Investigating MAMMAL’s Implicit Assumptions, Technical Limitations, and Unresolved Questions

Target Audience#

• Researchers and professionals interested in AI-driven biological simulation and Virtual Cell research.
• Readers interested in AI-based drug discovery, molecular design, protein and antibody function prediction, and biological foundation models.
• Anyone interested in understanding MAMMAL's historical significance, technical contributions, implicit assumptions, limitations, and open research questions.

Summary#

• MAMMAL is a generalist foundation model that demonstrates how diverse biological entities can be represented in a common format and processed across multiple drug discovery tasks using a single architecture.
• Its design rests on several important implicit assumptions, including the idea that biologically relevant information can be sufficiently preserved in sequence-based representations and that different biological domains and tasks share transferable patterns.
• The original paper reports state-of-the-art (SOTA) performance on 9 of 11 downstream tasks and includes limited evaluations involving previously unseen protein classes and new drugs. However, systematic generalization to unseen conditions, causal understanding of biological mechanisms, and dynamic prediction of cellular states remain insufficiently established.
• Representing diverse biological entities in a common language is not equivalent to learning the underlying principles governing biological phenomena.
• Key future challenges include rigorously validating the benefits of shared representations and integrating 3D structures, physical and biological mechanisms, temporal dynamics, and out-of-distribution generalization.

1. Introduction: What Has MAMMAL Achieved, and What Has It Not?#

MAMMAL (Molecular Aligned Multi-Modal Architecture and Language) is a foundation model designed to handle diverse biological data—including proteins, antibodies, small molecules, gene expression, and continuous numerical values—within a common representation framework.

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]

MAMMAL's primary contribution is demonstrating that different biological entities can be converted into common representations and processed across a broad range of drug discovery tasks using a single architecture.

The original paper reports state-of-the-art (SOTA) performance on 9 of 11 downstream tasks. [2]

This is an important research achievement.

However, a substantial conceptual gap remains between these results and claims such as:

• The model has developed a unified understanding of biological phenomena.
• The model understands the mechanisms of action of previously unseen drugs.
• The model can simulate cellular behavior.

To evaluate MAMMAL accurately, we need to distinguish three different levels of claims.

1.1. What the Paper Demonstrates#

MAMMAL demonstrates that multiple types of biological inputs can be processed by a shared model and that task-specific fine-tuning can achieve strong predictive performance across diverse benchmarks.

These tasks include:

• Small-molecule property prediction
• Drug–Target Interaction (DTI) prediction
• Protein–Protein Interaction (PPI) prediction
• Drug response prediction
• Antibody design

These results support the potential value of using a common architecture rather than relying exclusively on separate models for each biological entity or task.

1.2. What the Architecture Implicitly Assumes#

MAMMAL's design implicitly assumes that biologically essential information can be sufficiently preserved when molecules and cells are represented using amino acid sequences, SMILES strings, ranked gene expression sequences, and related formats.

It also assumes that learning across different tasks and biological domains through shared model parameters provides meaningful benefits.

However, successfully processing diverse tasks within one model does not necessarily imply that the model has learned biologically meaningful shared knowledge.

1.3. What Remains Insufficiently Demonstrated#

Important capabilities requiring further investigation include:

• Systematic generalization to unseen drugs, targets, cells, and experimental conditions
• Understanding of biological causal mechanisms
• Accurate representation of 3D structures and physical interactions
• Prediction of time-dependent cellular changes
• Prediction of responses to combined interventions
• Consistent effectiveness across real-world drug discovery workflows

The current evidence does not establish that MAMMAL lacks these capabilities.

Rather, these capabilities have not yet been sufficiently demonstrated.

Therefore, MAMMAL is more accurately characterized as a generalist foundation model that unifies biological data representations and prediction tasks, rather than a unified model of biology itself.

This distinction matters.

Being able to solve many biological tasks using shared representations is not equivalent to learning the common biological laws underlying those tasks.

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

To understand MAMMAL's historical significance, it is useful to examine the development of biological foundation models along multiple research directions rather than treating progress as a single linear sequence.

At least two major directions can be identified.

2.1. Direction One: Integrating Biological Entities and Tasks#

Between approximately 2019 and 2022, foundation models specialized for individual biological domains advanced rapidly.

Examples include:

• ESM for proteins
• MoLFormer for small molecules
• Geneformer for cellular representations

These models used large-scale domain-specific datasets to learn representations through Transformers and related architectures. [3][4][5]

Subsequently, models such as BioT5, Tx-LLM, and MAMMAL began integrating multiple biological data types and tasks. [1][6][7]

The central question along this research direction is:

"How many different biological entities and tasks can be handled within a common model?"

