Why AI Biology Needs Better Evaluation
By Li Lei on July 13, 2026
In AI biology, we love numbers.
Training loss.
Validation loss.
Perplexity.
AUROC.
AUPRC.
Correlation.
Benchmark rankings.
A new foundation model improves one of these numbers, and we often call it progress.
Sometimes it is.
But increasingly, I find myself asking a harder question:
What does better model performance actually mean for biology?
Imagine that a DNA foundation model is trained on billions or trillions of nucleotides.
Its training loss falls.
Its validation loss falls.
Its sequence likelihood improves.
The optimization worked.
But what exactly have we learned?
Does the model understand gene regulation better?
Can it predict the effect of a variant it has never encountered?
Does its representation transfer to another species?
Can it predict what happens after a genetic perturbation?
Can it identify a biologically meaningful candidate that a simpler model would miss?
And ultimately:
Would its prediction change what experiment a scientist chooses to do next?
Those are very different questions.
Recent evaluations of DNA and single-cell foundation models make this distinction increasingly important. Large pretrained models do not automatically outperform specialized models or even simple baselines across every biological task. Performance can depend strongly on training objective, architecture, pretraining diversity, biological context, task definition, and how evaluation data are separated from training data [1–8].
At the same time, models such as AlphaGenome and Evo 2 demonstrate that large biological models can achieve remarkable capabilities when model design and evaluation are tightly connected to meaningful biological questions [9,10].
So the lesson is not:
Foundation models do not work.
Nor is it:
Bigger models will solve biology.
The more interesting lesson is:
There is no single metric that tells us whether a foundation model understands biology.
In the first essay of this series, I argued that biologists need to move beyond simply using AI tools and learn to think in AI systems.
In the second, I argued that biology is not just data — biological meaning depends on context.
Now comes the next question:
How do we know when an AI result deserves to influence science?
A lower loss is not biological understanding.
A benchmark score is not biological truth.
A prediction is not a mechanism.
And a prediction is certainly not yet a discovery.
1. When does a prediction become a discovery?
I find it useful to think about AI biology evaluation as a six-level stack.
Figure 1. When does an AI prediction become a discovery?
AI biology evaluation can move from foundation-model health through task performance, generalization, biological validity, experimental validation, and finally scientific decision value.
At the bottom, we ask a very technical question:
Did the model learn its training objective?
At the top, we ask a much more scientific question:
Did the model help us make a better decision or discover something new?
Those are not the same thing.
A model can perform beautifully at Level 0 and disappoint at Level 1.
It can perform well on a benchmark and collapse when confronted with a new species or experimental condition.
It can generalize statistically while learning a shortcut rather than a meaningful biological relationship.
And it can generate a biologically plausible hypothesis that fails when tested experimentally.
That is why I think we need to stop talking about AI evaluation as if it were one number.
It is a chain of evidence.
2. What does training loss actually tell us?
Before asking whether a foundation model understands biology, we should ask:
Did training work?
During pretraining, researchers may monitor:
training loss
validation loss
cross-entropy
token or sequence likelihood
masked-token reconstruction
calibration
convergence
scaling behavior
These metrics are essential.
If training loss does not decrease, optimization may be failing.
If training loss continues improving while validation loss deteriorates, the model may be overfitting.
But there is a crucial distinction.
Figure 2. Lower loss does not automatically mean better biology.
Training metrics tell us whether a model is learning its optimization objective. They do not by themselves establish downstream biological capability.
A lower validation loss tells us:
the model has become better at its mathematical objective.
It does not automatically tell us:
the model understands enhancers better;
predicts gene expression better;
generalizes across species better;
predicts perturbations better;
or makes better scientific decisions.
This distinction is particularly important for foundation models.
A masked DNA language model, an autoregressive nucleotide model, and a model using different tokenization schemes may optimize very different objectives. Their losses may not even be directly comparable.
Recent benchmarking supports this caution.
Feng and colleagues evaluated several DNA foundation models across genomic and genetic applications and found substantial task-dependent differences. General-purpose models were competitive for some problems, such as pathogenic-variant identification, while performing less strongly on others, including gene-expression prediction and causal-QTL identification [1].
