From Sequence Alignment to Generative Molecules

1. The evolution of bioinformatics and computational biology, 2000–2026

2000–2007: the reference era. The draft human genome was published in 2001 [1]. It moved biology from studying single genes to studying whole genomes. The main tools were built on alignment and statistics. BLAST [2] for homology search, hidden Markov models for gene and domain finding, and phylogenetics shaped annotation pipelines. Structural bioinformatics used homology modelling and molecular mechanics force fields such as AMBER and CHARMM. Structure-based drug design meant rigid or semi-flexible docking against crystal structures, with empirical scoring functions.

2007–2015: the throughput era. Next-generation sequencing cut the cost per genome by several orders of magnitude. Data volume became the main problem [3]. The field shifted to read alignment (Burrows–Wheeler indexing), variant calling, RNA-seq quantification, ChIP-seq and GWAS. Systems biology added network inference and constraint-based metabolic modelling. Molecular dynamics became routine on GPUs, and free-energy perturbation (FEP) began to predict relative binding affinities with useful accuracy.

2015–2020: the representation era. Single-cell transcriptomics and atlas projects [4] produced high-dimensional, sparse data. This needed manifold learning, batch correction and probabilistic latent-variable models. Deep learning entered genomics through convolutional models of regulatory sequence and graph neural networks on molecular graphs. Computational biology started to become a machine-learning discipline.

2020–2026: the foundation-model era. AlphaFold2 [5] and RoseTTAFold [6] solved single-chain protein structure prediction for most practical uses. They combined multiple-sequence-alignment embeddings with attention-based geometric reasoning. Protein language models such as ESM-2/ESMFold [7] showed that evolutionary constraints can be learned from sequence alone. AlphaFold 3 [8] extended prediction to protein–ligand, protein–nucleic acid and modified-residue complexes using a diffusion-based architecture. RFdiffusion [9] made de novo design of binders and scaffolds routine. The 2024 Nobel Prize in Chemistry recognised computational protein design and structure prediction [10]. By 2026, the frontier is multimodal models that combine sequence, structure, omics and phenotypic imaging. The toolset has changed from mainly descriptive to mainly generative.

2. AI/ML versus traditional drug design and development

The traditional pipeline runs target identification, hit finding (HTS or fragment screening), hit-to-lead, lead optimisation, preclinical studies and clinical phases I–III. It takes about 10–15 years and costs well over US$1 billion per approved drug, once failures are counted [11]. Only about 10% of drugs entering phase I reach approval [12].

AI/ML competes at four points:

• Target discovery. Knowledge graphs, causal inference on human genetics, and perturbation-omics models rank targets by genetic support. Targets with human genetic evidence have measurably higher clinical success rates.

• Hit finding. Physics-based ultra-large docking already screened hundreds of millions of compounds [13]. ML surrogates (active learning, docking-score regressors) now triage multi-billion-compound make-on-demand libraries at a fraction of the compute. Deep learning found halicin, a structurally novel antibiotic, through a virtual screen [14].

• Design and optimisation. Generative models (VAEs, reinforcement-learning-guided SMILES/graph generators, equivariant diffusion over 3D coordinates) propose molecules conditioned on the pocket and on multi-parameter objectives (potency, selectivity, ADMET, synthesisability). Closed design–make–test–analyse loops with automated chemistry shorten cycle times.

• Development. ML supports ADMET and toxicity prediction, patient stratification, trial-site selection and synthetic control arms.

The evidence so far. An analysis of AI-derived molecules found phase I success rates of roughly 80–90%, well above historical averages. Phase II success was similar to conventional programmes, though the sample was small [15]. Insilico Medicine's TNIK inhibitor rentosertib, for idiopathic pulmonary fibrosis, went from AI-nominated target to phase 2a readout with published results [16, 17]. This is an important proof of concept, not proof of higher approval rates. The consistent pattern is that AI clearly compresses discovery timelines and improves early-stage molecular quality. It has not yet shown that it reduces the efficacy failures in phase II that dominate attrition. Those failures come mostly from the wrong biology, not bad chemistry.

