Abstract
In the last few years, biologists and computer scientists have claimed that the introduction of data science techniques in molecular biology has changed the characteristics and the aims of typical outputs (i.e. models) of such a discipline. In this paper we will critically examine this claim. First, we identify the received view on models and their aims in molecular biology. Models in molecular biology are mechanistic and explanatory. Next, we identify the scope and aims of data science (machine learning in particular). These lie mainly in the creation of predictive models which performances increase as data set increases. Next, we will identify a tradeoff between predictive and explanatory performances by comparing the features of mechanistic and predictive models. Finally, we show how this a priori analysis of machine learning and mechanistic research applies to actual biological practice. This will be done by analyzing the publications of a consortium—The Cancer Genome Atlas—which stands at the forefront in integrating data science and molecular biology. The result will be that biologists have to deal with the tradeoff between explaining and predicting that we have identified, and hence the explanatory force of the ‘new’ biology is substantially diminished if compared to the ‘old’ biology. However, this aspect also emphasizes the existence of other research goals which make predictive force independent from explanation.
This is a preview of subscription content, log in via an institution to check access.
Similar content being viewed by others
Notes
It should be emphasized that the effects of the use of computers in biology have been studied well before the studies we refer to. The use of computers and computational models has certainly boosted in the last few decades (see for instance Keller 2002, Chapter 8), but these have been mostly used as ‘crutches’ and instrumentally to achieve the aims that are traditionally ascribed to biology.
We do not commit to any particular thesis on this aspect.
‘Description’ does not mean necessarily a linguistic description; actually, the preferred means for expressing mechanistic description are diagrams (see Bechtel and Abrahamsen 2005 for preliminary arguments on this matter).
It is not necessary to recall in detail the complex structure of mechanistic explanations; for a full-fledged account of this issue, see (Craver 2007).
There is an interesting debate on what ‘relevant’ might possibly mean and the difference between leaving out relevant details from the model (i.e. incomplete models) and abstracting from irrelevant details (i.e. abstraction). See in particular (Levy 2014; Levy and Bechtel 2014; Love and Nathan 2015).
De Regt’s characterization of understanding is much more rich than this, but for the purpose at hand this definition is sufficient.
Please note that it does not mean that large models cannot be used to draw general predictions that can be also verified experimentally. For instance, if you model protein–protein interactions with network science, you will obtain a very large model that cannot be turned into an explanation—it is impossible to draw the exact causal narrative connecting all the entities. However, network science tools identify central hub, and one can draw the very general prediction that, if I knock-down a central hub, then the network—and hence the biological phenomenon—will be disrupted. However, this prediction does not help any researcher in elaborating a mechanistic explanation.
There are similarities with the notion of problem representation in a problem space as famously indicated by Newell and Simon (even though here we refer to the formulation made by Bechtel and Richardson). Bechtel and Richardson’s concept of problem is an instantiation of the four elements of the problem representation (initial state, goal, defining moves, path constraints). For machine learning, a problem is an instantiation of an underlying input–output relation (a function) which has been sampled in order to obtain the dataset to be supplied to the learning algorithm.
While there are algorithms that do not halt for some inputs or may require an unlimited amount of memory, they are not used in data science, so we will ignore them in what follows.
Algorithms in machine learning can be supervised, unsupervised and reinforcement learning. We will refer to supervised and unsupervised algorithms because they are relevant to ‘discover’ quantitative relations between inputs and outputs. On the other hand, reinforcement learning algorithm are the less relevant here, which are used especially in engineering rather than natural or social sciences.
Here we must remark that each learning algorithm builds a different kind of algorithmic model. Some learning algorithms build complex models composed by many submodels which are combined by a consensus subalgorithm. These are called ensemble algorithms and models.
“The generalization performance of a learning method relates to its prediction capability on independent test data. Assessment of this performance is extremely important in practice, since it guides the choice of learning method or model, and gives us a measure of the quality of the ultimately chosen model” (Hastie et al. 2009, p. 219).
The algorithm produces the model, which in turn is applied to test new data.
Given a universe of objects, a class is a subset of the universe whose elements share some features which makes the class relevant for some scientific or technological purpose.
