A Guide to Trivial Baselines in Computational Biology

Biology is extremely noisy. It is common for replicates of the same biological experiment to show low correlation, sometimes barely at all, even when the protocol, cell line, and every other controllable variable remain unchanged.

One active area of research, for example, is predicting the transcriptomic effect of a small molecule on a given cell line. The effect is measured relative to a population of control cells and is usually represented as a vector of (log) fold-changes across a panel of genes. Yet results from separate experiments, and sometimes even from different batches of the same experiment, can be largely uncorrelated, especially when the perturbation has only a weak effect.

This makes even the most trivial baselines surprisingly competitive, as there may be very little signal to recover after all. Such baselines are often omitted from studies, perhaps because they seem too obvious to be worth testing. Every few months, a new study shows that a handful of simple baselines outperform a large-scale model, whether in gene-expression perturbation prediction (Ahlmann-Eltze et al., 2025), single-cell foundation models (Kedzierska et al., 2025), or molecular property prediction (Jiang et al., 2021). This cuts both ways, though: rigorous baseline comparison is also how you find the cases where a large model genuinely earns its keep, as for some foundation models on perturbation response prediction (Cole et al., 2026).

This post catalogs common baselines that fall under the broad label of “trivial.” By trivial, I mean simple computations or transformations of the data, along with the most basic inference methods, such as a nearest-neighbor estimator with k=1. What follows is a task-by-task walkthrough, from expression prediction to time-series forecasting, with concrete baselines to compute for each.

A task-by-task baseline catalog

1. Expression prediction

This category covers expression prediction under perturbation, differentiation, disease, development, or transfer to a held-out cell type.

• Training-set meanPredict the average control expression profile across all training cells or samples. This is worth computing alongside a context-specific mean and can be useful for debugging. Suppose a model converges to the global training mean but not the relevant context mean. If the context mean performs better than the training mean, this may indicate that the model is failing to use its context encoder.
• Context meanPredict the average expression profile among training samples that share the same context. Context might refer to tissue, cell type, disease state, donor, experimental condition, or some combination of these. Several context means of increasing specificity can be defined to reveal which level of metadata accounts for most of the predictive performance. Pushed further, this idea of learning how responses vary across reference contexts can itself become a competitive predictor, matching much more expensive models while training in seconds (Svensson et al., 2026).
• Control expressionIn a perturbation study, predict that the intervention has no effect and return the expression profile of the control population. If the target is represented as a difference or log-fold-change vector, this is equivalent to predicting zeros across all genes.
• All-zero expressionWhen predicting raw expression rather than a difference vector, predict that expression is literally zero everywhere. This is distinct from the no-change baseline above. It is unlikely to be scientifically useful, but it can reveal whether the model has collapsed to a trivial solution or whether the loss function or metric is badly misaligned with the intended task.
• Nearest-neighbor profileFind the most similar cell or sample in the training set and use its observed expression profile as the prediction. The feature space and distance metric should always be reported. A neighbor found using raw counts is not the same as one found in PCA space.
• Additive perturbation effectsFor combinatorial interventions (e.g., knocking out two genes), predict the sum of the individual perturbation effects, when those effects are available. If a model cannot beat this baseline, it has not captured anything meaningful beyond the independent effects of the individual components (Ahlmann-Eltze et al., 2025).

2. Differential-expression gene-set prediction

• Random gene setPredict a random set containing the same number of genes as the true DE set. This provides a floor for precision, recall, overlap, or other ranking-based metrics. In practice, it may also be appropriate to sample genes based on their expression or detection frequency.
• A fixed set of common genesAlways predict the same set of genes. One option is to use a published list of housekeeping genes. Another is to use the genes that are most frequently identified as differentially expressed in the training set.
• Highly variable genesCompute the highly variable genes using the train set and always predict that same set. This tests whether the supposed DE signal is largely driven by genes that are variable under almost any condition.
• Empty setPredict that no genes are differentially expressed. This is useful when the true DE sets are small or sometimes empty. It is also an effective test of metric behavior: some metrics are undefined when both the predicted and true sets are empty, and libraries such as scikit-learn allow different conventions for handling these cases (e.g., the zero_division parameter in f1_score). Those choices can materially affect the reported result. It is also worth checking that a DE method returns close to nothing on data with no real biological difference, since several popular methods do not (Squair et al., 2021).

