The science behind precision synbiotics for autism.

1. Study overview

The study — “Precision synbiotics increase gut microbiome diversity and improve gastrointestinal symptoms in a pilot open-label study for autism spectrum disorder” — was published in mSystems (American Society for Microbiology) on April 25, 2024. It is the largest published investigation of personalized probiotic + prebiotic (synbiotic) intervention designed specifically for autistic individuals.

The core hypothesis: autistic individuals harbor a measurably distinct microbiome signature that can be identified via machine learning, and that personalized synbiotics (custom probiotics + matched prebiotics) formulated from that signature can increase microbial diversity and reduce GI symptoms.

2. Study design and cohort

This was a pilot open-label study — participants knew they were receiving the intervention. There was no placebo arm. The neurotypical controls (n=123) provided baseline microbiome samples for the ML classifier but did not receive the intervention.

IRB approval: Arizona State University, STUDY00012299. Registered on ClinicalTrials.gov before enrollment began.

Enrollment criteria

• ASD diagnosis: Confirmed via professional diagnosis (physician, psychologist, or school evaluation)
• Age range: 2.5–75 years
• Exclusions: Antibiotic use in the last 2 months; medication, supplement, or therapy changes in the last 2 months or planned during the first 3 months
• No exclusion for diet type, GI severity, or concurrent non-antibiotic medications — reflecting real-world conditions

Cohort demographics

Retention: 170 of 296 (57.4%) completed the full protocol. Dropout reasons among the 126 who left: lack of perceived benefit (34%), price (21%), customer service issues (3 cases). On average, 5–6 months elapsed between the baseline sample and the second timepoint sample receipt and processing.

3. Sequencing and bioinformatics

This study did not use standard 16S rRNA amplicon sequencing with a public pipeline. Instead, it used a proprietary metagenomic approach:

• Platform: Illumina NextSeq 550, 150 bp × 150 bp paired-end reads
• Mean sequencing depth: 17.15 million total reads (ASD cohort), 6.14 million reads (control cohort)
• Mean classified reads: 71% in both cohorts
• Reference database: Hand-curated database of ~23,000 microbial species, sourced from RefSeq, Greengenes, HMP, NCBI, and PATRIC
• Functional analysis: HUMAnN2 with MetaCyc and UniRef90 databases
• Gene family analysis: UniRef90 (201901b release); threshold >1,000 reads per kilobase per gene family

4. Machine learning classifier

The research team built a stochastic gradient boosting classifier to distinguish ASD microbiomes from neurotypical microbiomes using the baseline sequencing data.

Model specification

Results

A separate Random Forest analysis on the baseline cohort comparison had an out-of-bag error rate of 17.54%. From the top 50 features, 28 taxa reached significance after FDR correction and are described in the next section.

5. Baseline microbiome differences

The Random Forest analysis identified 28 taxa significantly different between ASD and neurotypical cohorts after false discovery rate (FDR) correction:

Elevated in ASD

The following genera/species were found at significantly greater proportions in the ASD cohort:

• Achromobacter
• Aeromonas
• Bacillus sp. (not B. subtilis specifically)
• Burkholderia
• Clostridium
• Cronobacter
• Klebsiella
• Micrococcus
• Rhodococcus
• Shigella
• Streptomyces

Reduced in ASD (higher in controls)

These taxa were at significantly larger proportions in the neurotypical cohort:

• Bacillus subtilis
• Faecalibacterium
• Fibrobacter
• Fusicatenibacter
• Lachnospira
• Lactococcus
• Paenibacillus
• Pseudoflavonifractor
• Ruminococcus
• Virgibacillus

Baseline diversity

Beta diversity (baseline)

• PERMANOVA: p = 0.065, R² = 0.00319 — no significant overall difference in community composition
• PCO1 axis (Wilcoxon): p = 0.0102 — significant separation
• PCO2 axis (Wilcoxon): p = 4.68 × 10^-12 — highly significant separation

Figure 3: Differential proportion of microbes between ASD and neurotypical cohorts at baseline

Taxa to the right are more abundant in ASD; taxa to the left are more abundant in neurotypical controls. 28 taxa reached significance after FDR correction.

Source: Phan et al., mSystems 2024, Fig. 3. CC BY 4.0.

6. Precision formulation pipeline

This is where the study departs from every other probiotic trial. Instead of giving everyone the same capsule, the team used a multi-step personalized pipeline:

Step 1: Metagenomic profiling

Each participant submitted a stool sample analyzed via Illumina NextSeq 550 sequencing (150 bp paired-end). Reads were aligned against the ~23,000 species reference database to produce a detailed species-level profile.

Step 2: Gap analysis

Each individual's profile was compared against the neurotypical reference panel and an internal database of healthy microbiome distributions, informed by the ML classifier's feature importance rankings. This identified which beneficial taxa were depleted and which potentially harmful taxa were elevated for that specific person.

