FALS Therapeutics: Prompting a Blueprint for a Biotech Discovering and Developing Disease-Modifying Therapies for Amyotrophic Lateral Sclerosis Using an Agentic AI Framework

Sándor Szalma

Abstract

The high cost and complexity of early-stage drug discovery create profound barriers for rare-disease programs with limited funding. Sporadic amyotrophic lateral sclerosis (sALS) exemplifies this challenge: despite affecting 90-95% of the ~32,000 prevalent ALS patients in the United States, no approved disease-modifying therapy exists specifically for sALS. Here we demonstrate FALS Therapeutics, a conceptual virtual biotech blueprint built and directed entirely through iterative prompting of Claude Sonnet 4.6 (Anthropic), a commercially available agentic large language model accessed through the claude.ai conversational interface without custom software infrastructure, MCP servers, or fine-tuning. A CEO agent received a USD 5M seed mandate and instantiated four AI direct-report agents: Head of Commercial, Head of Research, Head of CMC, and Head of Clinical Development, forming a Development Committee (DC) for stage-gate governance. The virtual organization completed four sequential work-streams: commercial landscape analysis, research master plan including target selection and molecule generation decision, clinical development master planning including a patient enrichment modeling strategy and combination trial proposal, and CMC manufacturing planning across four drug modalities. The blueprint was created by the AI agent via planning documents and interactive dashboards, presentation decks, figures, tables and code. Key outputs include an ~USD 791-812M global ALS market characterization, a 12-target ranked biological long-list with TBK1 identified as a novel Tier 1 target not in the initial mandate, a fully open-source AI tool stack saving ~USD 302,000/year in licensing costs, a Phase 1-3 clinical development plan with a four-layer patient enrichment model (FALS-PEM) estimated to reduce Phase 2b sample sizes by 40-60%, a Phase 2a and 2b combo trial of two assets in the pipeline proposed by the AI agent without prompting, and a consolidated CMC budget of ~USD 35.5M across all modalities from seed to NDA. A critical design feature - the iterative gap-identification and gap-closure loop directed by the CEO agent by prompts from expert human-in-the-loop - produced 18 governance documents without additional human specification. A systematic reference verification audit - the primary methodological contribution - flagged seven categories of AI-generated claims requiring caution and identified and corrected three errors introduced by AI hallucination. This work extends earlier published work of AI agents in drug discovery by demonstrating the lower bound of what expert-guided prompting of an unmodified conversational AI, without bespoke agent infrastructure, can achieve. It can generate the full portfolio of pre-IND planning documents for a rare-disease biotech, compressing months of work into days at minimal cost but expert review and continuous feedback via prompting is still crucial for ensuring scientific and technical validity.

ark:99999/genrxiv-2026-00004

Sándor Szalma Inxyte Informatics, Carlsbad, CA, USA FALS Therapeutics (Virtual); conceptual demonstration, April 2026

Keywords: agentic AI, prompt engineering, large language model, drug discovery, ALS, amyotrophic lateral sclerosis, virtual biotech, target identification, molecular design, patient enrichment, open-source tools, TDP-43, TBK1, STING, UNC13A, Claude, Anthropic, PRS, clinical development, CMC

1. Introduction

Drug discovery and development require integrating diverse evidence across biological scales and data modalities [1–3]. Bringing a single therapeutic from concept to clinic often requires more than a decade of development and costs on the order of billions of dollars [1, 2]. Despite these investments, approximately 90% of drug candidates entering Phase I trials never reach approval, with failures most often driven by insufficient efficacy and unexpected safety liabilities [3–6]. The integration of artificial intelligence (AI) into drug discovery has been proposed as a means to reduce this attrition by enabling faster, more data-driven decision-making [7, 8].

Several recent developments define the current landscape for AI-driven drug discovery. The Virtual Biotech by Zhang et al. is a multi-agent platform with four scientific divisions, eleven total agents, and over 100 MCP-server tools, demonstrated across clinical trial annotation (55,984 trials via 37,075 parallel agents), target evaluation (B7-H3 in lung cancer, USD 46 in API costs), and trial failure analysis (OSMRβ in ulcerative colitis, USD 54 in API costs) [9]. Biomni by Huang et al. further extended this paradigm with a general-purpose biomedical AI agent capable of autonomously executing research tasks across 25 biomedical domains including causal gene prioritization, drug repurposing, rare disease diagnosis, and microbiome analysis, integrating large language model reasoning with retrieval-augmented planning and code-based execution [10]. Zhavoronkov et al. articulate a “prompt-to-drug” vision - a pipeline where a plain-language prompt initiates a fully autonomous end-to-end drug development program encompassing hypothesis generation, candidate design, synthesis orchestration, and clinical planning [11].