Tang and colleagues similarly found that pretrained genomic-language-model representations did not consistently outperform conventional sequence representations or strong supervised models on biologically aligned regulatory-genomics tasks [2].
The important lesson is:
A foundation model can become a better model of its training data without necessarily becoming a better model of the biology we care about.
3. Foundation-model evaluation is a portfolio, not one score
Instead of searching for one universal metric, I think biological foundation models need several complementary types of evaluation.
Figure 3. Foundation-model evaluation is a portfolio, not one score.
Intrinsic, downstream, generalization, and biological evaluation answer fundamentally different questions.
Intrinsic evaluation
Did the model learn its pretraining objective?
Loss, likelihood, masked-token recovery and convergence belong here.
Downstream evaluation
Can the learned representation solve biologically meaningful tasks?
For DNA models, this might include:
variant-effect prediction,
enhancer classification,
gene-expression prediction,
splice prediction,
or regulatory activity.
For single-cell models:
cell-state representation,
perturbation response,
cell-type prediction,
or gene-expression reconstruction.
Generalization evaluation
Does the model still work when it encounters biology that is meaningfully different from what it saw during training?
New species?
New populations?
New tissues?
New sequence families?
New perturbations?
Biological evaluation
Do the predictions survive independent biological evidence?
Genetics?
Evolution?
Functional assays?
Perturbation?
Prospective validation?
These categories answer different questions.
That is why statements such as:
“Model A has lower perplexity, therefore it is better for biology”
are too strong.
Scale can create new capability.
But evaluation tells us what capability was actually created.
4. Zero-shot performance is useful — but it needs careful interpretation
One particularly interesting promise of foundation models is zero-shot prediction.
Imagine introducing a nucleotide variant into a sequence.
A genomic language model can compare the probability or likelihood assigned to the reference and alternative sequence.
If the variant disrupts a strongly learned sequence pattern, the model may assign the altered sequence a lower probability.
That creates a variant-effect score without training directly on pathogenicity labels.
Conceptually, this is powerful.
But recent work shows why evaluation matters.
Alfisi and colleagues systematically compared DNA foundation models for zero-shot variant-effect prediction and found that performance depended on more than model size. Architecture, training objective, sequence context, and diversity of pretraining data all mattered [3].
One particularly interesting observation was the value of multi-species training.
Evolution has already performed billions of years of natural experiments.
Training across diverse species may therefore expose models to patterns of biological constraint that are difficult to learn from one genome alone.
This suggests something deeper:
the composition of the training data may matter as much as the number of model parameters.
And it reinforces why foundation-model evaluation cannot be reduced to a scaling curve.
5. Strong baselines keep us scientifically honest
One of the most useful lessons from recent AI biology studies is remarkably simple:
Always compare a sophisticated model against a strong simple baseline.
This has become especially clear in single-cell biology.
Kedzierska and colleagues evaluated Geneformer and scGPT in zero-shot settings and found that their representations did not consistently outperform simpler established approaches [5].
Ahlmann-Eltze and colleagues examined prediction of transcriptional responses to genetic perturbation and found that several deep-learning approaches did not consistently outperform deliberately simple linear baselines [6].
This does not mean that single-cell foundation models have failed.
It means that a complicated model should earn its complexity.
If a very large model cannot outperform a linear model or strong task-specific baseline on the capability being claimed, that tells us something important.
Perhaps the objective needs improvement.
Perhaps the benchmark is too easy.
Perhaps the training data do not contain enough causal information.
Perhaps the representation captures cell identity well but not intervention response.
Emerging bioRxiv work is pushing this idea further. The scFME framework, for example, evaluates single-cell foundation models specifically for in-silico perturbation, comparing predicted changes against control, zero, and random perturbation baselines rather than simply asking whether embeddings appear structured [7].
That distinction matters.
A beautiful embedding is not necessarily a useful perturbation model.
A beautiful UMAP is not a discovery.
6. Is the model generalizing — or recognizing something familiar?
This is one of the most important problems in biological machine learning.
