History of AI in the Life Sciences
Computational biology has used learning algorithms for decades. The story is not a single breakthrough at AlphaFold 2 in 2020 but an arc from dynamic-programming sequence alignment to generative biology, with each generation inheriting methods, tasks, and failure modes from earlier work. Reading the field historically helps separate a genuinely new capability from a new implementation of an established biological problem.
Use this chapter to:
- Place current AI-biology systems in the longer history of computational biology.
- See why shared data, benchmarks, and biological tasks changed which methods mattered.
- Use history to calibrate new model claims without treating novelty as proof.
Summary: AI in biology did not begin with AlphaFold. The current foundation-model era grew from sequence alignment, profile HMMs, supervised genomics, deep regulatory models, and decades of structure-prediction benchmarks.
Key point: Methods become important when data, benchmarks, and biological decisions line up. Older methods rarely disappear; they become infrastructure for the next generation.
Bottom line: Use this chapter to read new capability claims against their predecessors. If a field lacks a blind benchmark, shared data standard, or clear failure mode, treat broad claims as early even when the model is new.
What is this field trying to solve? Place current AI-biology claims in the longer history of computational biology.
What is the core idea? Each generation inherits biological tasks, benchmarks, data constraints, and failure modes from earlier methods.
What is the current state of the field? Sequence alignment, profile models, supervised learning, deep learning, and generative systems remain active layers rather than a sequence of replacements.
What do we know, and what remains open? Blind benchmarks have established selected capability changes. Many fields still lack comparable evaluation cultures.
Why does this matter? Historical context helps distinguish a new capability from a new implementation or label.
Introduction
The conventional public narrative places the start of AI in biology in late 2020, when AlphaFold 2 produced single-chain protein structure predictions whose accuracy was comparable to experimental measurement on the CASP14 blind benchmark (Jumper et al., 2021). That date is right for the moment when biology entered the broad public consciousness as a field where deep learning would matter. It is wrong as a start date for the field.
Computational biology has used data-derived scoring and statistical learning for decades. Early practical tools included dynamic-programming sequence aligners; profile hidden Markov models later became widely used for protein-family modelling, homology search, and gene prediction. Support vector machines, random forests, logistic regression, and related methods were widely applied across genomics in the 2000s. Deep learning entered biology in the 2010s, found early traction in regulatory genomics, and produced high-profile advances in variant calling before AlphaFold 2. AlphaFold 2 was an inflection in which problems deep learning could solve, not in whether it should be used at all.
Reading the field requires reading this history. Three reasons:
Many capability claims have a prior generation. “End-to-end learning for transcription factor binding” is a newer framing of a problem previously approached with position weight matrices and other statistical models. Current systems may gain from scale, representation, and architecture while retaining earlier problems with confounding, measurement, and transfer. Knowing the prior generation’s failure modes is a practical way to identify where current claims may break.
The strongest breakthrough claims survive independent evaluation. CASP is the clearest example of a blinded community benchmark. MoleculeNet is a public benchmark suite, not a blind challenge, and BLAST scoring is a method rather than a benchmark. The transferable lesson is to match evaluation design to the claim instead of treating every leaderboard as equivalent.
New methods do not erase established infrastructure. Sequence-database growth encouraged faster search and indexing, while deep models added representations that can capture longer-range context. BLAST, profile HMMs, and classical statistical models remain useful because they answer different questions and offer different tradeoffs. The current generation’s limits are already visible; what follows it is not obvious.
The rest of this chapter walks each era in turn. The aim is not a complete bibliography. The aim is to give a reader who knows AI but not biology, or biology but not AI, enough context to read a 2026 capability claim against a 2006 one and tell which lessons still apply.
Timeline of AI in Biology
The milestones below show how computational biology became life sciences AI. The sequence is cumulative: later systems changed scale, representation, and experimental reach, while retaining earlier databases, search methods, statistical assumptions, and evaluation practices.
