Amyotrophic lateral sclerosis (ALS) is characterised by motor neuron (MN) death; however, astrocytes play a key role in disease pathogenesis. Developments in the field of artificial intelligence (AI) have the potential to impact drug discovery in multiple ways, including the rapid identification of drug repurposing candidates. A combination of natural language processing and deep learning algorithms was used to generate a knowledge graph based on scientific literature, omics and chemical databases, and other public sources with the aim to identify drug repurposing candidates for ALS. The aim of the study was to determine the effect of a cancer compound identified by AI, gefitinib, on MN survival, and to decipher its mode of action in in vitro and in vivo models of ALS. We used co-cultures of healthy motor neurons with ALS patient-derived astrocytes (iAstrocytes), obtained through a semi-direct conversion protocol, to assess the neuroprotective properties of gefitinib. Compound treatment led to a significant rescue of MNs cultured with ALS iAstrocytes and a significant reduction in the levels of cleaved TDP-43 fragments in ALS iAstrocytes. Our data suggest that gefitinib-mediated activation of autophagy decreased the 35 kDa fragments of TDP-43. In a proof-of-concept in vivo study in SOD1G93A mice, gefitinib treatment significantly delayed the onset of neurological symptoms, thus showing the potential of AI-augmented drug discovery for neurodegenerative disorders. ### Competing Interest Statement The research described in this manuscript was funded by BenevolentAI. DWS, AP, DPS and PJR were employed by BenevolentAI.

Efficacy of combined vandetanib and everolimus in vivo. A, Tolerability in NOD.SCID mice exposed to daily oral treatment with the combination of vandetanib and everolimus over 14 days, as assessed by body weight relative to day 0. Mean and SD of three mice per group are plotted. PO, oral administration. B, Schema for in vivo efficacy (survival) and tumor burden experiments in orthotopic PDX and GEMM allografts of mutant ACVR1-driven DIPG. Tx, treatment. C, Survival curves for mice (n = 16–22 per group) bearing HSJD-DIPG-007 orthotopic xenografts, treated with vandetanib (pink), everolimus (light orange), or the combination (dark red), compared with vehicle-treated controls (gray). D, Bar plot quantifying tumor burden as assessed by ddPCR for ACVR1R206H in mice treated with vandetanib (pink), everolimus (light orange), or the combination (dark red), expressed as a percentage of vehicle-treated controls (gray). E, Bar plot quantifying cellularity by human nuclear antigen (HNA)–positive cells in mice treated with vandetanib (pink), everolimus (light orange), or the combination (dark red), expressed as a percentage of vehicle control (gray). Mean and SD plotted. *, P < 0.05, adjusted t test. F, Survival curves for mice (n = 16–22 per group) bearing Nestin-Tv-a; Trp53fl/fl; Hist1h3bK27M, Acvr1R206H orthotopic allografts, treated with vandetanib (pink), everolimus (light orange), or the combination (dark red), compared with vehicle-treated controls (gray). G, Dot plot of vandetanib concentration in the normal brains and engrafted tumors of nude mice treated with combined vandetanib and everolimus, as assessed by mass spectrometry. The black horizontal line represents the median. H, Dot plot of vandetanib concentration in the normal brains and engrafted tumors of NSG mice treated with combined vandetanib and everolimus, as assessed by mass spectrometry. The black horizontal line represents the median.