Biological observations are rarely independent.
DNA sequences share homology.
Proteins belong to families.
Individuals share ancestry.
Cells may come from the same donor.
Species share evolutionary history.
Experiments share protocols and batch effects.
A random train-test split can therefore create the illusion of generalization.
Figure 4. Unseen rows do not necessarily mean unseen biology.
A random split may place highly homologous biological examples in both training and test sets. Biology-aware splits provide a stronger test of genuine generalization.
Suppose Gene A1 and Gene A2 are in training.
Gene A3 is placed in the test set.
But if A3 is highly homologous to A1 and A2, has the model really encountered new biology?
Not necessarily.
Recent bioRxiv work by Rafi and colleagues examined this problem explicitly, demonstrating how homology can create leakage between training and evaluation partitions in genome-trained sequence models [4].
This leads to one of the most important principles in biological AI evaluation:
A test example is not truly “unseen” simply because its row was absent from the training set.
Biological independence matters.
Depending on the scientific question, stronger evaluation may require holding out:
an entire chromosome,
a gene family,
a homology cluster,
a donor,
a population,
an environmental condition,
or an entire species.
These evaluations are harder.
But discovery happens precisely when models encounter something they have not already seen.
7. Did the model learn biology — or a shortcut?
Even if a model generalizes statistically, we still need to ask:
What did it actually learn?
Suppose an AI system predicts genes associated with drought tolerance.
Perhaps the top candidates are:
expressed in the relevant tissue,
responsive under drought,
near genetic associations,
supported by chromatin accessibility,
connected to stress-response pathways,
or evolutionarily conserved.
Our confidence increases.
But perhaps the model instead learned:
GC content,
gene length,
annotation density,
batch identity,
phylogenetic similarity,
or another feature correlated with the label.
Both models might produce high AUROC.
Only one may teach us useful biology.
This is where interpretability becomes valuable.
Not because every neural network must produce a simple human-readable explanation, but because probing model representations can reveal whether recognizable biological features have emerged.
Evo 2, for example, reported internal features associated with exon-intron boundaries, transcription-factor-binding sites, protein structural properties, and prophage-related sequence [10].
That is encouraging.
But interpretability is still not the final test.
A model feature that resembles biology is evidence.
It is not yet mechanism.
8. AlphaGenome: evaluation closer to biological measurement
AlphaGenome offers an interesting example of evaluation that is more closely aligned with biological observables.
Rather than producing only a generic embedding, AlphaGenome predicts thousands of genomic tracks from long DNA sequences, including:
gene expression,
transcription initiation,
chromatin accessibility,
histone modifications,
transcription-factor binding,
chromatin contacts,
and splicing [9].
Its variant-effect evaluations then ask how sequence changes alter these measurable molecular outputs.
That is scientifically attractive because the evaluation moves closer to quantities biologists actually measure.
The principle is broader than AlphaGenome:
The closer an AI evaluation is to a meaningful biological measurement or intervention, the stronger the scientific claim we can make.
Predicting a benchmark label is useful.
Predicting an experimentally measurable consequence is stronger.
Predicting a new experimental outcome prospectively is stronger still.
9. Can nature prove the model wrong?
For me, this is where AI biology becomes most exciting.
A good scientific prediction should eventually confront reality.
Imagine an AI system predicts that a regulatory sequence controls a drought-responsive gene.
What would test that hypothesis?
CRISPR perturbation?
Reporter assay?
Perturb-seq?
Expression profiling?
Chromatin accessibility?
Protein binding?
A field experiment?
The exact experiment depends on the biological question.
But the principle is universal:
A scientific AI system should ideally generate hypotheses that can be falsified.
We should be able to ask:
What result would prove this prediction wrong?
Loss cannot answer that.
AUROC cannot answer that.
A benchmark cannot completely answer that.
Eventually, nature has to answer.
That is why prospective validation is so powerful.
Make the prediction first.
Then collect new evidence.
Blind prediction challenges such as CASP became influential in structural biology because predictions were evaluated against experimentally determined structures that participants did not know when making their predictions.
The answer existed outside the model.