Abundance of data facilitates a better model selection and assessment, because the estimations of the performance of a model are more accurate (Hastie et al. 2009, p. 222). Infrequent but relevant cases can only be observed in sufficient numbers if the number of samples is large enough (Junqué de Fortuny et al. 2013, p. 216). For an important kind of algorithmic models such as maximum likelihood estimators, there are formal proofs that under mild conditions they are both asymptotically unbiased, i.e. the bias reduces to zero as the number of samples tends to infinity; and asymptotically efficient, i.e. the variance reduces to the lowest possible for any model as the number of samples tends to infinity (Sugiyama 2015, pp. 140–143). In some fields such as natural language processing, as the number of samples increases, a point is reached where all relevant cases are sufficiently covered in the dataset, so that memorization of examples outperforms models based on general patterns (Halevy et al. 2009, p. 9).
Please note that this concept of ‘black box’ differs from Latour’s, who refers to ready-made computer systems which are assumed to perform their function correctly (Latour 1987, pp. 3–4).
Any how-possibly model is in a sense a black-box model, because it establishes that there are components that are clearly involved in a phenomenon, but we do not know exactly how.
There are some machine learning methods like Bayesian networks (Spirtes et al. 1993) which can learn causal connections among variables from data. They can ascertain that some variables are causes of other variables, but they cannot say anything about the specific mechanism that is behind such causal connection. In other words, Bayesian networks by themselves cannot produce any explanations, since the specific mechanisms must be found by the scientist.
The real-world process and the model are categorically different. For trained machine learning models, they might not even be structurally similar, depending on the kind of algorithm that is used to learn the model. There are algorithms that aim to learn the structure of the biological interactions, which can be regarded as white box algorithms, while other algorithms do not try to yield a model which resembles the target biological system.
Well known cases include stock market prediction, modeling the spread of communicable diseases on a population, and recommendation systems for online marketing.
Please note once again that predictions here do not necessarily overlap with predictions in the mechanistic context.
Some data sets are so large and complex that it would be almost impossible for a human to find significant patterns without the help of algorithms which automatically elaborate models to fit the data (Dhar 2013, p. 68).
A difference between the desiderata of models f′ for explanation-based and data science approaches is that for the former we want f′ to have a causal structure that is similar (not understood in the philosophy of science technical sense) as much as possible to the phenomenon we are investigating, while for data science the predictive performances for new cases is the main goal.
Let’s consider a variation of the example of temperatures. We may say that the temperature tomorrow at 0:00 GMT in a particular weather station will be the average of the temperatures recorded on the same day of the year at 0:00 GMT in the same weather station, computed over the available weather data. We can diminish variance as follows. As the number of years with available data (the number of samples) increases, the variance diminishes because the output that f’ produces for unseen test data will be less sensitive to oscillations in the training dataset, i.e. extremely cold or extremely hot years in the historic record which is used for training.
This makes sense only if we assume that complex systems must be approached by taking into account the contribution of each of their components. There might be other approaches to complexity (e.g. systemic or holistic approaches) which may not require the attitude we are describing here.
It can be argued that there are algorithms which learn the structure of the biological system under investigation, so that we can relate somehow the work of algorithms to mechanistic descriptions. For example, the algorithm PARADIGM mentioned above does learn the structure of the interactions among the entities. But it cannot ascertain the specific mechanisms which underlie behind the interactions. So the lack of intelligibility comes from the inability of algorithms to learn those specific mechanisms rather than a dissimilarity between the learned interaction structure and the real one. There are other algorithms such as Prediction Analysis of Microarray (PAM, Tibshirani et al. 2002), which do not intend to learn the biological structure in any way, since they are aimed to prediction only. When algorithms like PAM are used, it means that scientists are not particularly interested in the structures, but in the predictions. In other words, when the biological network under investigation is too complex to obtain an explanatory mechanistic model, then the only option is to use a black box prediction algorithm which does not intend to learn the structure of the analyzed network.
Let us clarify again: this tradeoff between explaining and predicting is a consequence of the way machine learning deals with the bias-variance tradeoff. In other words, this tradeoff and the bias-variance tradeoff are different tradeoffs.
For network modeling in molecular biology, the number of variables is fixed by the number of detected compounds, so that there is no flexibility to choose the size of the model to be learned.
For a more thorough exposition of the structure of TCGA studies, see Ratti (2015).
The reason for doing this is that, when mechanistic sketches in this field are outlined, usually genes and mutations at the genetic level are considered central, and hence we decided to focus only on this important level because the epistemic processes applied to it are similar to the ones applied to other levels. Hence the analysis of other levels may be redundant.
Please note that experimental validation is different from the validation phase of machine learning procedures, where an algorithmic model is chosen according to its performance on a validation set, which is a subset of the available dataset.