MAMMAL made a substantial contribution to this direction.

It represents proteins, small molecules, antibodies, and cellular gene expression through structured prompts and supports classification, regression, and sequence generation within a single architecture.

2.2. Direction Two: Representing Biological Structure and Mechanisms#

In parallel, models such as Uni-Mol, ESM3, and RFdiffusion have focused on representing molecular 3D structures, functional properties, and structural constraints. [8][9][10]

Rather than treating biological entities solely as sequences, these approaches seek to incorporate information such as:

• Spatial arrangements of atoms
• Molecular geometry
• Structural constraints
• Protein function
• Structure-guided generation

The central question in this direction is:

"What information must a model represent to accurately predict biological phenomena?"

MAMMAL does not primarily pursue this second direction.

Its inputs are largely based on one-dimensional representations, including amino acid sequences, SMILES strings, and ranked gene expression sequences.

The original paper also identifies the challenge of combining the simplicity of a common sequence-to-sequence representation with the expressive capabilities of specialized models that generate 3D structures. [2]

2.3. Where Does MAMMAL Fit Between These Directions?#

The following table summarizes these research directions.

Research Direction	Representative Models	Main Characteristics
Representation learning for individual biological entities	ESM, MoLFormer, Geneformer	Specialized representation learning for proteins, small molecules, and cells
Integration of biological entities and tasks	BioT5, Tx-LLM, MAMMAL	Common architectures for heterogeneous biological data and tasks
Detailed representation of structures and functions	Uni-Mol, ESM3, RFdiffusion	Emphasis on 3D structures, functional properties, and structural constraints
Integration of richer modalities	BioMatrix	Integration of sequences, structures, and natural language
Biological mechanisms and generalization to unseen conditions	MAP	Knowledge-driven prediction of cellular responses to unseen drugs

BioMatrix and MAP were reported in 2026 and address research problems different from those emphasized by MAMMAL. [11][12]

They should not be interpreted as straightforward replacements or universally superior successors.

MAMMAL emphasizes breadth across biological entities and tasks.

ESM3 and BioMatrix emphasize greater depth in the modalities used to represent particular biological entities.

MAP emphasizes biological mechanism-related knowledge and prediction of cellular responses to previously unseen drugs.

These research directions are therefore complementary rather than simply hierarchical.

This historical perspective leads to a fundamental question:

"In biology, does the ability to make predictions using shared representations mean that a model has learned shared biological principles?"

This question is central to evaluating MAMMAL's implicit assumptions.

3. Seven Implicit Assumptions Underlying MAMMAL#

The following seven assumptions can be inferred from MAMMAL's architecture, training strategy, and evaluation framework.

Not all of them are explicitly stated as claims by the authors.

Some are hypotheses implied by the design choices and should therefore be evaluated independently.

3.1. Assumption 1: One-Dimensional Representations Preserve Sufficient Biological Information#

This is one of the most important assumptions underlying MAMMAL.

MAMMAL primarily represents:

• Proteins as amino acid sequences
• Small molecules as SMILES strings
• Cells as ranked gene expression sequences

Continuous numerical values are incorporated into the embedding space through learnable projection layers. [2]

The following table summarizes these representations and the information they do not explicitly encode.

Biological Entity	Main MAMMAL Representation	Information Not Explicitly Represented
Protein	Amino acid sequence	3D structure, conformational dynamics, post-translational modifications
Small molecule	SMILES	Molecular conformers and explicit 3D geometry; some stereochemical information can be encoded in SMILES
Cell	Ranked gene expression sequence	Detailed expression magnitudes, protein activity, spatial organization, cellular history
Molecular interaction	Combination of sequences	Binding poses, solvent effects, thermodynamic and kinetic states

This design introduces an inductive bias: many biological prediction tasks can be adequately addressed using sequence-based representations.

However, this assumption has important limitations.

For example, a protein's amino acid sequence can provide information for predicting its structure, but its actual conformation and function may depend on environmental conditions and binding partners.

Similarly, two cells with comparable ranked gene expression profiles may differ in protein phosphorylation states or other regulatory properties.

Therefore, the ability to infer 3D structures or cellular functions from sequences does not mean those properties are uniquely determined by the input representations.

The key question is not simply:

"Does adding 3D information improve prediction accuracy?"

A more fundamental question is:

"Does the current input representation contain sufficient information to uniquely determine the biological phenomenon being predicted?"

This question deserves direct experimental investigation.

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

MAMMAL uses shared Transformer parameters to learn from proteins, small molecules, gene expression, and other biological data.