| Year | Milestone | Why it matters |
|---|---|---|
| 1970 | Needleman-Wunsch global alignment (Needleman and Wunsch, 1970) | Dynamic programming provided an exact method for aligning biological sequences under a defined scoring scheme. |
| 1981 | Smith-Waterman local alignment (Smith and Waterman, 1981) | Local alignment made it possible to identify related regions inside otherwise different sequences. |
| 1990 | BLAST (Altschul et al., 1990) | Heuristic search made rapidly growing sequence databases practically searchable. |
| 1997–1998 | GENSCAN and profile-HMM methods (Burge and Karlin, 1997; Eddy, 1998) | Probabilistic sequence models supported gene prediction, protein-family modeling, and homology search. |
| 2015 | DeepBind and DeepSEA (Alipanahi et al., 2015; Zhou and Troyanskaya, 2015) | Deep learning began producing sustained benchmark gains in regulatory genomics. |
| 2016–2018 | DeepVariant and the precisionFDA challenge (Poplin et al., 2018) | A convolutional system achieved leading SNP performance in a public variant-calling challenge and transferred across evaluated sequencing technologies. |
| 2018–2020 | AlphaFold 1 at CASP13 (Senior et al., 2020) | Learned distance predictions moved protein-structure prediction beyond the previous generation while leaving substantial room for improvement. |
| 2020–2021 | AlphaFold 2 at CASP14 (Jumper et al., 2021) | A blinded community benchmark established a major advance in single-chain protein-structure prediction. |
| 2023 | RFdiffusion (Watson et al., 2023) | Diffusion-based protein generation was paired with experimental characterization of selected designs. |
| 2024 | AlphaFold 3 and the Nobel Prize in Chemistry (Abramson et al., 2024; Nobel Prize, 2024) | Interaction prediction expanded across biomolecular classes, while protein structure prediction and computational protein design received formal scientific recognition. |
The timeline is not a sequence of replacements. Alignment, search, profile models, supervised learning, structure prediction, and generative design remain different layers of contemporary workflows. The historical question is not which method won permanently, but which task, dataset, and evaluation made a method useful.
The Algorithmic Era (1970-1990)
Computational biology started before machine learning. A defining problem of the 1970s and 1980s was sequence comparison: given two protein or DNA sequences, identify their similarity and align them. The Needleman-Wunsch algorithm provided a dynamic-programming solution for global alignment (Needleman and Wunsch, 1970); the Smith-Waterman algorithm provided local alignment, which is useful when related regions appear within different flanking sequences (Smith and Waterman, 1981). These are not learning algorithms. They are exact methods that, given a scoring scheme for matches, mismatches, and gaps, compute an optimal alignment in quadratic time.
The scoring scheme was where statistics entered. Margaret Dayhoff’s PAM matrices and later the BLOSUM matrices were empirical: counts of how often each amino acid substituted for each other amino acid in evolutionarily related sequences. BLOSUM converted conserved blocks of related proteins into substitution scores that became a default language for protein comparison (Henikoff and Henikoff, 1992). These were data-derived scoring schemes, the proto-machine-learning of the field.
BLAST, the Basic Local Alignment Search Tool, made database-scale sequence search practical (Altschul et al., 1990). BLAST traded exhaustive optimal alignment for tractable heuristic search: it indexes query words, finds high-scoring seed matches, and extends them. BLAST remains widely used in bioinformatics pipelines, including workflows that prepare or inspect data for newer models.
The algorithmic era did not have “AI” in any modern sense, but it established a durable engineering lesson: a method’s practical value depends on accuracy, tractability, interpretability, and fit to the biological task. BLAST became infrastructure because its heuristic search made large sequence databases usable, not because it returned an exhaustive optimal alignment.
The Statistical Learning Era (1990-2012)
By the mid-1990s, the practical problem had shifted from comparing two sequences to modelling families of sequences. A protein family shares ancestry but diverges over time; positions in the family vary in how conserved they are, and the variation itself carries information about function. Profile hidden Markov models (HMMs), formalised in tools like HMMER and SAM, captured this position-specific conservation (Eddy, 1998). Pfam subsequently became a widely used database of protein families represented by profile HMMs.
The HMM era also drove gene prediction. GENSCAN used a generalised HMM to predict gene structure in genomic DNA, integrating signal-detection components (splice sites, start codons) into a probabilistic framework that could be trained on annotated genomes (Burge and Karlin, 1997). When the human genome was published in 2001, HMM-based annotation pipelines were the standard.
In parallel, microarrays produced large gene-expression datasets, and the field absorbed the supervised-learning toolkit from broader machine learning. Support vector machines were applied to cancer classification, gene-function prediction, and protein-fold recognition. Random forests, logistic regression, and naïve Bayes were also used across biological prediction tasks. The methods were general-purpose; biological domain knowledge entered through labels, features, study design, and interpretation.
Performance in this era often depended on feature engineering. To classify protein function with an SVM, researchers designed features such as amino-acid composition, predicted secondary structure, hydrophobicity profiles, and sequence-derived physicochemical properties. The approach made domain assumptions visible, but laborious feature design could also constrain what the classifier learned.
Deep Learning Enters Biology (2012-2018)
AlexNet won the ImageNet competition in 2012. Within three years, the architecture’s core idea, learning hierarchical features end-to-end from raw inputs, had migrated into biology.
Regulatory genomics was an early area of sustained activity. DeepBind (Alipanahi et al., 2015) used convolutional neural networks to predict sequence specificities and transcription-factor binding. DeepSEA extended the approach to chromatin features and non-coding variant effects at single-nucleotide resolution (Zhou and Troyanskaya, 2015). Basset learned chromatin-accessibility patterns across cell types (Kelley et al., 2016); DanQ added a recurrent layer to model motif interactions (Quang and Xie, 2016); Basenji moved toward longer-range, quantitative regulatory profiles (Kelley et al., 2018). Together, these studies showed that learned sequence representations could improve specific regulatory-genomics benchmarks, without establishing universal superiority over engineered features.