AI biology needs more tests with this character:
prospective prediction,
blind validation,
independent experiments,
orthogonal assays,
and perturbational testing.
Instead of asking only:
Can the model explain biology we already measured?
we should increasingly ask:
Can it correctly predict biology we have not measured yet?
That is much closer to discovery.
10. Does AI change the next scientific decision?
There is one final level of evaluation that I think receives too little attention.
Does the AI system improve a decision?
Scientists rarely build models because they simply want more predictions.
Eventually, we need to decide:
Which gene should we perturb?
Which variant should we prioritize?
Which molecule should we synthesize?
Which experiment should we run?
Which hypothesis deserves another six months of work?
Imagine two models ranking 10,000 genes.
Model A:
AUROC = 0.92
Model B:
AUROC = 0.89
Model A appears better.
But the laboratory can test only ten genes.
Among those ten predictions:
Model A contains two validated candidates.
Model B contains six.
Which model is more scientifically useful?
Probably Model B.
This suggests that AI biology may need to pay more attention to metrics such as:
top-k enrichment,
experimental hit rate,
number needed to test,
cost per validated candidate,
search-space reduction,
and information gained from the next experiment.
These metrics move us from prediction performance toward decision value.
And ultimately, decision value may be one of the most important measures of AI's contribution to science.
11. What about scientific AI agents?
Evaluation becomes even more difficult when we move from models to agents.
A model might perform one task:
predict a variant effect.
An AI agent might:
interpret a scientific question,
search literature,
retrieve data,
write analysis code,
choose statistical methods,
run tools,
interpret results,
generate hypotheses,
and recommend an experiment.
How do we evaluate such a system?
Only the final answer?
The retrieved evidence?
The analysis code?
The statistical choices?
The reproducibility?
The proposed experiment?
One subtle error early in the workflow can propagate through everything downstream.
ScienceAgentBench was developed specifically to test data-driven scientific agents using 102 tasks derived from 44 peer-reviewed publications [12].
Even advanced agent configurations solved only a fraction of the tasks reliably.
That does not mean scientific agents are unpromising.
It means we should be cautious about moving from impressive demonstrations to claims of autonomous scientific discovery.
The more autonomous an AI system becomes,
the more rigorous its evaluation must become.
Not less.
Human agreement is not the final benchmark
There is another complication.
Human expertise is essential.
But human agreement cannot always be the ultimate standard.
AlphaGo's famous Move 37 is a useful reminder.
The move initially appeared highly unusual to expert players.
Its value became clear later.
Science may eventually encounter its own versions of Move 37.
An AI system may propose:
an unexpected gene,
a surprising regulatory mechanism,
an unusual protein sequence,
a non-obvious molecular interaction,
or an experiment an experienced scientist would never have prioritized.
If our evaluation asks only:
Does the expert agree with the AI?
we may reject precisely the discoveries AI could help us make.
So human judgment matters.
But the stronger question is:
Does reality agree?
Does the molecule function?
Does the sequence regulate expression?
Does the protein fold?
Does the perturbation change the phenotype?
Does the result replicate?
The ultimate judge of scientific AI should not be AI itself.
And it should not always be human intuition.
It should be evidence.
From prediction to discovery
We are entering an extraordinary period in biology.
Models such as AlphaGenome and Evo 2 demonstrate capabilities that would have seemed extraordinarily ambitious only a few years ago [9,10].
Foundation models are learning increasingly rich representations of DNA, proteins, cells, and molecules.
Scientific AI agents are beginning to interact with multi-step research workflows.
The opportunity is real.
But greater capability creates a greater need for scientific discipline.
As models become more powerful, generating plausible predictions becomes easier.
Determining which predictions deserve our trust becomes harder.
So perhaps the most important question in AI biology is no longer simply:
Can AI make a prediction?
We should ask:
Did the model actually learn?
Then:
Does it work on the biological task?
Then:
Does it generalize to genuinely new biology?
Then:
Did it learn meaningful biology rather than a shortcut?
Then:
Can we test its prediction?
Then:
Did it improve the next scientific decision?
And eventually:
Did it help us discover something we did not know before?
