphasizes not only statistical relationships between drug structures and cellular responses but also prior knowledge about drug–gene relationships and mechanisms of action.

This approach seeks to improve prediction under unseen conditions by incorporating biologically meaningful relationships into learned representations.

However, using existing mechanistic knowledge is not equivalent to independently discovering and validating new causal mechanisms.

This distinction remains important.

While MAMMAL primarily emphasizes integrating diverse tasks through a common representation, MAP focuses on cellular response prediction and the use of mechanistic knowledge to improve generalization to unseen drugs.

This illustrates how the evaluation priorities of biological foundation models have diversified between 2024 and 2026.

11. ESM3, ADiT, UCE, and TxGemma: Diverging Research Directions#

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

Between 2024 and 2026, several approaches developed in parallel to address different challenges.

11.1. ESM3: Multimodal Learning Within Proteins#

ESM3 integrates protein sequence, structure, and function. [3]

Compared with MAMMAL, the range of biological entities it addresses is narrower.

However, it integrates multiple forms of information about proteins more deeply.

This represents an approach that prioritizes richer representations within a particular biological domain rather than increasing the number of biological entity types.

11.2. ADiT: Physical and Geometric Information#

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

Unlike MAMMAL, which addresses a broad collection of tasks, ADiT is closer to a specialist model focused on a particular prediction problem.

Molecular interactions depend on atomic distances, spatial configurations, and binding-site geometry.

Directly representing such physical and geometric information offers advantages that differ from those of sequence-centered models.

However, using all-atom representations does not automatically mean that a model fully obeys physical laws.

11.3. UCE: Universal Cellular Representations Across Species#

Universal Cell Embedding (UCE) focuses on learning cellular representations that can be used across different species. [15]

MAMMAL represents cellular gene expression as sequences of gene names ranked by expression levels.

UCE follows a different approach, seeking to place cells in a shared latent representation space across tissues and species.

The UCE study describes an integrated atlas containing 36 million cells, spanning more than 1,000 cell types and eight species. [15]

This represents a route toward generalized biological representations from the cellular perspective.

However, learning a shared representation of cells is not equivalent to predicting how those cells will change after interventions.

11.4. TxGemma: Connecting Therapeutic Tasks with LLM Reasoning and Interaction#

TxGemma addresses 66 therapeutic development tasks. [12]

Whereas MAMMAL emphasizes structured representations of biological data, TxGemma emphasizes general-purpose LLM capabilities and their connections to natural-language interaction, reasoning, and agentic workflows.

These represent two complementary directions.

One focuses on learning biological data using representations designed for the domain.

The other connects biological data and prediction models to language-based reasoning and research workflows.

Both may prove useful in practical drug discovery.

12. A Historical Framework: Progress Is Not a Single Linear Trajectory#

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

12.1. Stage One: 2019–2022 — Language Models for Individual Domains#

Representative models:

• ESM
• ProtTrans
• MoLFormer

Research at this stage focused on learning proteins, small molecules, and other biological entities as separate languages.

12.2. Stage Two: 2022–2024 — Alignment Across Modalities#

Representative models:

• MolT5
• MoleculeSTM
• Uni-Mol
• BioT5
• BioT5+

Research increasingly focused on connecting molecular structures, proteins, natural language, and 3D representations.

12.3. Stage Three: 2023–2024 — Generalization Across Biological Tasks#

Representative models:

• Geneformer
• scGPT
• Tx-LLM
• MAMMAL

During this period, research advanced in generalized cellular representations and models capable of addressing multiple biological entities or therapeutic tasks.

However, Geneformer and scGPT primarily focus on cellular representations and should not be classified as multi-entity Generalist Models in exactly the same sense as MAMMAL or Tx-LLM.

MAMMAL made the use of a Structured Biological Language for heterogeneous drug discovery tasks particularly explicit.

12.4. Stage Four: 2024–2026 — Native Multimodality#

Representative models:

• ESM3
• BioMatrix

This direction emphasizes integrating not only sequences but also 3D structures, functions, and natural language.

In addition to representing more biological entities, models increasingly aim to capture richer information within each entity.

12.5. Stage Five: Important Challenges in 2026 and Beyond — Generalization, Mechanisms, and Virtual Cells#

Related approaches:

• MAP
• UCE
• ADiT

The focus increasingly includes generalization to unseen drugs, cells, species, and experimental conditions.