This design reflects the expectation that joint training across biological domains can produce representations that benefit multiple tasks.

However, using a common neural network does not necessarily imply that the network develops biologically meaningful shared latent representations.

Several different explanations are possible:

• Genuine knowledge transfer occurs between biological domains.
• Different domains effectively learn separate internal features within the shared network.
• Performance improvements arise primarily from regularization effects associated with large-scale pretraining.
• Joint training introduces interference between tasks, potentially reducing performance in some domains.

These possibilities cannot be distinguished solely by examining final benchmark scores.

More informative experiments would compare:

• Single-domain training
• Multi-domain training
• Removal of selected modalities
• Restrictions on parameter sharing

Although MAMMAL reports strong benchmark performance, this alone does not establish the biological meaning of its shared representations.

3.3. Assumption 3: Learning Across Tasks Produces Mutual Benefits#

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

Its design implicitly assumes that learning across different tasks can improve the usefulness of shared representations.

However, information that is useful for classification is not necessarily identical to information that is useful for molecular generation.

Consequently, increasing the number of training tasks does not always improve performance.

Negative transfer may occur.

For example, a shared representation might improve one prediction task while reducing performance relative to a specialized model on another task.

To evaluate this assumption, aggregate benchmark scores are insufficient.

The important question is:

"Which tasks transfer useful knowledge to which other tasks, and by how much?"

Controlled experiments that isolate individual transfer relationships are necessary.

3.4. Assumption 4: Correlations Learned from Observational Data Remain Useful Under New Conditions#

This assumption is particularly important in drug discovery.

Many MAMMAL tasks learn from observed molecular properties and experimental measurements.

However, these datasets contain selection biases related to:

• Which compounds were synthesized
• Which targets were measured
• Which cell lines were tested
• Which experimental protocols were used

A model may therefore learn dataset-specific regularities rather than the underlying biological mechanisms.

Strong performance under a random train/test split does not guarantee comparable performance under conditions such as:

• Previously unseen chemical scaffolds
• Previously unseen target proteins
• Different cell types or cell lines
• Different laboratories or experimental procedures
• Distributions substantially different from the training data

These challenges are commonly discussed in terms of out-of-distribution (OOD) generalization.

MAMMAL should therefore be evaluated not only through random splits but also through systematic testing under unseen biological and experimental conditions.

3.5. Assumption 5: Static Molecular and Cellular Representations Are Sufficient for Predicting Biological Functions#

This assumption becomes especially important when considering Virtual Cell research.

MAMMAL primarily uses molecular or cellular information measured at particular points in time.

Real cells, however, are dynamic systems.

Even when two cells exhibit similar initial gene expression states, their subsequent responses may differ depending on:

• Drug dosage
• Exposure duration
• Surrounding cells
• Previous stimulation history

Cellular states also evolve over time after drug administration.

Therefore, a model that predicts a particular endpoint from a static cellular representation is fundamentally different from a model that predicts the temporal evolution of cellular states.

Accurate drug response prediction does not, by itself, demonstrate the ability to simulate cellular state transitions.

3.6. Assumption 6: Large-Scale Pretraining Leads to Useful Biological Knowledge#

MAMMAL uses approximately two billion pretraining samples. [2]

This represents substantial training scale.

However, the number of samples does not directly measure the amount of independent biological information available.

Large datasets may contain:

• Highly similar sequences
• Repeated measurements
• Overlapping biological examples
• Biased experimental conditions

Furthermore, research on single-cell foundation models published in 2026 reported that increasing pretraining dataset size did not always produce continued performance improvements and that clear data scaling laws were not consistently observed. [13]

This does not demonstrate that MAMMAL's pretraining is inefficient.

It does, however, suggest that the assumption "more data leads to deeper biological understanding" requires model- and task-specific validation.

Important considerations include not only dataset size but also:

• Diversity
• Independence
• Measurement quality
• Relevance to downstream tasks

3.7. Assumption 7: Strong Performance Across Many Tasks Implies Broad Practical Utility in Drug Discovery#

MAMMAL reports results obtained through task-specific fine-tuning across multiple benchmarks.

However, successful drug discovery requires satisfying multiple constraints simultaneously.

These include:

• Binding affinity
• Selectivity
• ADMET properties
• Manufacturability
• Clinical safety

Strong predictive performance on each individual task does not necessarily guarantee successful candidate selection when these requirements are combined.

For example, a molecule may exhibit strong binding affinity but remain unsuitable as a drug candidate because of toxicity, poor cellular uptake, or insufficient stability.

Prediction errors from individual models may also accumulate when multiple predictions are combined within a discovery workflow.

