REVIEW
How 'Foundational' Are Current Molecular Foundation Models?
Biomolecules · Submitted 29 Sep 2026 · 40 references
Pith's verdict: An honest, well-referenced survey that gives the molecular-ML field a usable checklist for 'foundational' claims; the survey conclusions hold, but the 'testable hypothesis' framing outruns the qualitative table labels.
T0 · AI review
Why Pith says this
Highlight any text to comment on it.
Referee report
The paper
The paper claims
No current molecular 'foundation model' fully satisfies the three testable criteria of generality, transferability, and out-of-distribution generalization, so 'foundational' should be treated as a hypothesis to test rather than a label…
The reading
What carries the argument
The central object is a three-criterion evaluation framework — generality, transferability, and out-of-distribution generalization — mapped onto four of the five properties of foundation models from the original report (multimodality, expressivity, memory capacity, and compositionality), with scalability deliberately set aside. The framework does its work by turning each model family into a row of Table 2, with labels '●' (criterion met), '◐' (partially met), '○' (not met), and '—' (not independently tested against a suitable baseline), each transferability row stating the adaptation budget and the baseline compared against, and each generalization row stating the split type. This forces claims about 'foundational' status to specify what evidence, benchmark, and baseline support them.
Extended reading notes
Core claim
The paper establishes that 'foundational' is best read not as a descriptive label assigned in advance but as a hypothesis about a model's capabilities, to be tested against three operational criteria: generality (one model spanning entity classes, property types, and task types), transferability (outperforming simpler, task-specific baselines with no or minimal fine-tuning), and out-of-distribution generalization (stable accuracy under distribution shifts such as scaffold splits, sequence dissimilarity, or perturbed structures). Applying these criteria to representative model families yields a mixed verdict: generality remains partial, with models spanning multiple entity classes typically restricted to a narrow range of properties and tasks and vice versa; transferability is difficult to establish consistently, with several small-molecule embedding models failing to beat fingerprint-based baselines while knowledge-injected models (CLAMP, CheMeleon, Mol-JEPA) succeed; and out-of-distribution generalization remains a major challenge, with the best-performing model in the BOOM benchmark showing an average OOD error roughly three times its in-distribution error. The paper reads the evidence as showing that success concentrates where target properties are consistently defined and training data are abundant with low label noise, and that scale alone is not the primary determinant of foundational capability, a claim it supports with examples like MiniMol matching larger models with two orders of magnitude fewer parameters.
Load-bearing premise
The verdict that no model satisfies all three criteria rests on the qualitative labels in Table 2, which assign 'met', 'partially met', or 'not tested' without a defined decision threshold; for several model families the relevant evidence is absent (marked '—'), so the conclusion partly reflects missing benchmarks rather than demonstrated failure, a fragility the paper itself acknowledges when it says the criteria are only as informative as the evaluations used to test them.
Editorial extensions
If this is right
- Any future model claiming 'foundational' status should be benchmarked against all three criteria with an explicit adaptation budget and a comparison to a simple, task-specific baseline, or the label should be withheld.
- Distribution-shift tests should be reported alongside conventional splits, since performance on conventional benchmarks does not predict accuracy under shifts in molecular structure, protein sequence, or conformational space.
- The two routes to cross-entity generality — mixed-resolution representations (proteins by residue, ligands by atom) and unified sequence languages — remain open, testable questions rather than demonstrated solutions.
- Progress toward molecular foundation models will depend on building coherent, well-characterized datasets and on incorporating prior chemical knowledge into models and tasks, not primarily on scaling model size.
Reading between the lines
- The same three criteria could discipline LLM-based chemistry assistants and agentic systems once they are evaluated for direct molecular capability, closing the gap the paper deliberately leaves open by excluding them.
- Peptides serve as a cheap discriminator: a model that spans atomistic small-molecule description and sequence-based protein description on peptides would supply direct evidence of genuine cross-entity generality, since peptides sit on the boundary between the two.
- If scale is not the primary lever, then investment patterns that favor ever-larger pretraining runs may be misdirected relative to spending on data curation, label quality, and shared evaluation infrastructure.