Predicting biological mechanisms, physical interactions, and dynamic cellular states also becomes increasingly important.

12.6. These Stages Do Not Represent a Simple Hierarchy#

It is important to emphasize that these stages are neither a strict chronology nor a ranking of model quality.

ESM3 and MAMMAL, for example, pursue different objectives.

BioMatrix's later publication does not mean that it includes every capability of MAMMAL.

Similarly, MAP's focus on unseen-drug cellular responses differs from ADiT's emphasis on molecular geometry.

A useful way to understand the evolution of biological foundation models is through two broad axes.

Axis One: Breadth of Biological Entities and Tasks

How many different biological entities and prediction tasks can a model address within a shared framework?

MAMMAL and Tx-LLM are representative examples.

Axis Two: Depth of Biological Representation

How comprehensively can a model represent 3D structures, physical interactions, biological mechanisms, and temporal dynamics?

ESM3, BioMatrix, ADiT, and MAP pursue different aspects of this direction.

MAMMAL is an important contribution along the first axis.

For Virtual Cells, however, both breadth and depth are likely to be necessary.

13. Unresolved Questions: What Should Be Validated Next?#

While recognizing MAMMAL's contributions, it is important to identify the capabilities that remain insufficiently validated.

Three particularly important areas are the value of shared representations, generalization to unseen conditions, and biological causality.

13.1. Generalization to Unseen Drugs, Targets, and Cells#

Generalization to unseen conditions can involve very different levels of difficulty.

For example:

Evaluation setting	Prediction problem	Main challenge
Known drug × Known cell	Unmeasured combinations of familiar entities	Interpolation
Unseen drug × Known cell	Drug without response measurements during training	Drug-cold generalization
Known drug × Unseen cell	Unseen cell types or cell lines	Cell-cold generalization
Unseen drug × Unseen target or cell	Both entities absent from training	Double-cold generalization
Unseen drug × New experimental conditions	Different doses, exposure times, or laboratories	Distribution shift
Unseen combinations of interventions	Combinations of drugs or genetic perturbations	Compositional generalization

MAMMAL evaluates some of these conditions.

However, robust performance across all of them has not been established.

In particular, settings in which both drugs and targets are unseen, temporal data splits, and generalization to new experimental facilities deserve further investigation.

Activity cliffs are another important consideration.

An activity cliff occurs when structurally similar compounds exhibit substantially different biological activities.

Such cases may be difficult for models that primarily learn general structural similarities.

13.2. Does Cross-Domain Knowledge Transfer Actually Occur?#

To validate MAMMAL's central design philosophy, it is necessary to test whether joint learning across different domains produces measurable benefits.

For example, consider comparing the following approaches on a Drug–Target Interaction task:

1. A model with separate encoders for proteins and compounds
2. A shared Transformer pretrained only on the target domain
3. A MAMMAL-style model jointly pretrained across multiple domains and tasks

Ideally, these models should be compared under similar data and computational budgets.

If the third approach consistently outperforms the others, it would support the value of integrated pretraining.

Varying the amount of fine-tuning data could further distinguish benefits in low-data transfer from improvements in final performance under data-rich conditions.

Ablation studies that remove particular pretraining tasks would also help determine which domains contribute knowledge to others.

Although the original MAMMAL paper reports strong performance across diverse benchmarks, a complete assessment of cross-domain knowledge transfer requires a systematic examination of controlled comparisons.

13.3. Has the Model Learned Biological Causality?#

Predictive accuracy and causal understanding are different properties.

For example, a model may correctly predict that Drug A inhibits the growth of Cell B without identifying whether the effect arises from inhibition of the intended target or from an off-target mechanism.

Intervention experiments are therefore important for evaluating mechanistic understanding.

Examples include:

• Predicting drug responses after knocking out a particular gene
• Predicting responses after changing the expression of a target protein
• Predicting combined effects of drugs and genetic perturbations
• Predicting responses in rescue experiments that restore disrupted functions

These evaluations help distinguish models that primarily learn observational associations from models capable of predicting the consequences of interventions.

Nevertheless, successful intervention prediction alone does not prove that the model's internal causal structure is completely correct.

Causal interpretation requires additional care.