Therefore, task-level prediction performance and the overall effectiveness of drug discovery decisions must be evaluated separately.

4. Technical Limitations and Current Validation Status#

The implicit assumptions discussed above raise an important question: which capabilities have MAMMAL's authors actually validated, and which remain open research problems?

An important distinction is that MAMMAL has not completely ignored generalization to unseen conditions.

The original paper includes:

• DTI benchmarks involving held-out protein classes
• Experimental evaluation using four drugs absent from the GDSC training data

These experiments provide some evidence of generalization beyond conventional random train/test splits. [2]

Therefore, describing MAMMAL as having no evaluation under unseen conditions would be inaccurate.

A more appropriate conclusion is that limited evidence exists, but broader and more systematic validation remains necessary.

4.1. Current Validation Status#

Capability	Current Validation Status	Remaining Challenges
Application across diverse tasks	Demonstrated	Adaptation beyond the 11 evaluated tasks
Generalization to unseen targets	Partially demonstrated	Broader target families and settings where both drugs and targets are unseen
Generalization to unseen drugs	Limited experimental evidence	Systematic evaluation across many novel compounds
Direct representation of 3D structures	Outside the primary scope	Integration of atomic and structural information
Prediction of temporal cellular changes	Insufficiently demonstrated	Time-series and multi-timepoint responses
Biological causal mechanisms	Insufficiently demonstrated	Prediction of intervention effects
Virtual Cell capabilities	Not demonstrated	State transitions, combined interventions, environmental dependence
Improvement in real-world drug discovery success rates	Not demonstrated	Prospective evaluation across drug discovery workflows

The following sections examine particularly important unresolved challenges.

4.2. How Far Can MAMMAL Generalize to Unseen Drugs, Targets, and Cells?#

Generalization to unseen conditions is among the most important capabilities to evaluate when assessing MAMMAL's practical value.

The phrase "predicting unseen data" can refer to several fundamentally different situations.

For conceptual clarity, these can be organized into five levels.

Level 1: Known Drugs and Known Cells

The model predicts unmeasured combinations of drugs and cells that have already appeared in the training data.

Because the individual drugs and cells are represented in training, this is generally a relatively easier setting.

Level 2: Unseen Drugs

The model predicts cellular responses to drugs for which no cellular response measurements were available during training.

In this setting, the model must infer responses from molecular structures or other relevant information.

Level 3: Unseen Targets or Cell Types

The model predicts responses involving target families or cell types absent from training.

Memorizing characteristics of previously observed targets or cells may not be sufficient.

Level 4: Unseen Drugs × Unseen Cells × New Conditions

Multiple factors change simultaneously, including:

• Drug identity
• Cell type
• Dosage
• Experimental conditions

The model must combine learned knowledge to predict previously unobserved biological settings.

Level 5: Unseen Combined Interventions and Temporal Dynamics

The model predicts time-dependent cellular responses to combinations of interventions.

Examples include administering two drugs simultaneously or introducing a genetic modification before administering a drug.

These five levels are a conceptual framework for discussing increasing generalization challenges, not an established formal classification.

MAMMAL primarily evaluates Level 1-type settings while providing limited evidence related to Levels 2 and 3.

However, robust performance under Levels 4 and 5 has not been established.

One important consideration is the use of random splits in the GDSC1 and GDSC2 evaluations.

When drug–cell combinations are randomly divided, the same drugs or cells may appear in both the training and evaluation sets.

In contrast, the original paper also evaluates DTI prediction using splits that hold out four protein classes.

This provides a more challenging assessment of generalization to unseen target groups than ordinary random splitting. [2]

Therefore, it would be incorrect to claim that MAMMAL has never been evaluated under OOD conditions.

Nevertheless, additional evaluation would be valuable, particularly using:

• Simultaneously held-out drugs and targets
• Temporal data splits
• New experimental conditions
• Broader collections of unseen compounds

4.3. What Does MAMMAL Actually Learn When Predicting Drug Responses?#

MAMMAL predicts drug sensitivity using gene expression information from cancer cell lines and molecular representations of drugs.

However, the predicted outputs are primarily experimental measurements such as IC50.

The model does not directly reconstruct how intracellular states change and eventually produce those measurements.

Consider the distinction between two questions.

Question A:

"What IC50 value should we expect when drug A is administered to cell B?"

Question B:

"Which molecular pathways change after drug A is administered, how do gene expression patterns subsequently evolve, and how do these changes ultimately lead to cell death?"

Question A concerns endpoint prediction.

Question B concerns mechanistic and dynamic modeling.

