The AI for Scientific Discovery Network+

The AI3SD Network+ was created to accelerate scientific discovery using new AAI techniques. Despite useful short-term impacts, this extensive space still has many unsolved challenges. It is vital that the trajectory of effort continues, encouraging further collaborations, helping funding recipients realize their potential, and creating resources for community. A core aspect our philosophy importance human element in AI, hence why mantra. Humans must remain loop any decision making requires ethical consideration or addition intelligence. approach combines best machine intelligence gain notable acceleration transparent, responsible, explainable AAI. We aim finish first term with a rich collection outputs, enabling members progress toward UN sustainability goals areas chemicals materials discovery. will sustainable entity only if organizers work together Artificial Intelligence Augmented Automated Investigation Scientific Discovery (AI3SD) established response UK Engineering Physical Sciences Research Council (EPSRC) late-2017 call promote cutting-edge research artificial groundbreaking discoveries. This article provides philosophical, scientific, technical underpinnings Network+, history different domains represented specific focus Network+. activities, covered year have highlighted significant challenges chemistry augmented space. These are shaping future directions concludes summary lessons learned running introduces plans landscape redrawn by COVID-19, including rebranding into AI 4 Network (www.ai4science.network). In perspective we describe inception (AI3SD1ai3-science-network [Internet].http://www.ai3sd.org/homeGoogle Scholar), funded (EPSRC2Home - EPSRC website [Internet].https://epsrc.ukri.org/Google Scholar) Innovation (UKRI3Home [Internet].https://www.ukri.org/Google late 2017, announced call4Automating Science Call Plus [Internet].https://epsrc.ukri.org/funding/calls/automatingsciencediscoverycallfornetworkplus/Google Scholar (AI) Our bid concentrated on chemical discovery, as these which had familiarity fields where it apparent major developments industry academia were already underway, offering huge potential sure discoveries techniques arising from target be transferable wider life sciences beyond. successful, awarded grant (EP/S000356/5EPSRCArtificial Grant [Internet].https://gow.epsrc.ukri.org/NGBOViewGrant.aspx?GrantRef=EP/S000356/1Google able due course appoint very network coordinator, past experience IT Utility (ITaaU) Network+6IT | Championing interdisciplinary collaboration digital economy [Internet].http://www.itutility.ac.uk/Google shown an essential difficult, multi-faceted role. Alongside appointment, also recruited diverse extremely experienced advisory board academic industrial institutions (the current all listed website7Advisory Board ai3-science-network [Internet].http://www.ai3sd.org/ai3sd-team/advisory-boardGoogle Scholar). note some amusement at Advisory meeting pointed out missed fact added fourth then could become AI4Science highly convenient abbreviation inspired us plan change name move more self-sustaining existence. While retain www.ai3sd.org URL, http://www.AI4Science.Network increasingly used point activities (further coincidences occurred because, checking, found URL www.ai4science.org not available, therefore obtained www.ai4science.network, although subsequent discovered held one own farsighted who present meeting!). vision centered community-wide engagement cutting edge both focused simply applying known learning data-mining problems two problem but instead growing truly collaborative environment multi-way interactions. Underlying emphasis recognition intelligent algorithms should augmenting goal (as described Box 1 Figure 1). exploit archived knowledge formulation feed higher-level symbolic representations, addressing philosophical issues causality uncertainty complex reasoning systems. Such combined