13.4. Is a Ranked Gene Expression Representation Sufficient?#

MAMMAL represents cellular gene expression as an ordered sequence of gene names.

This is an effective way to make gene expression compatible with Transformer-based processing.

However, rank-based representations may lose information about expression magnitude.

For example, two cells can have the same gene expression ranking but substantially different absolute expression levels or expression ratios.

If these differences are important for drug response, rankings alone may be insufficient.

Comparisons between rank-only representations and representations preserving quantitative expression values would help clarify the consequences of this compression.

13.5. Can the Model Estimate Its Own Uncertainty?#

In drug discovery, it is important not only to generate accurate predictions but also to know when those predictions may be unreliable.

This is particularly relevant for unseen compounds and targets, where models may produce confidently incorrect outputs.

Evaluation should therefore include:

• Calibration: Do predicted confidence levels correspond to actual correctness?
• Uncertainty estimation: Can the model appropriately quantify predictive uncertainty?
• Abstention: Can the model decline to make predictions under unreliable conditions?

These properties are important for practical deployment in drug discovery workflows.

13.6. Does the Model Create Value in Real Drug Discovery Workflows?#

Finally, performance on individual benchmarks must be connected to practical outcomes.

For example, compounds selected by a model could be synthesized and experimentally evaluated in a prospective study, then compared against candidates selected by conventional methods.

Relevant outcomes would include experimental hit rates, costs, computational requirements, and search efficiency.

This type of evaluation is essential for determining whether predictive improvements translate into meaningful scientific or operational value.

13.7. Priority Research Questions#

Priority	Research question	Suggested evaluation	What it would establish
High	Generalization to unseen conditions	Drug-cold, Target-cold, Double-cold, Temporal Split	Performance beyond observed distributions
High	Value of integrated pretraining	Compare single-domain, joint, and no pretraining	Whether cross-domain transfer occurs
High	Biological causality	Genetic perturbations, drug interventions, rescue experiments	Ability to predict intervention effects
Medium	Sufficiency of representations	Compare sequence-only, 3D-enhanced, and multi-omics inputs	Which information is essential
Medium	Gene expression representation	Compare ranked and quantitative representations	Effects of information compression
Medium	Experimental robustness	Evaluate across laboratories, assays, and cell line families	Sensitivity to batch and environmental effects
Medium	Predictive reliability	Calibration, uncertainty estimation, abstention	Ability to identify unreliable predictions
Medium	Computational efficiency	Compare training and inference costs at equivalent accuracy	Practical value of generalist architectures
Medium	Practical molecular generation	Generate novel antibodies and validate them experimentally	Whether generated sequences are functionally useful

These experiments would be valuable not only for MAMMAL but also for evaluating subsequent generations of biological foundation models.

14. MAMMAL from the Perspective of Virtual Cells#

The discussion becomes particularly interesting when viewed through the lens of Virtual Cells and biological simulation.

Drug discovery prediction models typically map specific inputs to specific outputs.

For example:

• Does Compound A bind to Protein B?
• What is the IC50 of Drug A in Cell B?
• How should a missing region of an antibody sequence be completed?

These are important prediction tasks.

Virtual Cells, however, aim to represent how cellular states change in response to interventions and over time.

14.1. Endpoint Prediction Versus Dynamic Modeling#

Consider the administration of Drug A to Cell B.

An endpoint prediction model might estimate drug sensitivity or cell viability at a particular measurement time.

A dynamic cellular model would instead attempt to predict the sequence of changes following treatment.

For example:

1. The drug interacts with a particular protein.
2. Protein activity or intracellular signaling changes.
3. Gene expression and metabolic states change.
4. Cellular proliferation, differentiation, or death is affected.

Predicting these connected processes requires capabilities that differ from predicting a single endpoint.

14.2. State Transition Models for Virtual Cells#

The dynamic prediction problem underlying a Virtual Cell can be expressed conceptually as:

[ x_{t+\Delta t}=F(x_t,u_t,e_t,\Delta t) ]

where:

• (x_t): Cellular state at time (t)
• (u_t): Intervention, such as drug treatment or genetic perturbation
• (e_t): Extracellular environment and surrounding conditions
• (\Delta t): Prediction time interval
• (F): State transition function

In this formulation, a model predicts the future cellular state from an initial state and intervention conditions.