The next domain was variant calling, the routine task of detecting genetic variants from short-read sequencing data. DeepVariant (Poplin et al., 2018) reframed variant calling as an image-classification problem: read pileups became images, and a convolutional network classified candidate sites. DeepVariant won the highest SNP performance in the 2016 precisionFDA Truth Challenge and demonstrated strong performance across sequencing technologies.
The third domain was protein structure prediction. AlphaFold 1 (Senior et al., 2020, debuting at CASP13 in 2018) used a deep residual network over distogram predictions, producing distances between residue pairs that were then assembled into structures. The result was a substantial step beyond prior physics- and template-based methods, enough to win CASP13 by a meaningful margin. But CASP13 also showed that the architecture was a stepping stone: AlphaFold 1’s accuracy on novel folds was still well short of experimental, and competitors were closing on the lead.
The defining feature of this era was that deep learning’s role expanded across sequence-input tasks. The CNN-and-RNN era produced a series of task-specific benchmark improvements rather than one CASP14-level result. Effect sizes varied by dataset, metric, baseline, and split design.
The Generative Breakthrough (2018-2024)
The next era is the one most readers know about. It was compressed: a five-year window in which structure prediction, design, and biomolecular interaction prediction each crossed thresholds that had been considered hard for decades.
Structure prediction. AlphaFold 2 (Jumper et al., 2021) won CASP14 in late 2020 with accuracy competitive with experimental structures for much of the single-chain benchmark. RoseTTAFold (Baek et al., 2021) followed with a three-track architecture. ESMFold (Lin et al., 2023) showed that protein language models can predict structures without explicit multiple sequence alignments, trading average accuracy for substantially faster inference in the evaluated workflows.
Structural proteome at scale. The AlphaFold Protein Structure Database (Varadi et al., 2024) released over 214 million predicted structures, effectively the predicted structural proteome of life. This made structure a free input to every downstream method, replacing the assumption that structures were rare and expensive. The community assessment of AlphaFold 2’s utility across structural-biology workflows (Akdel et al., 2022) documented both the genuine impact and the persistent gaps: disordered regions, conformational ensembles, large complexes, and ligand-bound states remained hard.
Protein design. RFdiffusion (Watson et al., 2023) adapted diffusion models to protein-backbone generation. Given structural constraints such as a binding-site geometry or desired fold, the system generated candidate backbones, and the study experimentally characterized selected designs. The defensible evidence is task- and assay-specific: computational generation can produce experimentally testable candidates, but success rates do not transfer automatically across targets or design objectives.
Variant interpretation. AlphaMissense (Cheng et al., 2023) scored approximately 71 million possible human single-amino-acid substitutions, categorizing 57% as likely benign, 32% as likely pathogenic, and 11% as ambiguous. Those are research-model categories, not clinical variant classifications.
Biomolecular interactions. AlphaFold 3 (Abramson et al., 2024) extended prediction beyond proteins alone to interactions involving nucleic acids, ions, small-molecule ligands, and post-translationally modified residues. The initial release was server-only. DeepMind released inference code and model parameters for academic, noncommercial use in November 2024, while Boltz and Chai provided separately developed public models under different licenses (Abramson et al., 2024).
External recognition. The 2024 Nobel Prize in Chemistry was shared between Demis Hassabis and John Jumper for protein structure prediction and David Baker for computational protein design (Nobel Prize, 2024).
The pattern of this era is worth naming. Advances combined larger datasets, problem-specific architectures, useful representations, and evaluation or experimental validation. AlphaFold used attention-based components, RFdiffusion used diffusion modeling, and AlphaMissense adapted structure-model representations. Architecture labels alone do not explain why a biological result transfers.
The Foundation Model Era (2021-Present)
The current era is defined by the same architectural family that drives large language models, applied across biological data modalities. The defining property is pre-training: a model is trained on a very large unlabeled corpus with a self-supervised objective (masked-token prediction, next-token prediction, contrastive prediction), then fine-tuned or queried for specific downstream tasks.
Protein language models. ESM-2 (Meta, 2022) and ESM-3 (EvolutionaryScale, 2024) are trained on hundreds of millions to billions of protein sequences with a masked-language-modelling objective. The learned representations transfer to structure prediction, mutation-effect prediction, function annotation, and design. ESMFold (above) is the protein-structure application of the ESM-2 representation.