That is the journey from prediction to discovery.
Training loss tells us whether optimization is working.
Benchmarks tell us whether a model performs on defined tasks.
Strong baselines tell us whether complexity adds value.
Generalization tests tell us whether the capability survives unfamiliar biology.
Biological validation tells us whether the prediction survives context.
Experiment tells us whether nature agrees.
Decision value tells us whether AI actually changes science.
None of these alone is enough.
Together, they form a much stronger definition of progress.
The future of AI biology will therefore not be defined only by who builds the largest model,
trains on the most data,
achieves the lowest loss,
or reaches the top of a leaderboard.
It will be defined by our ability to distinguish between:
a well-trained model,
a strong benchmark result,
a biologically meaningful prediction,
a plausible hypothesis,
reliable evidence,
and a real discovery.
Those are not the same thing.
Knowing the difference is scientific judgment.
And as AI becomes more powerful, I believe that judgment will become more valuable — not less.
References and Further Reading
[1] Feng, H., Wu, L., Zhao, B. et al. Benchmarking DNA foundation models for genomic and genetic tasks. Nature Communications 16, 10780 (2025). doi:10.1038/s41467-025-65823-8.
[2] Tang, Z., Somia, N., Yu, Y. & Koo, P. K. Evaluating the representational power of pre-trained DNA language models for regulatory genomics. Genome Biology 26, 203 (2025). doi:10.1186/s13059-025-03674-8.
[3] Alfisi, I., Ciapi, F., Baragli, M. & Magi, A. Benchmarking DNA foundation models for zero-shot variant effect prediction shows the importance of context, training, and architecture. Genome Biology (2026). doi:10.1186/s13059-026-04238-0.
[4] Rafi, A. M., Kiyota, B., Yachie, N. & de Boer, C. Detecting and avoiding homology-based data leakage in genome-trained sequence models. bioRxiv (2025). doi:10.1101/2025.01.22.634321. Preprint.
[5] Kedzierska, K. Z., Crawford, L., Amini, A. P. & Lu, A. X. Zero-shot evaluation reveals limitations of single-cell foundation models. Genome Biology 26, 101 (2025). doi:10.1186/s13059-025-03574-x.
[6] Ahlmann-Eltze, C., Huber, W. & Anders, S. Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. Nature Methods 22, 1657–1661 (2025). doi:10.1038/s41592-025-02772-6.
[7] Boylan, J., Solovyeva, E., Bouiller, T. et al. Single Cell Foundation Models Evaluation (scFME) for In-Silico Perturbation. bioRxiv (2025). doi:10.1101/2025.09.22.677811. Preprint.
[8] Weidener, L. S., Brkić, M., Jovanović, M., Ulgac, E. & Meduri, A. VCBench: A Multi-Dimensional Benchmark for Single-Cell Foundation Models. bioRxiv (2026). doi:10.64898/2026.06.18.733146. Preprint.
[9] Avsec, Ž., Latysheva, N., Cheng, J. et al. Advancing regulatory variant effect prediction with AlphaGenome. Nature 649, 1206–1218 (2026). doi:10.1038/s41586-025-10014-0.
[10] Brixi, G., Durrant, M. G., Ku, J. et al. Genome modelling and design across all domains of life with Evo 2. Nature 652, 1349–1361 (2026). doi:10.1038/s41586-026-10176-5.
[11] Kapoor, S., Cantrell, E. M., Peng, K. et al. REFORMS: Consensus-based Recommendations for Machine-learning-based Science. Science Advances 10, eadk3452 (2024). doi:10.1126/sciadv.adk3452.
[12] Chen, Z., Chen, S., Ning, Y. et al. ScienceAgentBench: Toward Rigorous Assessment of Language Agents for Data-Driven Scientific Discovery. International Conference on Learning Representations (ICLR) (2025).
Note on preprints
Several references above are bioRxiv preprints and have not yet undergone peer review. I include them because biological foundation-model evaluation is developing rapidly and some important methodological work appears first as preprints. Their findings should therefore be interpreted as emerging evidence rather than established consensus.
No comments:
Post a Comment