One may also say that this is the point where ‘data-intensive’ studies meet mechanistic studies, in the sense that studies such as the ones of TCGA are only one step towards the elaboration of mechanistic explanations. However, this observation may miss the importance and the role of bioinformatics tools. These are not just tools aimed at selecting a few entities by means of eliminative inductive strategies, but tools that are used to characterize the complexity of biological systems. In fact, we may use these tools to prioritize a few entities and elaborate a simple mechanistic models, but by doing this we would miss the complexity of biological systems and the other analyses that can be done on biological complexity without necessarily narrowing down just a small and local part of it, as we do when we just focus on a few cancer genes as part of the complexity of a particular type of tumor.
This aspect may be interpreted as being related to the pathway concept, as Ross (2018) points out when she says that “instead of identifying a particular explanatory target and ‘drilling down’, these maps [i.e. pathways representation] involve identifying a set of entities in some domain and ‘expanding our by tracing their causal connections” (p. 13).
One may argue that mechanistic philosophers’ requirements for a good explanation are in tension, but this is beyond the scope of the present paper.
This of course does not mean that more traditional forms of molecular biology cannot possibly coexist with machine learning-driven molecular biology.
References
Akbani, R., et al. (2015). Genomic classification of cutaneous melanoma. Cell, 161(7), 1681–1696.
Alberts, B. (2012). The end of “small science”? Science, 337(6102), 1583.
Bechtel, W., & Abrahamsen, A. (2005). Explanation: A mechanist alternative. Studies in History and Philosophy of Science Part C: Studies in History and Philosophy of Biological and Biomedical Sciences, 36(2), 421–441. https://doi.org/10.1016/j.shpsc.2005.03.010.
Bechtel, W., & Richardson, R. (2010). Discovering complexity—Decomposition and localization as strategies in scientific research. Cambridge, MA: The MIT Press.
Bertolaso, M. (2016). Philosophy of cancer. Dordrecht: Springer.
Bishop, C. M. (2006). Pattern recognition and machine learning. New York: Springer.
Boem, F., & Ratti, E. (2016). Towards a notion of intervention in big-data biology and molecular medicine. In G. Boniolo & M. Nathan (Eds.), Foundational issues in molecular medicine. London: Routledge.
Boniolo, G., & Campaner, R. (2018). Molecular pathways and the contextual explanation of molecular function. Biology & Philosophy, 33(3–4), 1–19. https://doi.org/10.1007/s10539-018-9634-2.
Callebaut, W. (2012). Scientific perspectivism: A philosopher of science’s response to the challenge of big data biology. Studies in History and Philosophy of Biological and Biomedical Sciences, 43(1), 69–80. https://doi.org/10.1016/j.shpsc.2011.10.007.
Carrier, M. (2014). Prediction in context: On the comparative epistemic merit of predictive success. Studies in History and Philosophy of Science Part A, 45(1), 97–102. https://doi.org/10.1016/j.shpsa.2013.10.003.
Chang, H. (2014). Epistemic activities and systems of practice: Units of analysis in philosophy of science after the practice turn. In L. Soler, S. Zwart, M. Lynch & V. Israel-Jost (Eds.), Science after the practice turn in the philosophy, history and social studies of science. Routledge.
Cox, D. R. (2001). Comment to ‘statistical modeling: The two cultures’. Statistical Science, 16(3), 216–218.
Craver, C. F. (2006). When mechanistic models explain. Synthese, 153(3), 355–376. https://doi.org/10.1007/s11229-006-9097-x.
Craver, C. (2007). Explaining the brain - Mechanisms and the mosaic unity of neuroscience. Oxford University Press.
Craver, C., & Darden, L. (2013). In search of mechanisms. Chicago: The University of Chicago Press.
De Regt, H. W. (2009). The epistemic value of understanding. Philosophy of Science, 76(5), 585–597. https://doi.org/10.1086/605795.
De Regt, H. W. (2015). Scientific understanding: Truth or dare? Synthese, 192(12), 3781–3797. https://doi.org/10.1007/s11229-014-0538-7.
De Regt, H. (2017). Understanding scientific understanding. Oxford: Oxford University Press.
Dhar, V. (2013). Data science and prediction. Communications of the ACM, 56(12), 64–73.
Douglas, H. E. (2009). Reintroducing prediction to explanation. Philosophy of Science, 76(4), 444–463. https://doi.org/10.1086/648111.