activity, when deployed, AI8Frankish K. Ramsey W.M. Cambridge Handbook Intelligence. University Press, 2014: 367Google Scholar,9Goebel R. Chander A. Holzinger Lecue F. Akata Z. Stumpf S. Kieseberg P. Explainable AI: New 42?.in: Tjoa A.M. Weippl E. Machine Learning Knowledge Extraction. Springer International Publishing, Cham (Switzerland)2018: 295-303Crossref Scopus (129) Google its foundation.Box 1DefinitionsIn refer (AAI). By AAI, mean application mathematical, algorithmic, considerations underpinning computational methods answering questions.Artificial refers complement augment intelligence, encompassing notions evolution context datasets.Augmented algorithmic assist solving, essentially extraction information large datasets, may structured.Scientific process product conducting inquiry, results add body knowledge. AI3SD, chosen view identification regular relationships underlying causes given domain model-based reasoning, verification them experimentation, while respecting societal constraints ethics philosophy. questions. datasets. structured. perspectives provide background elements (scientific AI), detail formation date, reflections so far how forward rest Network+’s time span. To set scene review nature evolved, least UK, since enlightenment period10Burns W.E. Enlightenment: An Encyclopedia. ABC-CLIO, 2003: 385Google Scholar, place disruption bring science. Beginning Bacon, simple inductivist viewed driven data. However, description actually been made, does ring true. As Whewell noted 1860,11Yeo Defining Science: William Whewell, Natural Public Debate Early Victorian Britain. Vol. 27. 2003Google Scholar,12Whewell W. On Philosophy Discovery: Chapters Historical Critical. J.W. Parker son, 1860: 598Google normally proceed unreasoning data novel hypotheses. Efficient usually experimentation directed understanding would affect understanding. Bringing order judgements hypothesis direct no easy task. For reason, philosophers science often distinguished between contexts made justification.13Popper Logic Discovery. Routledge, 2005: 545Google 14Reichenbach H. Experience Prediction: Analysis Foundations Structure Knowledge. Chicago 1938: 410Google 15Hoyningen-Huene varieties distinction justification.Stud. Hist. Philos. Sci. 1987; 18: 501-515Crossref (79) justifying clearly careful collection, ordering, analysis data, historically chaotic. owes much judgements, intuitions, hunches, creativity scientists involved, well chance encounters, proclivities, values, social settings. According Mendeleev, he literally dreamt up periodic table. theory special relativity supposedly came Einstein stared idly window moving tram. Intuition seems key. Benjamin16Benjamin A.C. mystery JSTOR.Philos. 1934; 1: 224-236Crossref put 1934, “act intuition … arise most readily control conspicuously absent”. Insights investigation. Their genesis offers explanation can unpredictable.17Kuhn T.S. structure discovery.Science. 1962; 136: 760-764Crossref PubMed (135) Insight precedes formal justification, cannot predicted basis mass scientist presented. attempts make use justification (for example, meta-DENDRAL18Buchanan B. Feigenbaum Lederberg J. Heuristic Programming Study Theory Formation Science., in: Second Joint Conference Intelligence.Presented Conferences Intelligence, London. 1971; : 40-48Google Scholar,19Lindsay R.K. Buchanan B.G. E.A. DENDRAL: case study expert system formation.Artif. Intell. 1993; 61: 209-261Crossref (107) treated something formalized; automatic algorithmic. Empirically, however, systems enjoy niche laboratory, tally skepticism even analytic whether Even Hempel20Hempel C.G. Thoughts limitations computer.in: Diagnosis Medicine. 