Microbial colonization of the human gut occurs soon after birth, proceeds through well-studied phases and is affected by lifestyle and other factors. Less is known about phage community dynamics during infant gut colonization due to small study sizes, an inability to leverage large databases and a lack of appropriate bioinformatics tools. Here we reanalysed whole microbial community shotgun sequencing data of 12,262 longitudinal samples from 887 children from four countries across four years of life as part of the The Environmental Determinants of Diabetes in the Young (TEDDY) study. We developed an extensive metagenome-assembled genome catalogue using the Marker-MAGu pipeline, which comprised 49,111 phage taxa from existing human microbiome datasets. This was used to identify phage marker genes and their integration into the MetaPhlAn 4 bacterial marker gene database enabled simultaneous assessment of phage and bacterial dynamics. We found that individual children are colonized by hundreds of different phages, which are more transitory than bacteria, accumulating a more diverse phage community over time. Type 1 diabetes correlated with a decreased rate of change in bacterial and viral communities in children aged one and two. The addition of phage data improved the ability of machine learning models to discriminate samples by country. Finally, although phage populations were specific to individuals, we observed trends of phage ecological succession that correlated well with putative host bacteria. This resource improves our understanding of phage–bacteria interactions in the developing early life microbiome. Reanalysis of 12,262 longitudinal infant gut microbiome samples using the Marker-MAGu pipeline revealed phage–bacteria dynamics over the first year of life.

Synergy of combined vandetanib and everolimus in vitro. A, Cell viability matrix for ACVR1R206H HSJD-DIPG-007 cells treated with distinct combinations of vandetanib (y-axis) and everolimus (x-axis) ranging from 0 to 20 μmol/L. B, Excess above Bliss matrix for ACVR1R206H HSJD-DIPG-007 cells treated with distinct combinations of vandetanib (y-axis) and everolimus (x-axis) ranging from 0 to 20 μmol/L. C, Bliss synergy map for ACVR1R206H HSJD-DIPG-007 cells treated with distinct combinations of vandetanib (y-axis) and everolimus (x-axis) ranging from 0 to 10 μmol/L. Heat maps colored according to the keys provided. D, Cell viability matrix for ACVR1R258G HSJD-DIPG-018 cells treated with distinct combinations of vandetanib (y-axis) and everolimus (x-axis) ranging from 0 to 20 μmol/L. E, Excess above Bliss matrix for ACVR1R258G HSJD-DIPG-018 cells treated with distinct combinations of vandetanib (y-axis) and everolimus (x-axis) ranging from 0 to 20 μmol/L. F, Bliss synergy map for ACVR1R258G HSJD-DIPG-018 cells treated with distinct combinations of vandetanib (y-axis) and everolimus (x-axis) ranging from 0 to 10 μmol/L. Heat maps colored according to the keys provided.

Validation of vandetanib as an ACVR1 inhibitor and everolimus as a P-gp inhibitor. A, Kinome selectivity map for vandetanib at 10 μmol/L, with red circles reflecting kinases inhibited by >65%. ACVR1 is labeled. The diameter of the circles is inversely proportional to the percentage of kinase activity remaining in the presence of an inhibitor. Taken from the Harvard Medical LINCS KINOMEscan database (https://lincs.hms.harvard.edu/). B, Bar plots showing the IC50 for various compounds against wild-type (gray) and mutant ACVR1: Q207E, R206H, and G328V (shades of purple). C, Western blot of ACVR1R206H HSJD-DIPG-007 cells treated with increasing concentrations of vandetanib in the presence of 0.1 μmol/L, 1.0 μmol/L, or 10 μmol/L everolimus. GAPDH is the loading control. D, Bar plots showing apparent permeability (Papp) of a BCRP substrate across a bidirectional cell monolayer (A–B and B–A) in the presence of various inhibitors. E, Bar plots showing apparent permeability (Papp) of a P-gp substrate across a bidirectional cell monolayer (A–B and B–A) in the presence of various inhibitors. Vandetanib, pink; everolimus, light orange; combination, dark red. Error bars represent SD of the mean. ***, P < 0.0001, ANOVA. n.s., not significant. F, Dot plot of vandetanib concentration (Conc) in the brains of normal mice treated with vehicle control (gray), vandetanib (pink), or the combination (dark red), as assessed by mass spectrometry. The black horizontal line represents the median.