- Because Table 2's 'met' versus 'partially met' threshold is undefined, two evaluators could score the same evidence differently; prespecifying a decision rule would make the framework fully operational.
Original abstract
Large-scale models have permeated the molecular sciences, yet what makes a model 'foundational' in this domain remains poorly defined. This paper proposes three testable criteria for assessing the foundational nature of molecular models: (i) generality across molecular entities, properties, and tasks; (ii) transferability to new applications with no or minimal task-specific retraining; and (iii) generalization beyond the training distribution. Applying these criteria to the state of the art reveals promising progress, particularly visible in biomolecular structure prediction and machine-learned interatomic potentials, although none of the approaches examined fully satisfies all three. Success is concentrated in domains where target properties are consistently defined and training data are abundant, with low label noise relative to physically meaningful variation. More broadly, progress in the molecular sciences appears to depend less on model scale alone than on the quality and structure of available data, as well as the incorporation of prior knowledge into models, prediction tasks, or downstream applications. This work shifts the notion of a molecular foundation model from a descriptive label to a testable hypothesis, offering a framework for assessing current models and guiding future developments.
The compiled record
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.
The abstract, sentence by sentence
Large-scale models have permeated the molecular sciences, yet what makes a model 'foundational' in this domain remains poorly defined. This paper proposes three testable criteria for assessing the foundational nature of molecular models: (i) generality across molecular entities, properties, and tasks; (ii) transferability to new applications with no or minimal task-specific retraining; and (iii) generalization beyond the training distribution. Applying these criteria to the state of the art reveals promising progress, particularly visible in biomolecular structure prediction and machine-learned interatomic potentials, although none of the approaches examined fully satisfies all three. Success is concentrated in domains where target properties are consistently defined and training data are abundant, with low label noise relative to physically meaningful variation. More broadly, progress in the molecular sciences appears to depend less on model scale alone than on the quality and structure of available data, as well as the incorporation of prior knowledge into models, prediction tasks, or downstream applications. This work shifts the notion of a molecular foundation model from a descriptive label to a testable hypothesis, offering a framework for assessing current models and guiding future developments.
background contribution method result interpretation main claim
| # | sentence | role | the authors' own work | states the main conclusion | strength |
|---|---|---|---|---|---|
| s0 | Large-scale models have permeated the molecular sciences, yet what makes a model 'foundational' in this domain remains poorly defined. | gap 0.64 | 0.07 | 0.03 | asserted 0.99 |
| s1 | This paper proposes three testable criteria for assessing the foundational nature of molecular models: (i) generality across molecular entities, properties, and tasks; (ii) transferability to new applications with no or minimal task-specific retraining; and (iii) generalization beyond the training distribution. | contribution 0.97 | 0.88 | 0.11 | asserted 0.98 |
| s2 | Applying these criteria to the state of the art reveals promising progress, particularly visible in biomolecular structure prediction and machine-learned interatomic potentials, although none of the approaches examined fully satisfies all three. | result 0.98 | 0.83 | 0.29 | asserted 0.93 |
| s3 | Success is concentrated in domains where target properties are consistently defined and training data are abundant, with low label noise relative to physically meaningful variation. | result 0.67 | 0.56 | 0.12 | asserted 0.99 |
| s4 | More broadly, progress in the molecular sciences appears to depend less on model scale alone than on the quality and structure of available data, as well as the incorporation of prior knowledge into models, prediction tasks, or downstream applications. | interpretation 0.96 | 0.45 | 0.55 | asserted 0.67 |
| s5 | This work shifts the notion of a molecular foundation model from a descriptive label to a testable hypothesis, offering a framework for assessing current models and guiding future developments. | contribution 0.78 | 0.89 | 0.30 | asserted 0.99 |
What was examined, and what was not
- arXiv listing abstract v1
- full text
- figures and tables
- references and citing sentences
- relations to other papers
- P abstract
- S sections
- T body sentences
- C citations
- R cross-paper relations
Figures
Figures reproduced from arXiv:2609.37550 by the authors.
References
Discussion
Disagree with the review, or know something it missed? Leave a comment, highlight any passage above to comment on it, or challenge this review formally.