It may also be possible to use a predicted state as input for subsequent predictions, thereby simulating cellular evolution over time.

MAMMAL's principal demonstrated capabilities do not center on learning such a state transition function.

Instead, they involve predicting specific properties or experimental endpoints from molecular and cellular inputs.

Therefore, describing MAMMAL itself as a Virtual Cell would be inappropriate.

14.3. Four Capabilities Required for Virtual Cells#

Four capabilities are particularly important.

1. Sufficiency of State Representations

Are observed features such as gene expression sufficient to predict future cellular states?

If not, additional information—including protein activity, metabolism, and spatial context—may be necessary.

2. Prediction of Temporal Dynamics

A Virtual Cell must predict not only cellular states at isolated time points but also changes over different time intervals.

Even when short-term predictions are accurate, errors may accumulate during long-term simulation.

3. Generalization to Unseen Interventions

Models must predict cellular responses to drugs, genetic perturbations, and combinations of interventions that were not observed during training.

MAP's emphasis on predicting responses to unseen drugs is relevant to this challenge.

4. Consistency Across Biological Scales

Molecular changes must be connected consistently to cellular and potentially tissue-level outcomes.

For example, a change in target protein activity may alter gene expression and ultimately affect cellular phenotypes.

A useful model should represent these relationships without producing contradictory predictions across levels.

14.4. Could MAMMAL Become a Component of a Virtual Cell?#

MAMMAL is not a complete Virtual Cell.

However, it could potentially serve as a component within a broader Virtual Cell architecture.

For example, it might provide modules for molecular interaction prediction, antibody property estimation, or particular cellular response predictions.

Such a system could combine MAMMAL-like Generalist Models with Specialist Models for 3D structures, physical interactions, and temporal cellular dynamics.

However, connecting multiple models does not automatically produce a coherent biological simulation.

The meaning of each model's inputs and outputs, error propagation, differences in temporal scales, and consistency of state variables would all require validation.

15. What Should the Next Generation of Biological Foundation Models Aim For?#

Comparing MAMMAL with subsequent research suggests at least four major directions for future development.

15.1. Integration Across Biological Entities and Tasks#

This is the direction emphasized by MAMMAL.

Processing proteins, small molecules, antibodies, cells, and related data through shared representations or architectures is important for integrating drug discovery workflows.

However, integration should not become an end in itself.

Its value should be established by determining which prediction tasks or generalization capabilities actually improve through joint learning.

15.2. Integration of 3D Structure, Physics, and Geometry#

BioMatrix, ESM3, and ADiT represent different aspects of this direction.

Combining sequence information with explicit structural and physical representations is important for understanding molecular functions and interactions.

However, not every prediction task necessarily requires 3D information.

Determining which representations are essential for particular biological problems is itself an important research question.

15.3. Biological Mechanisms and Generalization to Unseen Conditions#

MAP represents one direction in this area.

Using knowledge of relationships among drugs, targets, genes, and cellular responses may improve predictions beyond observed experimental conditions.

However, even models incorporating known biological relationships must be evaluated independently for causal predictive capabilities.

15.4. Integration with LLM Reasoning, Interaction, and Agentic Workflows#

TxGemma represents this direction.

Drug discovery involves more than generating numerical predictions.

Researchers also need to formulate hypotheses, compare experimental conditions, coordinate multiple models, and interpret results.

Connecting biological prediction models with natural-language reasoning and research workflows could therefore be valuable.

However, a fluent explanation generated by an LLM should not be confused with a scientifically valid explanation.

15.5. Are Generalist and Specialist Models Competitors?#

MAMMAL demonstrates the potential of Generalist Models to address heterogeneous biological tasks through shared architectures.

Specialist Models may nevertheless retain advantages for problems requiring detailed structural, physical, or temporal representations.

Future research may therefore evolve in more than one direction.

Rather than placing every function in a single increasingly large model, architectures may combine generalist components for information integration with specialist components for particular biological problems.

Ultimately, the important criteria are not whether a system contains one model or several.

They are predictive accuracy, generalization to unseen conditions, computational efficiency, and scientific validity.

16. How Should the Evaluation of Biological Foundation Models Evolve?#

The preceding analysis suggests that evaluating biological foundation models solely through benchmark performance will become increasingly insufficient.

Traditional evaluation has emphasized the number of downstream tasks on which a model achieves strong predictive performance.