1985: 115-122Google trying pin down inductive “logic confirmation,” argued capable genuine For, argued, real involves introduction concepts theoretical principles such embedded. classical programming architectures impose frameworks. So computers exceed pace they test theories, beyond them, rendering concept introduced 1962 Kuhn17Kuhn unable shift paradigms. Hempel's finally seem outdated, however. There recently changes brute power various come under banner learning. automated inspect vast quantities along dimensions allows identify surprising correlations, akin Bacon's inductivism. Doing puts frame genuinely capacity sense. They unnoticed patterns anomalies just within firmly Either way, manifests tool philosophies mind. With mind, now reflect emergence current, third wave UK. several waves followed hype, winter 2).19Lindsay Scholar,37Bringsjord Govindarajulu N.S. intelligence.in: Zalta E.N. Stanford Encyclopedia Philosophy. Metaphysics Lab, University, 2020https://plato.stanford.edu/archives/sum2020/entries/artificial-intelligence/Google 38Duan Y. Edwards J.S. Dwivedi Y.K. era Big Data – evolution, agenda.Int. Inf. Manag. 2019; 48: 63-71Crossref (420) 39Dietterich T.G. Steps robust intelligence.AI Magazine. 2017; 38: 3-24Crossref (50) Scholar,40Brannigan attributional model discovery.Soc. Stud. 1989; 19: 601-613Crossref (7) Scholar,41Hey T. Tansley Tolle Fourth Paradigm: Data-Intensive Microsoft Research, 2009: 287Google Scholar,21Shortliffe E.H. Mycin: knowledge-based computer program applied infectious diseases.Proc. Annu. Symp. Comput. Appl. Med. Care. 1977; 66-69Google 22Myers J.D. Pople H.E. Miller R.A. CADUCEUS: computerized diagnostic consultation internal medicine.Proc. 1982; 44-47Google 23Allgood Rodriguez O. Bédorf Fournier P.-A. Kadurin Zhavoronkov MacKinnon Rosengarten Kremliovsky M. Chang et al.Artificial Healthcare. AAIH: Alliance Healthcare, 2019: 45Google 42Todd M.H. Computer-aided organic synthesis.Chem. Soc. Rev. 2005; 34: 247-266Crossref (114) 43Noordik J.H. CAOS/CAMM services: integrated assisted tools.in: Gauglitz G. Software-Entwicklung der Chemie 3 [Internet]. Berlin Heidelberg, 1989: 449-465http://link.springer.com/10.1007/978-3-642-74373-3_51Crossref 24Pensak David Corey E.J. LHASA—logic heuristics synthetic analysis.in: Computer-Assisted Organic Synthesis American Chemical Society, 1977: 1-32https://doi.org/10.1021/bk-1977-0061.ch001Crossref 44The logic synthesis: multistep synthesis carbogenic molecules (Nobel lecture) 1991 Angewandte Edition English Wiley Online Library [Internet].https://onlinelibrary.wiley.com/doi/abs/10.1002/anie.199104553Google 25Law Zsoldos Simon Reid D. Liu Khew S.Y. Johnson A.P. Major Wade Ando H.Y. Route designer: retrosynthetic utilizing rule generation.J. Chem. Model. 2009; 49: 593-602Crossref (134) 45Deng L. Yu Deep learning: applications.SIG. 2014; 7: 197-387https://doi.org/10.1561/2000000039Crossref (1817) 46Goodfellow I. Bengio Courville Learning. MIT 2016: 801Google 47Gil Greaves Hendler Hirsh Amplify intelligence.Science. 346: 171-172Crossref (72) Although latter part twentieth century,48Kayala M.A. Baldi P.F. predict reactions.in: Shawe-Taylor Zemel R.S. Bartlett P.L. Pereira Weinberger K.Q. Advances Neural Information Processing Systems 24 Curran Associates, Inc., 2011: 747-755http://papers.nips.cc/paper/4356-a-machine-learning-approach-to-predict-chemical-reactions.pdfGoogle hindered extravagant claims regarding capabilities computers. Both winters broken combination advent approaches AI49Segler M.H.S. Waller M.P. Neural-symbolic retrosynthesis reaction prediction.Chem. Eur. 23: 5966-5971Crossref (201) unveiling exciting AI,50Schwaller Gaudin Lányi Bekas C. Laino “Found Translation”: predicting outcomes reactions neural sequence-to-sequence models.Chem. 2018; 9: 6091-6098Crossref reignited fire excitement technologies. reviewing claims, Brannigan concluded 1989 programs available (yet) achieved autonomy discovery.51Staff How tech giants investing Techworld.https://www.techworld.com/picture-gallery/data/tech-giants-investing-in-artificial-intelligence-3629737/Google At beginning twenty-first century, Government invested e-science (often referred cyber-science outside generated character than e-science). 