Step 3: Custom synbiotic blend

• Probiotic component: 4–8 strains per participant, selected from >100 scientifically studied ingredients. Species and strains chosen based on the clinical literature curation database for their “ability to modulate the microbiota–gut–brain axis and improve ASD-specific characteristics”
• Prebiotic component: 1–2 types per participant, chosen to selectively feed target organisms
• Each formula was different — no two participants received the same blend

Post-intervention taxa that increased

After 3 months, participants showed significant increases in:

• Bacillus subtilis — increased to the point of being no longer significantly different from neurotypical controls
• Pseudoflavonifractor sp. An85 — same, converged with controls
• Bifidobacterium breve
• Lactobacillus reuteri
• Lactobacillus plantarum
• Lactobacillus brevis
• Enterococcus faecium (not a probiotic ingredient — endogenous increase)
• Enterococcus faecalis (not a probiotic ingredient — endogenous increase)

Figure 6: Longitudinal microbiome trends — taxa that shifted after supplementation

Post-supplementation, key taxa like B. subtilis and Pseudoflavonifractor converged with neurotypical levels. Several Lactobacillus and Bifidobacterium species increased as expected from probiotic supplementation.

Source: Phan et al., mSystems 2024, Fig. 6. CC BY 4.0.

7. Microbiome outcomes post-intervention

The primary microbiome endpoints were changes in alpha diversity from baseline (T1) to 3-month post-supplementation (T2) among the 170 completers.

Figure 2: Diversity comparisons — ASD vs. neurotypical, and longitudinal ASD changes

Left panels: alpha diversity (Shannon, richness, evenness) at baseline and post-supplementation. Right panels: PCoA ordination showing community-level shifts. After supplementation, ASD diversity was no longer significantly different from neurotypical controls.

Source: Phan et al., mSystems 2024, Fig. 2. CC BY 4.0.

Beta diversity (post-intervention)

The overall community composition (beta diversity) showed a shift toward the neurotypical cluster on PCO axes, though the ASD T1 vs. T2 comparison did not reach significance — meaning the shift was directional but modest at the whole-community level.

8. GI symptom outcomes (GSRS)

GI symptoms were measured using the Gastrointestinal Symptom Rating Scale (GSRS), a validated clinical instrument with five subscales scored 1–7 (1 = no discomfort, 7 = very severe).

Baseline GSRS scores

Post-supplementation

The composite GSRS score significantly decreased from T1 to T2 (Wilcoxon test, p < 0.05), indicating a reduction in GI symptom severity. Per the PGIA GI-specific question, 52% of participants reported GI improvement.

This is meaningful because GI symptoms affect up to 90% of autistic individuals and are correlated with behavioral challenges, sleep disruption, and reduced quality of life.

9. Behavioral and developmental outcomes

Multiple validated instruments were used. Here's what each one found:

PGIA (Parent Global Impressions of Autism)

A parent-reported measure where families rate change from baseline on a scale of −3 (much worse) to +3 (much better).

PGIA by domain (≥50% reported improvement)

SRS-2 (Social Responsiveness Scale, 2nd edition)

Baseline mean: 80.08 (SD 10.36). Distribution: 29% moderate, 6% severe, 6% mild-to-moderate.

Result: No significant change post-supplementation.

SCARED (Screen for Child Anxiety Related Disorders)

Result: No significant change post-supplementation.

Figure 7: Longitudinal survey assessments — GSRS, PGIA, SRS-2, and SCARED

GSRS showed significant improvement (GI symptoms decreased). PGIA showed 62% parent-reported improvement but scores are within placebo range. SRS-2 and SCARED showed no significant changes.

Source: Phan et al., mSystems 2024, Fig. 7. CC BY 4.0.

10. Metabolic pathway analysis

Using HUMAnN2 functional profiling, the study identified significant differences in metabolic pathways between ASD and neurotypical microbiomes at baseline:

Elevated in ASD microbiomes

• Biosynthesis of amino acids, cofactors, vitamins, fatty acids/lipids, nucleotides, tetrapyrrole
• Degradation of carboxylates
• Fermentation of pyruvate, alcohols, and fermentation to short-chain fatty acids
• LPS (lipopolysaccharide) biosynthesis superpathway — associated with gut inflammation
• Polymyxin resistance — antibiotic resistance pathway

Reduced in ASD microbiomes

• Carbohydrate degradation
• Sulfur oxidation
• Thiamine biosynthesis
• Peptidoglycan biosynthesis (higher in controls)
• 4-aminobutanoate (GABA) degradation — significantly lower in ASD vs. controls

Post-supplementation pathway changes

Several pathways significantly decreased after intervention, converging toward neurotypical levels:

• Petroselinate biosynthesis — decreased, no longer significantly different from NT
• L-aspartate and L-asparagine biosynthesis — decreased
• Unsaturated fatty acid biosynthesis — decreased
• Purine deoxyribonucleoside degradation — decreased
• Purine nucleotide salvage — decreased

Figure 4: Differential metabolic pathways between ASD and neurotypical cohorts

Pathways to the right are more abundant in ASD; pathways to the left are more abundant in neurotypical controls. Note elevated LPS biosynthesis and polymyxin resistance in ASD, and reduced GABA degradation.