This remains important.

However, drug discovery and Virtual Cells require additional evaluation dimensions.

Evaluation dimension	Central question
Task Breadth	How many different tasks can the model address?
Entity Breadth	How many types of biological entities can it represent?
Modality Depth	How comprehensively does it represent sequences, 3D structures, functions, and omics data?
Cross-Domain Transfer	Does joint learning across domains provide measurable benefits?
OOD Generalization	Can the model generalize to unseen drugs, targets, cells, and experimental conditions?
Mechanistic Validity	Can it accurately predict biological mechanisms and intervention effects?
Temporal Prediction	Can it predict the evolution of cellular states over time?
Uncertainty	Can it assess the reliability of its predictions?
Experimental Validity	Can independent experiments confirm its predictions?
Practical Utility	Does it improve real scientific or drug discovery workflows?

A Perspective published in Nature Methods on September 4, 2026, also discusses the importance of reproducibility, generalization, and replicability in evaluating biomedical foundation models. [16]

A related 2026 review highlights realistic data splits, out-of-distribution evaluation, and uncertainty assessment as important considerations. [17]

These considerations suggest that reporting strong average performance is not enough.

It is equally important to understand under which conditions a model succeeds and under which conditions it becomes unreliable.

Such evaluations are essential both for scientific understanding and for practical deployment.

17. Conclusion: The Most Important Question MAMMAL Leaves Behind#

MAMMAL is not the most advanced 3D structure model available in 2026.

Nor is it a model that fully understands biological causality or cellular temporal dynamics.

Nevertheless, it demonstrates an important direction for biological AI: representing proteins, antibodies, small molecules, gene expression, and continuous numerical values through a common Structured Biological Language and applying that representation to diverse prediction and generation tasks.

Its pretraining on approximately 2 billion samples and reported SOTA performance on 9 of 11 downstream tasks provide significant evidence for the practical value of this approach. [2]

Furthermore, the study includes evaluations involving held-out protein classes and a limited set of unseen drugs.

It is therefore not a study restricted entirely to conventional randomly split benchmarks.

However, an important assumption underlies MAMMAL's design:

By converting heterogeneous biological entities into a common computational representation and jointly training on multiple tasks, a model can acquire generalizable predictive capabilities beyond those available from learning each domain separately.

MAMMAL's results provide some support for this hypothesis.

However, they do not automatically establish that:

• Shared biological mechanisms have emerged within its learned representations.
• The model can reliably generalize across broadly unseen drugs, targets, cells, and experimental conditions.
• It adequately represents 3D structures and physical molecular interactions.
• It can causally predict changes in cellular states following interventions.
• It can simulate the temporal evolution of cells.

Each of these capabilities requires its own validation.

Research developments through 2026 show that progress beyond MAMMAL is occurring along multiple paths.

BioMatrix advances the native integration of sequences, structures, and natural language.

MAP emphasizes prediction of cellular responses to unseen drugs using biological mechanistic knowledge.

ADiT focuses on atomic-level structure and geometry.

UCE explores cellular representations that generalize across species.

TxGemma connects therapeutic tasks with natural-language interaction, reasoning, and agentic workflows.

These approaches do not simply replace MAMMAL.

Rather, they extend different aspects of biological representation, prediction, and generalization beyond the breadth of entities and tasks that MAMMAL helped integrate.

MAMMAL is therefore best understood not as a completed unified model of biology, but as an important historical example of a Generalist Architecture that demonstrated the feasibility of unifying heterogeneous biological entities and tasks.

For Virtual Cells and biological simulation, the next challenge is particularly clear:

How can we combine breadth across biological entities with depth in representing 3D structure, physics, biological mechanisms, and temporal dynamics?

An even more important question is how to validate the capabilities that such integration is expected to produce.

The ability to solve many biological prediction tasks through a common architecture is not equivalent to understanding the common biological laws underlying those tasks.

What must a shared representation learn in order to make biologically correct predictions beyond the conditions observed during training?

This is perhaps the most important question emerging from MAMMAL's achievements and limitations.

It is also a central challenge for the next generation of biological foundation models and Virtual Cell research.

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] Rosen, Y. et al., “Universal cell embedding provides a foundation model for cell biology,” Nature, 656, pp. 183–191, 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/