3. Gene-regulatory networks and graph inference

This category includes gene-regulatory network inference, cell-cell communication networks, and protein-interaction prediction.

When inferring a graph

• CoexpressionUse Pearson or Spearman correlation, partial correlation, or another simple measure of association as the edge score. Coexpression is not regulation, but if a sophisticated regulatory model cannot outperform a correlation matrix, that distinction may not matter much in practice (Pratapa et al., 2020).
• Random graphGenerate a random graph with the same number of nodes and approximately the same number of edges or same degree distribution as the reference graph. Matching graph statistics matters, as a very sparse graph is not a meaningful comparison for a dense graph.
• Fully connected graphPredict every possible edge. This can expose evaluation metrics that heavily reward recall without adequately penalizing false positives.

When using an existing graph as part of a model

Suppose a model incorporates a graph such as the STRING protein-interaction network.

• Node degreeFor gene-level predictions, compare the model against the weighted or unweighted degree of each gene in the input graph. Many biological labels correlate with connectivity. Gene essentiality, for example, is often associated with node degree (Jeong et al., 2001).
• Expression levelCompare against the input gene's expression level, especially when expression is explicitly supplied as a node feature. It is easy to assume that the graph is doing the heavy lifting when the model has simply learned to prioritize highly expressed genes.
• A graph-free ablationTrain the same model, or the closest meaningful equivalent, without the graph. This is not quite a trivial baseline but helps determine if the graph contributes anything at all.
• Cell-level baselinesIf the final prediction is made at the cell or sample level, the expression and DE baselines above all apply.

4. Representation learning and embedding quality

This is especially relevant when evaluating embeddings produced by foundation models.

• Raw or normalized expressionIf foundation-model embeddings are used for a downstream task such as cell-type classification, drug-response prediction, or drug repurposing, train the same downstream model directly on raw or normalized expression values.
• PCACompare against a 50-100 component PCA. PCA is usually extremely good at filtering noise from expression data. If possible, keep the number of components the same as the dimensionality of the learned embedding. It is a surprisingly durable baseline against learned embeddings (Bendidi et al., 2024).
• Random embeddingsReplace the learned embeddings with random features or embeddings produced by an untrained model. This establishes whether performance comes from the learned representation itself, architectural inductive bias, or some downstream classifier.
• Simple metadataUse obvious covariates such as donor, batch, or tissue. If metadata is highly predictive or correlates with the downstream label, then strong probe performance does not necessarily demonstrate meaningful biological reasoning.

5. Sequence-to-function prediction

This includes variant-effect prediction and other attempts to predict molecular function directly from DNA, RNA, or protein sequence.

• No change from the reference sequencePredict that a mutation or sequence modification has no effect relative to the unmutated sequence. This can be a strong baseline when most variants are neutral.
• k-mer modelsTrain a simple model using k-mer counts. This helps determine whether the signal captured by a larger sequence model comes entirely from short motifs or whether the model has learned meaningful long-range dependencies.
• Gene- or protein-identity baselineWhen constructing data splits, it is often preferable to hold out entire genes or proteins. This provides a more convincing test of generalization. If variants from the same gene or protein appear in both the training and test sets, include the gene- or protein-identity baseline. Otherwise, a model may succeed by memorizing that some genes tend to receive higher scores than others (or are more studied and thus better represented in the data, such as BRCA1 variants).
• Conservation scoresFor variant-effect prediction, include simple evolutionary conservation scores whenever they are available and relevant.