In the original paper, the authors conducted 72-hour cell viability experiments involving four new drugs and reported agreement between predicted and experimentally observed rankings of drug potency. [2]

This is an important validation beyond numerical benchmark comparisons.

However, the result does not by itself demonstrate that the model accurately estimates IC50 values for every individual cell line or that it has learned the drugs' mechanisms of action.

4.4. Correct Predictions Do Not Necessarily Imply Correct Biological Mechanisms#

This limitation is not unique to MAMMAL. It is a broader issue affecting statistical prediction models.

Suppose a model correctly predicts that drug A inhibits proliferation in cell B.

That does not necessarily mean the model has correctly identified why the inhibition occurs.

Potential explanations include:

• Inhibition of the intended target
• Off-target effects
• Dependence on the cell's metabolic state

These explanations must be distinguished experimentally.

To establish that a model has learned biological mechanisms, it is important to test whether it can predict the effects of specific interventions rather than merely reproduce observed correlations.

Related studies on single-cell perturbation prediction have reported cases in which generalization to unseen cellular conditions or interventions remains inadequate. [14][15]

This is an important challenge for biological foundation models more broadly, not just for MAMMAL.

4.5. Are Shared Representations Actually Beneficial?#

One of the most direct ways to test MAMMAL's central hypothesis is to distinguish the benefits of a common architecture from the benefits of joint pretraining.

For example, consider three approaches to predicting protein–compound interactions.

1. Train protein and compound models separately and combine their representations at the final prediction stage.
2. Use a shared Transformer architecture but restrict pretraining to the target domain.
3. Use joint pretraining across multiple biological domains and tasks, as in MAMMAL.

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

Performance should also be compared under different amounts of downstream training data.

This could reveal whether MAMMAL's main benefit lies in:

• Transfer learning under limited data
• Final predictive performance with abundant data
• Both

However, based on the main paper examined in this investigation, it cannot be concluded that a complete set of such controlled experiments has already been established.

A detailed audit of supplementary materials and publicly available code would be necessary.

4.6. Molecular Generation Is Not the Same as Designing Experimentally Useful Molecules#

MAMMAL's antibody design benchmark evaluates the recovery of correct amino acids in Complementarity-Determining Regions (CDRs).

This evaluation is based on Amino Acid Recovery (AAR).

The paper reports improvements for CDRH3, providing evidence of useful sequence generation capabilities. [2]

However, high AAR does not necessarily mean that newly generated antibodies possess strong binding affinity, selectivity, or stability.

Additional experiments would be necessary to determine whether generated antibodies:

• Bind to new targets
• Improve binding affinity relative to existing antibodies
• Avoid unintended target binding
• Meet development-related requirements such as expression, aggregation, and thermal stability
• Maintain performance under changes in experimental conditions or antigen structures

The sequence recovery results alone do not establish all of these properties.

5. How Should MAMMAL's Evaluation Results Be Interpreted?#

The original MAMMAL paper reports strong results across multiple benchmarks.

However, the capabilities directly evaluated must be distinguished from broader conclusions that might be inferred from those results.

5.1. What Does SOTA Performance on 9 of 11 Tasks Mean?#

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

This is important evidence that a common architecture can be applied competitively across diverse tasks.

However, the comparisons are limited to models and results available under the respective benchmark settings.

Furthermore, baselines, dataset splits, and fine-tuning conditions vary across tasks.

Therefore, achieving SOTA performance on 9 of 11 tasks does not mean that MAMMAL outperforms every existing model across all areas of drug discovery.

5.2. Comparison with AlphaFold 3#

The MAMMAL paper also includes a comparison with AlphaFold 3. [2][16]

This comparison requires careful interpretation.

On the MAMMAL side, the evaluation uses a fine-tuned binding prediction classifier.

On the AlphaFold 3 side, a confidence score derived from structural prediction is used as a proxy for binding likelihood.

These are not identical modeling objectives.

Consequently, stronger performance by MAMMAL on a specific binary classification task does not imply that MAMMAL generally outperforms AlphaFold 3 in predicting structures or biomolecular complexes.

5.3. What Can and Cannot Be Concluded from the Evaluation Results?#

Evaluation Result	What It Supports	What It Does Not Establish
SOTA on 9 of 11 tasks	Competitiveness on selected established benchmarks	Best performance across all drug discovery tasks
Comparison with AlphaFold 3	A fine-tuned MAMMAL classifier can outperform an AF3 confidence proxy on a specific binding classification task	Superiority over AF3 in structure or complex prediction
Improved AAR for antibody CDR generation	Strong recovery of amino acids in known sequence contexts	Guaranteed binding or drug development suitability of new antibodies
Drug response prediction on GDSC	Prediction of drug sensitivity in specific cell line and drug datasets	Prediction of clinical responses in unseen patients
Experiments with four new drugs	Reproduction of potency rankings in a limited drug set	Established quantitative accuracy across novel compounds generally
Successful common architecture	Multiple data types and tasks can be processed by one model	Formation of a unified representation of biological mechanisms

The distinction between demonstrated benchmark performance and broader biological understanding is especially important.

