Home AI ServicesAI-Powered Digital Library-to-Library Interaction Screening

ICDMO AI Services

AI-Powered Digital Library-to-Library Interaction Screening

High-Throughput Many-to-Many PPI Prediction

Run 'many-to-many' screening between two gene or protein libraries using AI-PPI deep learning and MEGADOCK docking to build candidate interaction networks. Screening scale is agreed against available compute during scoping.

N × M

Pairwise scale

3–6 wk

Typical turnaround

Ranked

Cross-matrix output

Batched

Compute allocation

Illustrative image representing computational infrastructure

No Reagents for the In Silico Phase

Confidence-Scored Results

Publication-Ready Figures

Validation Planned With Partners

Response Within One Business Day

Service Overview

Unlike 'one-to-many' approaches, ICDMO's digital library-to-library screening evaluates pairwise interactions between two protein libraries (bait library A × prey library B). The platform runs in two tiers: MEGADOCK for fast initial docking, then AI-PPI for finer secondary analysis, narrowing a large candidate set down to a ranked shortlist. Whole-genome and transcription factor digital libraries are available as starting points. The output is a set of interaction hypotheses ranked by model confidence — no experimental material is needed for the computational phase, and confirmation still requires laboratory work.

Schematic: screening runs as a cascade. Cheap filters remove the bulk, expensive scoring is applied only to survivors, and the attrition at every stage is reported.

Technical Advantages

✓

Ready-to-Use Libraries

Whole-genome and transcription factor digital libraries are available as starting points, so screening can begin without building a physical library first. Species coverage is confirmed during scoping.

✓

Two-Stage AI Strategy

MegaDock initial screening + AI-PPI deep learning secondary analysis to maximize precision while maintaining high throughput.

✓

Cross-Species Capability

Study virus–host interactions and inter-species protein networks with the same computational pipeline.

✓

Publication-Quality Visualization

Downstream AOS analysis provides 3D protein models (overview + close-up) with binding site analysis, H-bonds, and distance information.

✓

Experimental Validation Support

Results feed directly into Y2H, Y1H, Co-IP, EMSA, BiFC, and dual-luciferase assay validation services.

✓

Extreme Throughput

Screen up to 10 million+ interaction pairs. Turnaround scales from 7 days (10K pairs) to 65 days (8–10M pairs).

Methodology

The platform performs a "many-to-many" interaction screen by: (1) Modeling 3D protein structures for every gene in both libraries using AI-based structure prediction; (2) Running MEGADOCK all-to-all protein docking to generate initial interaction scores; (3) Applying the AI-PPI deep learning model for high-precision secondary analysis; (4) Delivering ranked interaction pairs with gene IDs, modeling scores, and binary interaction calls (1 = predicted interaction, 0 = no interaction), sorted by descending score. The total interaction score represents the predicted docking score between two proteins — higher score indicates greater likelihood of interaction.

Sample Submission Requirements

1Bait Library: CDS (coding sequence) file required

2Prey Library: CDS (coding sequence) file required

3If a specific database version is required, specify clearly at submission

4For cross-species studies, provide species annotation alongside CDS files

Comparison: Digital vs. Traditional Library Screening

Digital Library-to-Library (AI-Powered)

▸Principle: Computational models simulate structure/sequence binding using deep learning

▸Materials: Protein sequences or structural models only

▸Throughput: Extremely high — millions of interactions in days

▸Cost: Zero experimental reagent cost

▸Timeline: 7–65 days depending on pair count

Traditional Yeast Library-to-Library (Y2H)

▸Principle: Experimental yeast growth reporter system

▸Materials: Constructed bait and prey plasmid libraries required

▸Throughput: Low — weeks to months for comparable coverage

▸Cost: Significant reagent and labor cost

▸Accuracy: High confidence but susceptible to biological background noise

What We Need From You

1Both libraries to be crossed: sequences, structures or compound sets, with the format of each stated

2The relationship being searched for — binding, inhibition, substrate turnover or interaction

3Any known positive and negative pairs, which anchor the scoring and allow retrospective validation

4Selection constraints and how many pairs you can realistically test downstream

5Computational budget and deadline, since pairwise screening scales as the product of the two libraries

How the Analysis Is Performed

1Preparation and standardisation of both libraries with consistent parameters across the matrix

2Tiered evaluation: cheap scoring across the full matrix, progressively more expensive methods on survivors

3Structure prediction or docking for the shortlisted pairs, depending on the modality involved

4Score normalisation across the matrix so that rows and columns remain comparable

5Retrospective validation against known pairs where such data exists

6Clustering of hits to distinguish genuine specificity from promiscuous rows and columns in the matrix

What You Receive

1Ranked pair matrix with scores for every combination evaluated

2Shortlisted pairs with structural models and interface or binding-mode analysis

3Heatmaps and network visualisations of the cross-library results

4Assessment of score distribution, including where the ranking is and is not discriminating

5Recommended experimental testing order for the prioritised pairs

Interpretation and Limitations

Cross-library screening multiplies both the opportunity and the false-positive burden: a matrix of ten thousand pairs will contain high-scoring pairs by chance alone. We report score distributions and flag promiscuous entries rather than presenting the top of the matrix as a hit list. All pairs are hypotheses for experimental testing, and deliverables are for research use only.