Clinical experience of combined vandetanib and everolimus in a 7-year-old girl with ACVR1-mutant DIPG (case 1). A, Timeline summarizing the clinical history of case 1, treated at the Royal Marsden Hospital, London, United Kingdom. *Imaging scans presented in B. Dx, diagnosis. B, Sequential head MRI scans: diagnostic MRI showing diffuse heterogenous tumor infiltration and anatomic distortion of the pons, including partial invagination of the basilar artery (i) with a single focus of intratumoral enhancement in the right central pons (ii); MRI scan 6 weeks after radiotherapy (RT) showed considerable reduction in tumor infiltrative bulk (iii) with the small focus of enhancement persisting in the right pons (iv); the MRI scan after three cycles of everolimus single agent constituted also the baseline scan prior to starting vandetanib/everolimus, and it showed increased T2 signal abnormality in the right pons (v) associated with mildly increased focal enhancement (vi); the MRI scans after three cycles (vii, viii) and six cycles (ix, x) of vandetanib/everolimus showed modest enlargement and progressive infiltration of the right pons (vii, ix) with concomitant increase of the associated focal enhancement (viii, x) but nevertheless with relatively stable appearances/lack of progression elsewhere. The child started cycle 7 of vandetanib/everolimus but then developed worsening symptoms (slurred speech, ataxia), and treatment was permanently discontinued 11 months after initial diagnosis. No further systemic treatment or reirradiation was administered. The child could be weaned off steroids completely, and her balance and activity levels improved significantly. Eight months after discontinuation of vandetanib/everolimus, she developed further clinical and radiologic progression (local and metastatic) and died a month later (OS, 20 months). C, Macroscopic transverse cross section of pons and cerebellum from the brain autopsy specimen from case 1. D, Hematoxylin and eosin–stained image from the inferior left pons (×40), with a higher magnification inset (×200) of region 1, showing diffuse infiltration of tumor cells. E, Protein structure representation of KRAS showing mutant G12A residue (shaded red), generated in COSMIC-3D (https://cancer.sanger.ac.uk/cosmic3d/). Allelic fractions are provided for mutant reads identified in distinct areas of the pons taken at autopsy. F, MAPK pathway activation in distinct autopsy regions measured by quantitative capillary electrophoresis for phospho-ERK1/2T202/Y204, plotted as a ratio of respective phosphorylated/total protein compared with normal cerebellum.

Supplementary Figures S1-S3

Identification of Fusobacterium spp. and subspecies in human colorectal tumors using Fusobacterium genus 16S V2 - V4 amplicon cloning and sequencing, as well as amplicon-targeted direct sequencing

Migration of HCP1 CRC cells in a Boyden chamber system.

Morphological changes in THP-1 monocytes induced by co-culture with F. nucleatum ssp. animalis under normoxic conditions.

Identification of Fusobacterium spp. in normal colorectal mucosal and tumor specimens.

CCL20-specific ELISA confirming differential expression of CCL20 in human colorectal tumors.

Amyotrophic lateral sclerosis (ALS) is a poorly treated multifactorial neurodegenerative disease associated with multiple cell types and subcellular organelles. As with other multifactorial diseases, it is likely that drugs will need to target multiple disease processes and cell types to be effective. We review here the role of Janus kinase (JAK)/Signal transducer and activator of transcription (STAT) signalling in ALS, confirm the association of this signalling with fundamental ALS disease processes using the BenevolentAI Knowledge Graph, and demonstrate that inhibitors of this pathway could reduce the ALS pathophysiology in neurons, glia, muscle fibres, and blood cells. Specifically, we suggest that inhibition of the JAK enzymes by approved inhibitors known as Jakinibs could reduce STAT3 activation and modify the progress of this disease. Analysis of the Jakinibs highlights baricitinib as a suitable candidate due to its ability to penetrate the central nervous system and exert beneficial effects on the immune system. Therefore, we recommend that this drug be tested in appropriately designed clinical trials for ALS.

Expression of CCL20 in human CRC cell lines and THP-1 monocytes induced by co-culture with F. nucleatum ssp.

Supplementary Figure Legends describe the identification of Fusobacterium by an alternative method and confirmation of CCL20 protein expression by ELISA and functional assays in additional cell lines.