Source: Phan et al., mSystems 2024, Fig. 4. CC BY 4.0.

Gene family analysis

A Random Forest on gene families had an error rate of 15.35%. Of the top 50 discriminating gene families: 6 were higher in ASD, 44 were lower. Most of the depleted gene families were annotated from Ruminococcus spp., Fusicatenibacter saccharivorans, and Faecalibacterium prausnitzii — key butyrate producers and gut-health organisms.

Figure 5: Differential gene families between ASD and neurotypical cohorts

44 of the top 50 discriminating gene families were depleted in ASD. Most originated from butyrate-producing species critical for gut barrier integrity.

Source: Phan et al., mSystems 2024, Fig. 5. CC BY 4.0.

11. Correlations

The study tested several correlations between microbiome metrics, behavior, and nutrition:

Non-significant: Shannon, richness, and evenness were not significantly correlated with overall nutritional assessment, PGIA at baseline, SRS-2, or GSRS.

Subgroup analyses

Figure 1: Microbes associated with ASD subpopulations

Microbial differences within the ASD cohort based on social responsiveness severity (SRS-2) and dietary patterns. Prevotella, Bacteroides, and Fusicatenibacter distinguished severe from non-severe groups.

Source: Phan et al., mSystems 2024, Fig. 1. CC BY 4.0.

• SRS-2 severity: Random Forest identified significant microbial differences between severe vs. non-severe social responsiveness groups. Prevotella spp., Bacteroides, and Fusicatenibacter were higher in the severe group.
• Diet (GFCF vs. inclusive): Dairy/gluten consumers had higher Bifidobacterium spp., Lactococcus lactis, and Streptococcus spp. No significant beta diversity differences by diet.
• Gender: No significant differences in microbial alpha or beta diversity between sexes in either cohort.
• Anxiety (SCARED): No significant microbial species differentiated cohorts based on anxiety status.

12. Safety and tolerability

• No serious adverse events (SAEs) attributed to the intervention
• Only 6% reported any worsening on the PGIA overall scale
• Dropout was logistical, not safety-related: 34% cited lack of perceived benefit, 21% cited price, 3 cases related to customer service
• Exclusion criteria (no antibiotics in prior 2 months, no medication changes) helped ensure a clean baseline

13. Limitations — in full

The authors themselves list these limitations. We believe in showing the full picture:

1. Open-label, no placebo arm. Participants knew they were receiving the intervention. Placebo effect and expectation bias cannot be ruled out — especially for subjective measures like PGIA.
2. PGIA scores are within placebo range. The overall PGIA improvement (0.36 ± 0.55) is nearly identical to the placebo group in Adams et al. 2013 (0.34 ± 0.54). The authors explicitly state: “Survey to measure improvements is comparable to a placebo effect.”
3. SRS-2 and SCARED showed no significant change. Standardized behavioral instruments did not detect improvement — only the parent-reported PGIA did.
4. Gender imbalance. ASD cohort was 79.7% male vs. 52.0% in controls. While no diversity differences by sex were found in this dataset, evidence from other studies suggests gender may influence the ASD microbiome.
5. No control data on diet, GI, or behavior. The neurotypical controls provided microbiome samples only. No dietary, GI, or behavioral data was collected from controls, limiting comparison.
6. 42.6% dropout rate. Only 170 of 296 completed the full protocol. Completers may systematically differ from dropouts — particularly since 34% dropped out for “lack of perceived benefit,” which biases the remaining sample toward those who felt it was working.
7. Each participant got a different formula. This is inherent to the personalized approach but makes it impossible to attribute outcomes to any specific strain or combination.
8. ML classifier doesn't generalize. The stochastic gradient boosting model achieved AUC 0.95 on training data but only 21.8% accuracy on validation — it overfits and cannot be used as a diagnostic tool.
9. No long-term follow-up. Whether benefits persist after stopping supplementation is unknown.
10. Not a randomized controlled trial (RCT). The gold standard for causal claims. The authors describe this as a pilot study. A proper RCT would be needed to establish efficacy.

14. Full citation and disclosures

Conflicts of interest

Transparency matters. Here are the declared conflicts:

• J. Phan, D. Nair, S. Jain, T. Montagne, S. Dietsche, K. Blanchard, and S. Treadwell are employed by Flore Inc. (formerly Sun Genomics Inc.)
• D.C. Calvo was co-supported by Flore Inc. (formerly Sun Genomics Inc.)
• J. Adams and R. Krajmalnik-Brown are co-founders of Gut-Brain Axis Therapeutics Inc., which is developing microbiome-based treatments for autism and other disorders

The study was conducted through IRB collaboration between Sun Genomics, Arizona State University, and the Biodesign Center.