A 2026 Nature Methods perspective on benchmarking biomedical foundation models also discusses the importance of reproducibility, generalization, and replicability. [17]

MAMMAL should therefore be assessed not only by average benchmark performance or the number of SOTA results but also by:

• Performance under unseen conditions
• Reproducibility
• Experimental validity
• Practical usefulness

6. What Should Be Validated Next?#

Based on the preceding analysis, several experimental directions are especially important for evaluating MAMMAL's central hypothesis.

6.1. Prioritizing Future Experiments#

Priority	Research Question	Recommended Experiment	What It Would Reveal
P0	Generalization to unseen conditions	Drug-cold, Target-cold, Double-cold, Temporal Split	Whether predictions extend beyond memorized training distributions
P0	Value of integrated pretraining	Compare single-domain, multi-domain, and no-pretraining approaches	Whether meaningful cross-domain knowledge transfer occurs
P0	Biological causality	Gene knockout, drug intervention, rescue experiments	Whether the model captures intervention effects rather than correlations
P1	Information loss in representations	Compare sequence-only, 3D-augmented, and omics-augmented models	Which information is essential for prediction
P1	Ranked gene expression representations	Compare ranked representations with representations preserving expression magnitudes	What information is lost through compression
P1	Transfer across laboratories and experimental batches	Split datasets by laboratory, measurement platform, or cell line lineage	Dependence on experimental conditions
P1	Prediction reliability	Calibration, uncertainty estimation, abstention	Whether unreliable predictions can be identified
P2	Practical usefulness of molecular generation	Generate novel antibodies and validate them experimentally	Whether sequence generation translates into molecular design
P2	Computational efficiency	Compare compute, training data, and maintenance costs at matched accuracy	Practical value of a common model

Three questions deserve particular attention:

1. Can MAMMAL generalize to genuinely unseen conditions?
2. Does integrated pretraining produce transferable biological knowledge?
3. Can the model predict causal effects of biological interventions?

6.2. Evaluating Generalization to Unseen Conditions#

The first priority is to distinguish performance on familiar drug–target combinations from performance involving unseen drugs or targets.

Relevant evaluation designs include:

• Drug-cold splits
• Target-cold splits
• Double-cold splits
• Temporal splits

These tests can help determine whether the model relies primarily on regularities within the observed data distribution or has acquired knowledge that remains useful under new conditions.

6.3. Evaluating the Value of Integrated Pretraining#

The second priority is to compare:

• Single-domain pretraining
• Multi-domain joint pretraining
• No pretraining

This comparison would help distinguish whether MAMMAL's performance advantages arise from:

• Its common architecture
• Knowledge transfer across biological domains
• Larger amounts of training data
• Greater model capacity

Comparisons under matched computational resources and training data budgets would also be valuable.

6.4. Evaluating Biological Causality#

The third priority is to test predictions under biological interventions.

Relevant experiments include:

• Gene knockout
• Drug treatment
• Rescue experiments

For example, suppose the model predicts that a particular gene is important for a drug response.

One could then test whether the model correctly predicts how the drug response changes when that gene is knocked out.

Agreement between predicted and experimentally observed intervention effects would provide evidence beyond conventional correlation-based prediction.

Together, these evaluations could clarify whether MAMMAL is primarily a successful multitask learning model, whether it has learned transferable representations across biological domains, and whether those representations support intervention prediction.

Before increasing model scale, it would also be important to compare MAMMAL with simpler statistical methods, specialized Graph Neural Networks (GNNs), and single-domain pretrained models under identical data and evaluation conditions.

7. What Is Missing from MAMMAL from a Virtual Cell Perspective?#

When considering biological simulation and Virtual Cell research, MAMMAL's limitations become more apparent.

A Virtual Cell model would ideally represent how a cell evolves from one state to another in response to time, interventions, and environmental conditions.

Conceptually, this can be expressed 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 manipulation
• (e_t): Extracellular environment
• (\Delta t): Time interval
• (F): Function describing cellular state transitions

MAMMAL's central demonstrated capability is not the learning of such a state-transition function.

Instead, its main demonstrated applications involve predicting specific properties or experimental endpoints from molecular and cellular inputs.