Genomic foundation models. Evo (Arc Institute, 2024) and related models train on whole-genome sequences with a next-token-style objective, producing representations that transfer to regulatory annotation, variant-effect prediction, and generative design of sub-genomic elements. The scaling laws look qualitatively similar to those in language modelling, though the question of whether genomic-scale generation is biologically meaningful in the same way text generation is meaningful remains open.
Single-cell foundation models. Geneformer (Theodoris et al., 2023), scGPT, and scFoundation are pre-trained on tens of millions of single-cell transcriptomes. They transfer to cell-type annotation, perturbation prediction, and trajectory inference. The cells, tissues, and systems biology chapters of this handbook treat their evaluation in detail; the short version is that they are clearly useful but the evidence that they exceed strong task-specific baselines is mixed.
Multimodal models. Boltz-2 (MIT, 2024-2025), AlphaFold 3, and Chai-1 increasingly combine structural, sequence, and chemical inputs in a single model. The current direction is toward biology-wide foundation models that handle sequence, structure, interaction, and modification in a unified representation.
The field is young: the term “biology foundation model” became routine in 2022-2023. Whether scaling will continue to deliver, whether emergent biological reasoning will appear in larger models, and whether the right pre-training objective for biology is the language-model objective at all are all open questions. The evidence is not yet conclusive.
What Fifty Years of History Tells Us
Three durable patterns emerge from the arc.
Blind benchmarks are the most credible evidence. CASP is the model: results submitted before targets are revealed, evaluated by independent assessors, published in a community-wide assessment. Every era’s breakthrough method has been validated this way: AlphaFold 1 at CASP13, AlphaFold 2 at CASP14, RoseTTAFold against AlphaFold 2 (with shared targets and consistent assessment). Capability claims in biology that have no equivalent blind-benchmark evidence are claims, not results. The cells and systems chapters, therapeutics chapters, and benchmarks chapter all describe what blind-benchmark coverage looks like in each subfield, and what areas still lack it.
Each era’s breakthrough becomes the next era’s preprocessing step. BLAST is still inside foundation-model training pipelines; profile HMMs are still inside Pfam classification; AlphaFold predictions are inputs to AlphaMissense, RFdiffusion, and structure-based design pipelines. The lesson is not that the old method was wrong; it is that each generation expands the substrate the next generation can use. Methods are not replaced; they are demoted to infrastructure.
The next breakthrough is difficult to forecast from the current paradigm. Transformers did not exist in 2010, and later diffusion and foundation-model approaches were not obvious continuations of earlier computational biology. The fields that have had a CASP14-style benchmark inflection provide one template. Other fields may advance through different combinations of data, experiments, and evaluation.
The handbook’s organisation reflects this history. The molecular-biology chapters cover problems where the breakthrough has happened. The cellular and systems-biology chapters cover problems where it has partly happened. The therapeutic-discovery chapters cover problems where translation is still the harder constraint than prediction. The automation chapters cover the laboratory infrastructure that any next breakthrough will depend on. Reading any one of those chapters against this history is a practical way to calibrate where a capability claim sits in the arc, and what the prior generation’s failure modes still imply.
How to Use This History
History is useful for calibration, not prediction. The demonstrated pattern is that AI in biology becomes consequential when a method, dataset, benchmark, and decision align. BLAST made sequence search practical. Profile HMMs made protein-family and gene-structure work routine. DeepVariant showed that deep learning could beat established variant-calling pipelines. AlphaFold 2 changed structure prediction because CASP exposed the improvement under blinded conditions. RFdiffusion and related design tools matter because laboratory validation turned generation into measurable design evidence.
The practical lesson is narrower than “the next model will solve the next problem.” It is that cellular perturbation, regulatory genomics, therapeutic translation, and closed-loop laboratories need their own benchmark cultures before broad claims deserve the same weight as structure prediction. Scaling, better data, and new architectures may improve biology models, but the historical record argues for benchmark discipline rather than architectural determinism.
For practitioners, three operational implications follow from this history:
Always ask which generation a method belongs to. A 2024 paper that uses an LSTM for a sequence task is using an older architecture. That is not automatically a problem (the data may suit it; baselines matter), but it is information. Conversely, a foundation-model approach to a problem with 200 labelled examples is an architecture choice without a data justification.
Match the evaluation to the claim. CASP is blinded. CAMEO is continuous. PoseBusters tests pose validity. BindingDB hold-outs and GuacaMol are public benchmark designs. These resources answer different questions and should not be collapsed into one evidence tier. The Benchmarks for Bio AI chapter separates them.
Treat “AI-powered” branding as a question, not an answer. Every era of this history has had branding that absorbed the previous era’s methods under the new label. “AI-powered” pipelines in 2026 routinely include BLAST, profile HMMs, and CNN classifiers from earlier eras. The question is whether the methods that justify the label produce the outcomes the claim implies. The Evaluation Principles for Life Sciences AI chapter sets out a workable framework for that question.