3. AI/ML compared with classical bioinformatics and computational biology

The distinction is less "old versus new" and more inductive bias versus learned representation.

In drug discovery the two are converging, not competing. AlphaFold models are trained on the Protein Data Bank and MSAs, both products of 25 years of bioinformatics. Co-folding predictions are now checked with physics-based relaxation, MD and FEP. Hybrid approaches include ML force fields (neural network potentials near quantum accuracy), ML-accelerated FEP, and generative design scored by docking or free-energy calculations. For example, a diffusion model proposes and physics filters. Classical methods still set the ground truth for mechanism, and ML supplies speed and chemical-space coverage.

Important limitations remain. Co-folding models can produce physically implausible poses and may memorise training ligands. Benchmark leakage between training and test sets inflates reported performance. Public bioactivity data are noisy and biased toward well-studied targets. Regulators now expect credibility assessment of AI models tied to their context of use [18].

4. Outlook

The next step is not a single model but integrated platforms. These combine human genetics for target confidence, foundation models for structure and design, physics for validation, and automated labs for rapid experimental feedback. The deciding metric is not how many molecules are designed. It is whether AI can raise the probability of clinical efficacy.

By 2026, that answer is still open.

References

1. _{^{Lander ES, et al. Initial sequencing and analysis of the human genome. Nature. 2001;409:860–921.}}

2. _{^{Altschul SF, et al. Basic local alignment search tool. J Mol Biol. 1990;215:403–410.}}

3. _{^{Stephens ZD, et al. Big Data: Astronomical or Genomical? PLoS Biol. 2015;13:e1002195.}}

4. _{^{Regev A, et al. The Human Cell Atlas. eLife. 2017;6:e27041.}}

5. _{^{Jumper J, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596:583–589.}}

6. _{^{Baek M, et al. Accurate prediction of protein structures and interactions using a three-track neural network. Science. 2021;373:871–876.}}

7. _{^{Lin Z, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science. 2023;379:1123–1130.}}

8. _{^{Abramson J, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024;630:493–500.}}

9. _{^{Watson JL, et al. De novo design of protein structure and function with RFdiffusion. Nature. 2023;620:1089–1100.}}

10. _{^{The Nobel Prize in Chemistry 2024 (Baker; Hassabis and Jumper). NobelPrize.org.}}

11. _{^{Paul SM, et al. How to improve R&D productivity: the pharmaceutical industry's grand challenge. Nat Rev Drug Discov. 2010;9:203–214.}}

12. _{^{Wong CH, Siah KW, Lo AW. Estimation of clinical trial success rates and related parameters. Biostatistics. 2019;20:273–286.}}

13. _{^{Lyu J, et al. Ultra-large library docking for discovering new chemotypes. Nature. 2019;566:224–229.}}

14. _{^{Stokes JM, et al. A deep learning approach to antibiotic discovery. Cell. 2020;180:688–702.}}

15. _{^{Jayatunga MKP, et al. How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons. Drug Discov Today. 2024;29:104009.}}

16. _{^{Ren F, et al. A small-molecule TNIK inhibitor targets fibrosis in preclinical and clinical models. Nat Biotechnol. 2025;43:63–75.}}

17. _{^{Xu Z, et al. A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial. Nat Med. 2025.}}

18. _{^{US FDA. Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products (Draft Guidance). January 2025.}}

19. _{^{Vamathevan J, et al. Applications of machine learning in drug discovery and development. Nat Rev Drug Discov. 2019;18:463–477.}}

Disclaimer

_{^{This article is for general educational and informational purposes only. It does not constitute scientific, medical, regulatory, legal or investment advice. It does not endorse any company, product, platform or therapeutic candidate. Descriptions of technologies, clinical programmes and regulatory positions reflect published literature as understood at the time of writing and may have changed. Figures on clinical success rates and costs are estimates from cited studies and depend on their methods and samples. Readers should consult the original sources and qualified professionals before relying on any information here. The views expressed are the author's own and do not represent any employer or affiliated organisation.}}