7.1. Endpoint Prediction Versus Dynamic Modeling#

In a MAMMAL-style prediction task, the model receives molecular or cellular information and produces an output such as:

• IC50
• Binding likelihood
• A classification result

In contrast, dynamic prediction for a Virtual Cell requires forecasting future cellular states under interventions, environmental influences, and elapsed time.

The model must also be able to advance from a predicted state to subsequent states and validate whether those states are biologically plausible.

Therefore, endpoint prediction and dynamic modeling represent different research problems.

7.2. Four Requirements for Virtual Cell Modeling#

At least four major requirements can be identified.

First, state sufficiency.

The observed cellular features must contain sufficient information to predict future cellular states.

If important variables—such as protein activity or cellular history—are absent from the input, accurate prediction may become difficult.

Second, temporal evolution.

The model must predict changes in genes, proteins, and cellular states across different time intervals.

Rather than predicting only a final endpoint, it should represent the trajectory through which cellular states evolve.

Third, responses to interventions.

The model must predict responses to previously unseen drugs, genetic manipulations, and combined interventions.

This requires capabilities beyond ordinary prediction within the training distribution.

Fourth, consistency across biological scales.

The model should connect molecular-level changes with cellular- and tissue-level outcomes without contradictions.

For example, drug binding to a protein, changes in signaling pathways, changes in gene expression, and changes in cellular function should be represented consistently.

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

MAMMAL could potentially serve as a component of a larger Virtual Cell system.

For example, it might be used as a module for predicting molecular interactions or specific cellular responses.

However, there is currently insufficient evidence that MAMMAL alone can realize a Virtual Cell.

A Virtual Cell would require capabilities involving:

• Dynamic state transitions
• Causal intervention effects
• Environmental dependence
• Consistency across biological scales

MAMMAL provides a potentially valuable foundation for integrating biological representations and tasks, but it is not itself a demonstrated Virtual Cell model.

8. Additional Considerations for Evaluating MAMMAL's Historical Significance#

The preceding analysis suggests four considerations that should be emphasized when positioning MAMMAL within the history of biological foundation models.

8.1. A Common Language Does Not Necessarily Imply Shared Knowledge#

MAMMAL implements a shared prompt syntax and a common model architecture.

However, the extent to which knowledge is genuinely shared across biological domains remains a separate empirical question.

The ability to process heterogeneous biological entities within one model does not automatically establish that the resulting representations encode biologically meaningful shared principles.

8.2. Integrating Information Is Different from Modeling Biological Phenomena#

The inclusion of 3D representations, as pursued by models such as BioMatrix, is an important development. [11]

However, adding 3D structure does not automatically provide an understanding of causality or dynamics.

Likewise, supporting richer modalities is not equivalent to learning more accurate biological mechanisms.

Expanding multimodal representations is therefore important, but it is not sufficient to solve every challenge in biological modeling.

8.3. MAMMAL Has Partially Validated Generalization#

The original MAMMAL paper includes DTI evaluations involving held-out protein classes and experimental validation using four new drugs. [2]

These provide meaningful evidence related to generalization.

Therefore, it would be inaccurate to state that MAMMAL's generalization capabilities have never been evaluated.

At the same time, these results do not establish broad predictive capability across arbitrary unseen drugs, targets, cells, and combined interventions.

The demonstrated scope and the remaining validation gaps must be distinguished.

8.4. The History of Biological AI Is Not a Simple Linear Progression#

MAMMAL, BioMatrix, ESM3, and MAP address different research objectives.

The fact that BioMatrix or MAP appeared later does not imply that they universally outperform MAMMAL.

MAMMAL's value lies in demonstrating a common architecture for diverse biological entities and tasks.

In contrast:

• BioMatrix emphasizes integration of sequences, structures, and natural language.
• ESM3 emphasizes integration of protein sequence, structure, and function.
• MAP emphasizes biological mechanism-related knowledge and prediction of cellular responses to unseen drugs. [9][11][12]

These approaches should be understood as distinct research directions rather than a simple ranking.

9. Overall Assessment: How Should MAMMAL Be Evaluated?#

The preceding discussion can be summarized as follows.