Douglas, H., & Magnus, P. D. (2013). State of the field: Why novel prediction matters. Studies in History and Philosophy of Science Part A, 44(4), 580–589. https://doi.org/10.1016/j.shpsa.2013.04.001.
Frické, M. (2015). Big data and its epistemology. Journal of the Association for Information Science and Technology, 66(4), 651–661.
Gerlee, P., & Lundh, T. (2016). Scientific models. Basel: Springer.
Glennan, S. (2017). The new mechanical philosophy. Oxford: Oxford University Press.
Golub, T. (2010). Counterpoint: Data first. Nature, 464(7289), 679. https://doi.org/10.1038/464679a.
Halevy, A., Norvig, P., & Pereira, F. (2009). The unreasonable effectiveness of data. IEEE Intelligent Systems, 24(2), 8–12.
Hanahan, D., & Weinberg, R. A. (2000). The hallmarks of cancer. Cell, 100(1), 57–70.
Hanahan, D., & Weinberg, R. A. (2011). Hallmarks of cancer: The next generation. Cell, 144(5), 646–674.
Hastie, T., Tibshirani, R., & Friedman, J. (2009). The elements of statistical learning (2nd ed.). New York: Springer.
Hempel, C., & Oppenheim, P. (1948). Studies in the logic of explanation. Philosophy of Science, 15(2), 135–175.
James, G., Witten, D., Hastie, T., & Tibshirani, R. (2013). An introduction to statistical learning. New York: Springer.
Junqué de Fortuny, E., Martens, D., & Provost, F. (2013). Predictive modeling with big data: Is bigger really better? Big Data, 4(1), 215–226.
Kaplan, D. M., & Craver, C. F. (2011). The explanatory force of dynamical and mathematical models in neuroscience: A mechanistic perspective. Philosophy of Science, 78(4), 601–627. https://doi.org/10.1086/661755.
Keller, E. F. (2002). Making sense of life: Explaining biological development with models, metaphors and machines. Cambridge, MA: Harvard University Press.
Latour, B. (1987). Science in action. Cambridge, MA: Harvard University Press.
Leonelli, S. (2011). Packaging data for re-use: Databases in model organism biology. In P. Howlett & M. S. Morgan (Eds.), How well do facts travel? The dissemination of reliable knowledge. Cambridge, MA: Cambridge University Press.
Leonelli, S. (2012). Introduction: Making sense of data-driven research in the biological and biomedical sciences. Studies in History and Philosophy of Biological and Biomedical Sciences, 43(1), 1–3. https://doi.org/10.1016/j.shpsc.2011.10.001.
Leonelli, S. (2016). Data-centric biology. Chicago: University of Chicago Press.
Levins, R. (1966). The strategy of model building in population biology. In E. Sober (Ed.), Conceptual issues in evolutionary biology (pp. 18–27). Cambridge, MA: MIT Press.
Levy, A. (2014). What was Hodgkin and Huxley’s achievement? British Journal for the Philosophy of Science, 65(3), 469–492. https://doi.org/10.1093/bjps/axs043.
Levy, A., & Bechtel, W. (2013). Abstraction and the organization of mechanisms. Philosophy of Science, 80(2), 241–261. https://doi.org/10.1086/670300.
Lombrozo, T. (2011). The instrumental value of explanations. Philosophy Compass, 6(8), 539–551. https://doi.org/10.1111/j.1747-9991.2011.00413.x.
Love, A. C., & Nathan, M. J. (2015). The idealization of causation in mechanistic explanation. Philosophy of Science, 82(December), 761–774. https://doi.org/10.1086/683263.
Machamer, P., Darden, L., & Craver, C. (2000). Thinking about mechanisms. Philosophy of Science, 67, 1–25.
Matthewson, J., & Weisberg, M. (2008). The structure of tradeoffs in model building. Synthese, 170(1), 169–190. https://doi.org/10.1007/s11229-008-9366-y.
Morange, M. (1998). A history of molecular biology. Cambridge, MA: Harvard University Press.
Morgan, M., & Morrison, M. (Eds.). (1999). Models as mediators. Cambridge, MA: Cambridge University Press.
Pietsch, W. (2015). Aspects of theory-ladenness in data-intensive science. Philosophy of Science, 82(5), 905–916.
Press, G. (2013). A very short history of data science. Forbes. http://www.forbes.com/sites/gilpress/2013/05/28/a-very-short-history-of-data-science/. Accessed 12 June 2016.