Humans are colonized with commensal bacteria soon after birth, and, while this colonization is affected by lifestyle and other factors, bacterial colonization proceeds through well-studied phases. However, less is known about phage communities in early human development due to small study sizes, inability to leverage large databases, and lack of appropriate bioinformatics tools. In this study, whole genome shotgun sequencing data from the TEDDY study, composed of 12,262 longitudinal samples from 887 children in 4 countries, is reanalyzed to assess phage and bacterial dynamics simultaneously. Reads from these samples were mapped to marker genes from both bacteria and a new database of tens of thousands of phage taxa from human microbiomes. We uncover that each child is colonized by hundreds of different phages during the early years, and phages are more transitory than bacteria. Participants' samples continually harbor new phage species over time whereas the diversification of bacterial species begins to saturate. Phage data improves the ability for machine learning models to discriminate samples by country. Finally, while phage populations were individual-specific, striking patterns arose from the larger dataset, showing clear trends of ecological succession amongst phages, which correlated well with putative host bacteria. Improved understanding of phage-bacterial relationships may reveal new means by which to shape and modulate the microbiome and its constituents to improve health and reduce disease, particularly in vulnerable populations where antibiotic use and/or other more drastic measures may not be advised.

Clinical characteristics of the study patients with CRC.

During the current pandemic, the vast majority of COVID-19 patients experienced mild symptoms, but some had a potentially fatal aberrant hyperinflammatory immune reaction characterized by high levels of IL-6 and other cytokines. Modulation of this immune reaction has proven to be the only method of reducing mortality in severe and critical COVID-19. The anti-inflammatory drug baricitinib (Olumiant) has recently been strongly recommended by the WHO for use in COVID-19 patients because it reduces the risk of progressive disease and death. It is a Janus Kinase (JAK) 1/2 inhibitor approved for rheumatoid arthritis which was suggested in early 2020 as a treatment for COVID-19. In this review the AI-assisted identification of baricitinib, its antiviral and anti-inflammatory properties, and efficacy in clinical trials are discussed and compared with those of other immune modulators including glucocorticoids, IL-6 and IL-1 receptor blockers and other JAK inhibitors. Baricitinib inhibits both virus infection and cytokine signalling and is not only important for COVID-19 management but is "non-immunological", and so should remain effective if new SARS-CoV-2 variants escape immune control. The repurposing of baricitinib is an example of how advanced artificial intelligence (AI) can quickly identify new drug candidates that have clinical benefit in previously unsuspected therapeutic areas.

AbstractSomatic mutations in ACVR1 are found in a quarter of children with diffuse intrinsic pontine glioma (DIPG), but there are no ACVR1 inhibitors licensed for the disease. Using an artificial intelligence–based platform to search for approved compounds for ACVR1-mutant DIPG, the combination of vandetanib and everolimus was identified as a possible therapeutic approach. Vandetanib, an inhibitor of VEGFR/RET/EGFR, was found to target ACVR1 (Kd = 150 nmol/L) and reduce DIPG cell viability in vitro but has limited ability to cross the blood–brain barrier. In addition to mTOR, everolimus inhibited ABCG2 (BCRP) and ABCB1 (P-gp) transporters and was synergistic in DIPG cells when combined with vandetanib in vitro. This combination was well tolerated in vivo and significantly extended survival and reduced tumor burden in an orthotopic ACVR1-mutant patient-derived DIPG xenograft model. Four patients with ACVR1-mutant DIPG were treated with vandetanib plus an mTOR inhibitor, informing the dosing and toxicity profile of this combination for future clinical studies.Significance:Twenty-five percent of patients with the incurable brainstem tumor DIPG harbor somatic activating mutations in ACVR1, but there are no approved drugs targeting the receptor. Using artificial intelligence, we identify and validate, both experimentally and clinically, the novel combination of vandetanib and everolimus in these children based on both signaling and pharmacokinetic synergies.This article is highlighted in the In This Issue feature, p. 275