2009, Hey al.52Chemical companies ready seize opportunities Accenture [internet].https://www.accenture.com/gb-en/insights/chemicals/process-reimagined-chemicalsGoogle edited “The Paradigm, Discovery,” essays exploring range topics science, covering every discipline. his article, Gray proposed exploration paradigm, following empirical studies, methods, simulation. published program, 2009. References document appear concentrate resource befits initial middleware). references drug resolve databases rather facilitating actual molecule three mentions influence denied, particularly changing way wide researchers work; Computing Committee (UKCRC) recommended benefit technologies better facilitate collaboration, sharing datasets repositories (e.g., research.” data-driven, cloud-based services-based modern Internet-enabled commerce infrastructure. infrastructure, driven. 2, early applications automating aspects healthcare MYCIN CADUCEUS)37Bringsjord computer-aided design routes unknown DENDRAL, CAOS LHASA).19Lindsay 53Reyes transform Medium.https://medium.com/@joshdelosreyes/how-ai-will-transform-the-chemical-industry-5771c1cbc964Date: 2019Google 54Artificial reality engineers engineering page [internet].https://www.chemengonline.com/artificial-intelligence-new-reality-chemical-engineersGoogle Corey,55AI [Internet].https://www.soci.org/chemistry-and-industry/cni-data/2018/4/ai-in-industryGoogle formalized analysis, pertinent synthesis. Other active development,41Hey none yet challenge graduate chemist. Typically, either human-defined rules automatically extracted (with varying degrees success). 2014, area deep learning,56editor ISSPowerful antibiotic time. Guardian [Internet].https://www.theguardian.com/society/2020/feb/20/antibiotic-that-kills-drug-resistant-bacteria-discovered-through-aiGoogle methodology enormous amounts significantly enhanced possible demonstrate networks, trained scale graphical processing unit (GPU) computing. Significant advances respect representing unstructured strings molecular structures, meaningful algorithms.57Murray P.M. Tyler S.N.G. Moseley Beyond numbers: charting space.Org. Process. Res. Dev. 2013; 17: 40-46Crossref (48) time, seen contributions discovery: “However, broader AI-based represent hypotheses, reason models hypothesis-driven reduce error-prone bottleneck discovery.”58Reymond J.-L. project.Acc. 2015; 722-730Crossref (254) 2011, Kayala Baldi59Lipinski C.A. Lombardo Dominy B.W. Feeney P.J. Experimental estimate solubility permeability development settings1PII original article: S0169-409X(96)00423-1. originally Advanced Drug Delivery Reviews 23 (1997) 3–25.1.Adv. Deliv. 2001; 46: 3-26Crossref (7506) presented nets reactions, promising being conducted 2017.60Gorse A.-D. Diversity Medicinal Chemistry Space Bentham Publishers, 2006https://www.ingentaconnect.com/content/ben/ctmc/2006/00000006/00000001/art00004Google Scholar,61#Envision2030: 17 world persons disabilities United Nations Enable [Internet].https://www.un.org/development/desa/disabilities/envision2030.htmlGoogle Enumerating multitude straightforward (if large) task, gives connecting starting molecule, intermediate nodes nodes. select optimum route (let alone work) challenge, needs solved support demand revolutionize green circular economies. meet outcome reaction. (AAI) surpass applications62The UK’s Strategy GOV.UK [internet].https://www.gov.uk/government/topical-events/the-uks-industrial-strategyGoogle 63230 Startups [Internet].https://blog.benchsci.com/startups-using-artificial-intelligence-in-drug-discoveryGoogle 64Frey J.G. Bird C.L. Willoughby Smart spaces transfer.in: Howlett R.J. Gabrys Musial-Gabrys Roach through Transfer 2012 Springer, 2013: 31-38https://doi.org/10.1007/978-3-642-34219-6_4Crossref (2) and, large-scale heterogeneous derived interactions,65Frey meetings: experimenting space.in: 11-17https://doi.org/10.1007/978-3-642-34219-6_2Crossref 66Frey Human smart