Our work here explores a complementary and more immediately accessible entry point: what can be achieved through iterative natural-language prompting of a single, unmodified commercially available conversational AI, without any custom software infrastructure? This approach trades the scalability of parallel agent dispatch and MCP tooling (demonstrated by [9] and [10]) for zero engineering overhead and immediate reproducibility by any organization with a commercial AI subscription. We deliberately position FALS Therapeutics not as competing with those systems but as demonstrating the lower bound of what is immediately accessible. As a side note, we should emphasize that historically, machine learning tools have always been essential for the work of computational scientists (computational biologists, computational chemists, etc.) in pharmaceutical discovery and development efforts. The new generation of emerging machine learning tools (referred as AI tools) are a natural progression, and their utilization should be paramount to the success of the computational scientist. Nevertheless, we hear great deal of negative attitudes in pharmaceutical organizations while many enthusiastic efforts adapting these tools should also be acknowledged [12]. As with any new technology, it needs to find its use and value in the drug discovery and development workflows and this manuscript is just an attempt to show some possibilities while still cautioning about ensuring correct context is maintained - for example, human-in-the-loop with their expertise, experience, intuition, and innovative mind-set.

We chose amyotrophic lateral sclerosis as the disease focus for three reasons. First, it represents a profound unmet medical need: no disease-modifying therapy exists for sporadic ALS, which accounts for 90-95% of all cases [13, 14]. Second, the disease has a rich, well-characterized genetic and molecular landscape, providing abundant public data for AI-driven target identification [15–17]. Third, the ALS drug development field has a well-documented history of clinical trial failures driven by patient heterogeneity and poor translational models, providing a clear set of lessons an informed AI should internalize [18, 19].

Here we report the construction, operation, and outputs of FALS Therapeutics - a conceptual virtual biotech with a virtual USD 5M seed budget and a mission to develop disease-modifying therapies for sALS. We describe four sequential work-streams executed by AI agents, the agent skills deployed in each, the scientific and commercial outputs generated, and the self-correcting governance mechanisms that emerged without explicit specification. We also conduct a systematic reference verification audit, explicitly flagging seven categories of claims requiring caution, and formally documenting one hallucinated source attribution corrected in this version.

2. Methods

This section describes the AI system and access mode, the comparison framework used to position FALS Therapeutics relative to related systems, the virtual organization architecture, and the agent skills deployed across work-streams.

2.1 AI System and Access Mode

All outputs were generated using Claude Sonnet 4.6 (Anthropic, PBC, San Francisco, CA), accessed via the claude.ai web interface in April 2026. Claude Sonnet (version 4.5) is also the primary scientist agent model used in the Virtual Biotech framework [9] and is deployed by Biomni as the LLM backbone for its generalist agentic architecture [10], enabling direct methodological comparison across three distinct deployment approaches. No external MCP servers, custom APIs, agent SDKs, or fine-tuning were applied. All agent roles were instantiated through natural-language prompting within a single continuous conversation context.

The platform provides Claude with integrated web search, code execution, file creation, and artifact generation capabilities. Code execution was used extensively for document generation (Node.js with the docx library), data visualization (HTML/React interactive dashboards), and scientific figure production. All generated files were delivered as downloadable Word documents (.docx), PowerPoint presentations (.pptx), and interactive HTML artifacts rendered inline in the conversation.

2.2 Comparison with Related AI Drug Discovery Systems

Table 1 positions FALS Therapeutics within the current landscape of AI-driven drug discovery platforms, comparing it against the Virtual Biotech [9], Biomni [10], and the prompt-to-drug framework [11].

Table 1. Positioning of FALS Therapeutics relative to related AI drug discovery systems. Virtual Biotech data from Zhang et al. 2026 [9]; Biomni from Huang et al. 2025 [10]; prompt-to-drug from Zhavoronkov et al. 2026 [11].

2.3 Virtual Organization Architecture

The virtual biotech was initialized through a CEO founding prompt specifying disease mission, seed funding (USD 5M), and the requirement to build a Development Committee. Four Head-level agent roles were instantiated by name and function through subsequent prompts: Head of Commercial, Head of Research, Head of CMC, and Head of Clinical Development (Figure 1). These agents were not separately instantiated model instances; Claude adopted distinct functional personas within the conversation.