Evaluation Dimension	Assessment	Evidence and Remaining Considerations
Common biological input representation	Strong contribution	Implements structured prompts and integrates multiple domains
Support for diverse tasks	Strong contribution	Competitive performance across 11 benchmarks
Cross-domain knowledge transfer	Promising but not fully established	Requires better isolation of the contribution of integrated pretraining
Generalization to unseen conditions	Partial evidence	Evaluated on held-out target groups and a limited set of new drugs
3D and physical representations	Clear limitations	Primarily sequence-based representations
Biological causality	Insufficiently demonstrated	Requires intervention and mechanistic validation
Suitability for Virtual Cell research	Potentially useful	Dynamic state-transition capabilities have not been demonstrated
Real-world drug discovery utility	Further validation required	Lacks broad prospective experimental and decision-making evaluation

9.1. What Is MAMMAL Fundamentally Assuming?#

MAMMAL's central implicit assumption can be summarized as follows:

"By converting biological entities into common computational representations and jointly learning diverse tasks, a model can acquire generalizable predictive capabilities beyond those achievable by learning each task independently."

MAMMAL's reported results provide some support for this hypothesis.

However, extending the hypothesis to stronger claims requires additional evidence.

Such claims include:

• The model has learned biological mechanisms.
• The model can broadly extrapolate to unseen conditions.
• The model can simulate cellular dynamics.

Each of these claims requires separate experimental validation.

9.2. What Matters Most for Future Research?#

The most important next step is not necessarily to increase the number of modalities or model parameters.

Rather, the key question is:

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

Answering this question requires systematic evaluation of:

• Generalization to unseen drugs and targets
• Knowledge transfer between biological domains
• Causal effects of interventions
• Temporal changes in cellular states

It is also important to recognize that a common model may not always be the best solution.

For specific molecular structures or physical phenomena, specialized models may achieve higher performance.

Therefore, understanding the relationship between generalist and specialist models is another important direction for future research.

10. Conclusion: What Questions Has MAMMAL Left Open?#

MAMMAL demonstrates a concrete approach to handling diverse biological entities and tasks within a common architecture.

This contribution is significant in the history of biological foundation models.

However, representing biological data within a common language is not the same as understanding the underlying biological mechanisms.

MAMMAL's design rests on several implicit assumptions:

• Sequence-based representations preserve essential biological information.
• Different biological domains share transferable regularities.
• Joint learning across tasks produces useful knowledge transfer.
• Relationships learned from observational data remain useful under unseen conditions.

Although MAMMAL's results provide partial support for these assumptions, they have not all been comprehensively validated.

The original paper includes evaluations involving held-out protein classes and a limited set of new drugs.

Nevertheless, systematic generalization to unseen conditions, causal understanding of biological mechanisms, and prediction of time-dependent cellular responses remain important areas for further investigation.

These limitations become especially significant when considering Virtual Cell research.

A Virtual Cell requires more than the integration of different molecular and cellular data types.

It also requires the ability to represent temporal evolution, responses to interventions, and consistency across molecular, cellular, and tissue scales.

MAMMAL may provide valuable components for such systems, but it is not itself a demonstrated unified simulation of cellular biology.

In this sense, MAMMAL is better understood not as a completed unified model of biology, but as a generalist foundation model that demonstrates the potential of unifying biological entities and tasks while highlighting deeper challenges involving knowledge transfer, causality, dynamics, and generalization to unseen conditions.

Its significance lies not only in its ability to process diverse biological data using a shared architecture.

It also helps clarify a more fundamental question for the next generation of biological AI:

"Is converting biological entities into common representations sufficient, or must models explicitly learn 3D structures, physical constraints, biological mechanisms, and temporal dynamics to make accurate predictions under unseen conditions?"

This question represents both an unresolved challenge emerging from MAMMAL and a central research problem for future AI-driven drug discovery and Virtual Cell modeling.

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] 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/

[4] 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

[5] 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

[6] 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/

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

[8] Zhou, G. et al., “Uni-Mol: A Universal 3D Molecular Representation Learning Framework,” International Conference on Learning Representations (ICLR), 2023. https://arxiv.org/abs/2303.16982

[9] 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/

[10] Watson, J. L. et al., “De novo design of protein structure and function with RFdiffusion,” Nature, 620, pp. 1089–1100, 2023. https://www.nature.com/articles/s41586-023-06415-8

[11] 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

[12] 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

[13] “Evaluating the role of pretraining dataset size and diversity on single-cell foundation model performance,” Nature Methods, 2026. https://www.nature.com/articles/s41592-026-03120-y

[14] “PertEval-scFM,” 2025. https://arxiv.org/search/?query=PertEval-scFM&searchtype=all

[15] “Benchmarking algorithms for generalizable single-cell perturbation response prediction,” Nature Methods, 2026. https://www.nature.com/articles/s41592-025-02980-0

[16] Abramson, J. et al., “Accurate structure prediction of biomolecular interactions with AlphaFold 3,” Nature, 630, pp. 493–500, 2024. https://www.nature.com/articles/s41586-024-07487-w

[17] 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