Questions We Are Asked Before Scoping

How large can the matrix be?

▸Tiered evaluation makes matrices of millions of pairs tractable, with expensive methods reserved for survivors. The practical limit is usually the budget and deadline rather than the method.

Can you cross different modalities?

▸Yes — compound sets against protein panels, peptide libraries against receptor families, or protein libraries against each other. The scoring approach changes with the modality and is stated in the proposal.

How do we avoid chasing artefacts?

▸By examining the score distribution rather than the ranking alone. Entries that score well against everything are flagged as promiscuous, and we recommend testing pairs that show selectivity within the matrix.

Technical Workflow

Scientific Consultation

We discuss the objective, the decision the result has to support and the data you already hold. Nothing is quoted before that is clear.

Feasibility Review

We check that your inputs, target class and available evidence can support the approach, and say plainly where they cannot.

Technical Proposal

Scope, methods, deliverables, timeline and price are confirmed in writing, along with the limitations we expect to report.

Computation

Both libraries are prepared, the full matrix is evaluated with fast scoring, and survivors progress through progressively more rigorous methods with normalisation across the matrix.

Report and Follow-Up

You receive the ranked matrix, the shortlisted pairs with structural analysis, and guidance on which results are discriminating and which are noise.

Service Timeline & Deliverables

Interaction Pair Count	Turnaround	Deliverables
10,000 < X ≤ 500,000 pairs	7 working days	Excel file with gene IDs, modeling scores, interaction status (1/0), sorted by descending score
500,000 < X ≤ 1,000,000 pairs	15 working days	Complete interaction dataset with ranked predictions
1,000,000 < X ≤ 2,000,000 pairs	20 working days	Full analysis report + interaction network visualization
2,000,000 < X ≤ 4,000,000 pairs	25 working days	Comprehensive dataset with prioritized hits
4,000,000 < X ≤ 6,000,000 pairs	40 working days	Full dataset + downstream AOS analysis available
6,000,000 < X ≤ 8,000,000 pairs	50 working days	Complete screening report
8,000,000 < X ≤ 10,000,000 pairs	65 working days	Complete screening report + optional 3D visualization

Is This the Right Service for You?

Choose this when

Two libraries must be crossed against each other rather than screened singly

A panel of targets needs profiling against a compound or sequence set

Specificity within a matrix matters as much as raw score

Consider another service when

One target and one library is the actual problem

AI Multimodal Protein–Compound Screening

Protein–protein pairs are the subject

Protein Interaction & Binding Site Prediction

What Goes Wrong, and How We Handle It

These are the failure modes that produce confident-looking results which do not survive the next experiment. Each one is designed against rather than discovered afterwards.

High scores arising by chance in a large matrix

Score distributions are analysed and the expected number of chance hits is stated before any pair is recommended.

Rows or columns that score well against everything

Promiscuous entries are flagged, and selectivity within the matrix drives selection rather than absolute score.

Compute spent uniformly across a matrix that does not need it

Tiered evaluation reserves expensive methods for survivors, which is what makes large matrices affordable.

A Worked Example

Profiling a compound set across a target family

The situation

The team needed selectivity information across twelve family members for 4,000 compounds.

What the work involved

All receptors were prepared under one protocol, the full matrix was scored quickly, survivors progressed to precision docking, and scores were normalised so rows and columns stayed comparable.

What came back

A ranked matrix with selectivity ratios per compound, forty promiscuous compounds flagged, and eleven compounds predicted to discriminate between the two most similar family members.

Illustrative of the shape this work takes. Details are generalised, no client or target is identified, and scope and outcome vary with the system in question.

What Published Work Reports

The results below were published by independent groups and are cited to show what these computational approaches have achieved in the literature. They are not ICDMO project results, and outcomes vary substantially between targets, datasets and programmes. We do not promise comparable results for any particular project.

A 22% confirmed hit rate from docking 138 million compounds

In a landmark 2019 study, ~138 million make-on-demand compounds were docked against the dopamine D4 receptor. Of 549 compounds selected and tested, 122 were confirmed active — a 22% hit rate — including ligands in the sub-nanomolar range. The same work screened ~99 million compounds against AmpC β-lactamase, where 5 of 44 tested compounds were confirmed (11%), one of which was among the most potent non-covalent AmpC inhibitors reported at the time.

Lyu et al., Nature, 2019 ↗

Hit rate rises with rank, which is why library scale matters

The same body of work showed that the probability of a compound being confirmed active increases as its docking rank improves, and that larger libraries push better-scoring molecules into the selected set. This is the practical argument for screening make-on-demand space rather than a physical catalogue — and also the reason scoring accuracy, not library size alone, remains the limiting factor.

Lyu et al., Nature, 2019 ↗