The iterative strategy employed a five-step governance loop per work-stream: (1) mandate specification; (2) agent execution and deliverable production; (3) CEO-level gap analysis comparing deliverable against mandate; (4) directed gap-closure prompt; and (5) DC integration of agent outputs. This mirrors the plan-execute-review cycle of agentic AI systems [20, 21] and the Scientific Reviewer agent function in the Zhang et al. system [9] but implemented through structured prompting rather than dedicated agent infrastructure.

2.4 Agent Skills Deployed

Figure 2 illustrates how the four AI agents – Head of Commercial, Head of Research, Head of CMC and Head of Clinical Development – directed by CEO closely mimic the workflow of a drug discovery and development biotech organization. The flow chart is linear but in reality, their work is iterative and many activities are happening in parallel. The outputs from each agent are enumerated in the figure. Worth noting that many activities or even planning can only take place after certain decisions are reached at the different stage gates. In our prompt-to-blueprint study we focused on artifacts which can be indeed created before stage gate decisions are made or can be carried out under certain assumptions and taking some risks of potential redundant work. For example, four modalities were identified and planned for in molecular generation and CMC planning stages assuming that the initial five targets will all be validated and downstream activities will be initiated by DC decisions. Hence, we will describe and focus on the 18 planning documents in the Result section.

3. Results

The four Head-level agents and the CEO produced 18 governance and planning documents across four sequential work-streams, summarized below by agent role.

3.1 Virtual Organization and Governance

FALS Therapeutics was established through a founding CEO prompt specifying disease mission, seed funding, and a Development Committee structure by the author. Four Head-level agents were instantiated sequentially with prompts from the author of the study. Each prompt describing the scope of the AI sub-agents were given as high level – while the author has many years of pharmaceutical research and development experience, many areas of the R&D processes are of cursory familiarity. The CEO AI agent refined the prompts before submitting to the sub-agents and developed a critical emergent feature - not specified in the initial mandate – to identify gaps in the inputs and outputs and direct gap-closure loop activities. After the Head of Research submitted its Stage 1 plan, the CEO conducted a systematic eleven-point gap analysis and identified eight specific missing elements, including the biology mastery document itself, quantitative go/no-go criteria, modality rationale, a candidate naming convention (established as FALS-T[n]-[modality]-[seq], e.g. FALS-T1-SM-001), target engagement biomarker specifications per target, and an IND ownership map across DC members. This pattern was subsequently applied after every agent deliverable.

This gap-identification pattern parallels the Scientific Reviewer agent in the Zhang et al. system [9], which formally critiques scientist agent outputs against three criteria (query alignment, evidence strength, analysis thoroughness) before CSO synthesis. In our implementation, the review function is performed by the CEO agent through structured prompting. The resulting critiques were substantively similar to the Supplementary Note II example in Zhang et al., which ran to 45 pages for a single agent output - suggesting that the core quality-assurance function is achievable without dedicated reviewer agent infrastructure, though with lower systematic rigor.

3.2 Head of Commercial: ALS Landscape Analysis

The Head of Commercial agent characterized unmet medical need, sized the market, and mapped the competitive pipeline.

3.2.1 Unmet Medical Need

The commercial analysis confirmed a profound unmet need in sALS. Sporadic ALS accounts for 90-95% of all cases [13]. With approximately 32,000 prevalent cases in the US and ~5,000 new diagnoses per year [22], and median survival of 3-5 years [23], the disease burden is substantial. Only three drugs are approved for ALS in the US: riluzole (1995) [24], edaravone (2017) [25], and tofersen/Qalsody (2023, restricted to the ~2% of patients with SOD1 mutations) [26]. AMX0035 (Relyvrio), approved under accelerated approval in 2022, was withdrawn in April 2024 following failure of the Phase 3 PHOENIX trial [27, 28].

3.2.2 Market Sizing and Competitive Pipeline

The global ALS therapeutics market was estimated at USD 791-812 million in 2024 across independent market research sources [29–31]. Twelve clinical-stage assets were identified across multiple mechanisms. No Phase 2b or later clinical asset specifically targets TDP-43 nuclear transport, the STING innate immune pathway, or HDAC6 in ALS - the three primary mechanistic white spaces identified for FALS Therapeutics. The competitive landscape interactive dashboard generated by the Head of Commercial is shown in Figure 3 and the PowerPoint presentation is in Supplementary File 1.

3.2.3 Tools Utilized

The Head of Commercial agent deployed Web search (FDA filings, ClinicalTrials.gov, SEC, PubMed), carried out market sizing analysis and competitive pipeline mapping; generated PowerPoint using Node.js (pptxgenjs), and created and interactive HTML dashboard.