6. Small-molecule property prediction

• Molecular weightUse molecular weight as the only input, or at least measure its direct association with the target.
• Molecular fingerprintsTrain a simple model using a standard molecular fingerprint.
• Simple physicochemical descriptorsFit a linear model, regularized linear model, or random forest using a small collection of standard descriptors such as molecular weight, logP, polar surface area, hydrogen-bond counts, formal charge, and number of rotatable bonds. Descriptor-based models like these are competitive with, and often beat, graph-based models on standard benchmarks (Jiang et al., 2021).

7. Trajectories, dynamics, and time-series prediction

This includes cell differentiation trajectories, pseudotime, and disease or treatment time courses. No single trajectory-inference method dominates across datasets and topologies, which makes these baselines a useful reference point regardless of which method is used (Saelens et al., 2019).

• No-change baselineCarry the last observed state forward unchanged. This is the first baseline to compute for almost any time-series problem. Many systems change slowly enough that persistence is difficult to beat over short prediction horizons.
• Linear extrapolationEstimate a straight-line trend or velocity from the last two or more observed time points and extrapolate it into the future. This provides a basic test of whether the model captures nonlinear dynamics or merely continues the most recent direction of change.
• Training-set mean trajectoryPredict the average trajectory observed in the training set, ignoring the specific starting state. A more refined version computes the mean trajectory within the relevant cell type, treatment group, disease state, or other context.
• Regression toward the meanPredict that the system moves toward the global or context-specific mean. This is a useful baseline because noisy measurements naturally tend to be followed by less extreme ones. A model may appear to predict recovery, stabilization, or convergence when it is mainly exploiting regression to the mean.

The oracle, or experimental ceiling

Finally, one of the most useful reference points is an oracle baseline, or more realistically, an estimate of the experimental ceiling. This is the same idea as the “noise ceiling” used in other empirical sciences to bound how well any model could possibly fit noisy data (Nili et al., 2014).

In many biological settings, replicate-to-replicate agreement provides an empirical estimate of the maximum performance one could reasonably expect. If two experimental replicates correlate at only 0.5, a model achieving 0.45 may be performing extremely well, even though a correlation of 0.45 looks unimpressive in isolation.

The experimental ceiling and the random baseline define the upper and lower goalposts. Together, they make it possible to report performance on a normalized scale, where a score of 0 corresponds to random performance and a score of 1 reaches the estimated experimental ceiling.

Closing thoughts

Trivial baselines are cheap. Discovering months later that the trivial baseline wins is considerably more expensive. It is a good habit to compute these baselines first and log them alongside your other metrics, so that the point of reference is always visible no matter your approach.

References

[1] 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).

[2] Bendidi, I. et al. Benchmarking transcriptomics foundation models for perturbation analysis: one PCA still rules them all. arXiv:2410.13956 (2024).

[3] Cole, E. et al. Foundation models improve perturbation response prediction. bioRxiv (2026).

[4] Jeong, H., Mason, S. P., Barabási, A.-L. & Oltvai, Z. N. Lethality and centrality in protein networks. Nature 411, 41–42 (2001).

[5] Jiang, D. et al. Could graph neural networks learn better molecular representation for drug discovery? A comparison study of descriptor-based and graph-based models. Journal of Cheminformatics 13, 12 (2021).

[6] 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).

[7] Nili, H., Wingfield, C., Walther, A., Su, L., Marslen-Wilson, W. & Kriegeskorte, N. A Toolbox for Representational Similarity Analysis. PLoS Computational Biology 10, e1003553 (2014).

[8] Pratapa, A., Jalihal, A. P., Law, J. N., Bharadwaj, A. & Murali, T. M. Benchmarking algorithms for gene regulatory network inference from single-cell transcriptomic data. Nature Methods 17, 147–154 (2020).

[9] Saelens, W., Cannoodt, R., Todorov, H. & Saeys, Y. A comparison of single-cell trajectory inference methods. Nature Biotechnology 37, 547–554 (2019).

[10] Squair, J. W. et al. Confronting false discoveries in single-cell differential expression. Nature Communications 12, 5692 (2021).

[11] Svensson, V. et al. Back to basics: Observed statistics are sufficient to predict drug responses. bioRxiv (2026).