Ratti, E. (2015). Big data biology: Between eliminative inferences and exploratory experiments. Philosophy of Science, 82(2), 198–218.
Ratti, E. (2016). The end of “small biology”? Some thoughts about biomedicine and big science. Big Data & Society, no. July–December:1–6.
Ratti, E., & López-Rubio, E. (2018). Mechanistic models and the explanatory limits of machine learning. In [2018] PSA 2018: The 26th Biennial meeting of the philosophy of science association (Seattle, WA; 1–4 November 2018). http://philsci-archive.pitt.edu/view/confandvol/confandvolPSA2018.html.
Rice, C. C. (2016). Factive scientific understanding without accurate representation. Biology and Philosophy, 31(1), 81–102. https://doi.org/10.1007/s10539-015-9510-2.
Ross, L. N. (2018). Causal concepts in biology: How pathways differ from mechanisms and why it matters. [Preprint]. http://philsci-archive.pitt.edu/id/eprint/14432. Accessed March 13, 2018.
Shmueli, G. (2010). To explain or to predict? Statistical Science, 25(3), 289–310.
Sloan, P. (2000). Completing the tree of descartes. In P. Sloan (Ed.), Controlling our destinies—Historical, philosophical, ethical, and theological perspectives on the human genome project. Notre Dame: University of Notre Dame Press.
Spirtes, P., Glymour, C., & Scheines, R. (1993). Causation, prediction, and search. New York: Springer.
Stevens, H. (2013). Life out of sequence—A data-driven history of bioinformatics. Chicago: Chicago University Press.
Stevens, H. (2015). Networks: Representations and tools in postgenomics. In S. Richardson & H. Stevens (Eds.), Postgenomics—Perspective on biology after the genome. Durham: Duke University Press.
Stevens, H. (2017). A feeling for the algorithm: Working knowledge and big data in biology. Osiris, 32(1), 151–174. https://doi.org/10.1086/693516.
Strasser, B. (2011). The experimenter’s museum—GenBank, natural history, and the moral economies of biomedicine. Isis, 102(1), 60–96.
Strevens, M. (2008). Depth—An account of scientific explanation. Cambridge, MA: Harvard University Press.
Sugiyama, M. (2015). Introduction to statistical machine learning. Burlington, MA: Morgan Kaufmann.
Tabery, J., Piotrowska, M., & Darden, L. (2015). Molecular biology. In E. N. Zalta (Eds.), The stanford encyclopedia of philosophy (Summer 2018 Edition).
Tibshirani, R., Hastie, T., Narasimhan, B., & Chu, G. (2002). Diagnosis of multiple cancer types by shrunken centroids of gene expression. Proceedings of the National Academy of Sciences of the United States of America, 99(10), 6567–6572.
Vaske, C. J., Benz, S. C., Sanborn, J. Z., Earl, D., Szeto, C., Zhu, J., et al. (2010). Inference of patient-specific pathway activities from multi-dimensional cancer genomics data using PARADIGM. Bioinformatics, 26(12), i237–i245.
Weinberg, R. A. (1985). The molecules of life. Scientific American, 253(4), 48–57. https://doi.org/10.1038/scientificamerican1085-48.
Weinberg, R. (2010). Point: Hypotheses first. Nature, 464(7289), 678. https://doi.org/10.1038/464678a.
Weinberg, R. A. (2014). Coming full circle-from endless complexity to simplicity and back again. Cell, 157(1), 267–271. https://doi.org/10.1016/j.cell.2014.03.004.
Weisberg, M. (2006). Forty years of “the strategy”: Levins on model building and idealization. Biology and Philosophy, 21(5), 623–645. https://doi.org/10.1007/s10539-006-9051-9.
Acknowledgements
The authors would like to thank David Teira, Enrique Alonso, and the participants to the workshop “Making sense of data in the sciences” in Hannover, and in particular Federica Russo and Sara Green, for their valuable comments and suggestions. They are also grateful to the editor and four anonymous reviewers for their constructive feedback.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Both authors have contributed equally to this work.
Electronic supplementary material
Below is the link to the electronic supplementary material.
Rights and permissions
About this article
Cite this article
López-Rubio, E., Ratti, E. Data science and molecular biology: prediction and mechanistic explanation. Synthese 198, 3131–3156 (2021). https://doi.org/10.1007/s11229-019-02271-0
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11229-019-02271-0