3.3 Head of Research: Research Pipeline, Biology Mastery, Tool Stack and Biological Intelligence, Molecular Design Planning, in vitro and in vivo and CRO Strategy

The Head of Research work-stream spanned pipeline design, biology synthesis, tool-stack optimization, target ranking, molecular design, medicinal chemistry, and preclinical assay/CRO planning, detailed in the seven subsections below.

3.3.1 Full Pipeline Development

The Head of Research AI agent first developed a Research Pipeline with 11 stages based on the guidance from the Head of Commercial AI agent’s report according to the prompting from the CEO AI agent. Detailed plans, including in silico and CRO-based in vitro and in vivo work, were developed and presented through an interactive dashboard (Figure 4a) and report.

The CEO AI agent then re-checked the 11-point mandate and identified 8 gaps - subsequently, the Head of Research created a document and interactive dashboard addressing these gaps (Figure 4b).

3.3.2 Biology Mastery

The Head of Research agent created a report based on ALS literature synthesis (PubMed, Elicit) mapping out the biological mechanism underlying the pathophysiology and a drug failure registry (9 ALS trials) (Supplementary File 2).

3.3.3 Open-Source AI Tool Stack

The initial tool stack specified three licensed AI platform categories totaling approximately USD 450,000 per year expense. Following a CEO directive to preserve seed capital for wet chemistry and in vitro assays, a replacement analysis identified fully open-source equivalents at ~USD 3,000-8,000 per year in cloud GPU compute costs only: Boltz-1/Boltz-2 (MIT) replacing AlphaFold3 commercial [32, 33]; REINVENT4 (AstraZeneca, Apache 2.0) replacing Chemistry42 [34]; GNINA 1.3 (Apache 2.0) replacing Glide/Schrödinger [35]; OpenFE/RBFE (MIT) supplementing Boltz-2 for late-stage FEP; and ADMETlab 3.0, ADMET-AI, SwissADME, pkCSM, ProTox-3.0 (all free) for ADMET prediction [36–40] (Figure 5).

Interestingly, the Head of Research AI agent report recommended one human hire - a Senior Computational Chemist - with listing the necessary skills to operate the open-source stack. At ~USD 140K fully loaded, the net saving after that hire is approximately USD 302,000 per year, reallocated entirely to CRO synthesis, iPSC-MN assays, ADME profiling, and in vivo pharmacology (Supplementary File 3).

Figure 5. Open-Source AI Tool Stack — Cost Optimization Dashboard. Tool-by-tool comparison dashboard showing licensed tools replaced by open-source alternatives across 6 categories (Structure prediction, Target ID, Molecular design, ADMET, Docking/FEP, Data management). Bar chart of cost savings per category. Performance gap assessment (SMALL/MODERATE/SIGNIFICANT) for each replacement. Total saving: USD 442K/year software costs; net USD 302K/year after computational chemist hire. Source: Head of Research, FALS Therapeutics; In Silico Tool Cost Optimization Report, April 2026.

3.3.4 Layer 1 Biological Intelligence Sweep - 12-Target Long-List

The biological intelligence sweep integrated seven open-source data platforms: OpenTargets, GWAS Catalog [15], Project MinE, Mendelian randomization literature [41], Answer ALS, STRING v12, and PubMed/Elicit AI. The sweep identified 28 candidate targets, producing a ranked 12-target long-list (Table 2) for DC Gate 1 consideration (full report is in Supplementary File 4). Four targets are proposed for Tier 1 status; TBK1 is the most consequential new addition - not in the initial mandate by the Head of Commercial AI agent - identified through convergent evidence across all four data streams. The proposed targets were then illustrated on the ALS pathophysiology diagram (Supplementary Figure 1).

Table 2. Ranked 12-target long-list for DC Gate 1 submission. NEW★ indicates targets identified by the Layer 1 sweep not present in the initial mandate. Genetic score based on GWAS significance, Mendelian randomization evidence (Duan et al. 2024 [41] for TBK1), and OpenTargets association score.

3.3.5 Molecular Design Planning

The Head of Research planned the modalities to address the five top targets (four Tier 1, and one Tier 2) with full analysis and supporting evidence for each decision (Supplementary File 5). The modality selection decisions are as follows: T1 (TDP-43/CK1δ/ε) - oral small molecule kinase inhibitor with CNS penetration, with intrathecal ASO backup; T2 (TBK1/STING) - oral small molecule kinase inhibitor with >20-fold selectivity over IKKβ; T3 (UNC13A) - intrathecal splice-switching antisense oligonucleotide (2’-MOE/LNA gapmer); T4 (G3BP1/2) - oral PROTAC degrader targeting the G3BP1 NTF2L domain; T5 (HDAC6) - oral selective small molecule hydroxamate inhibitor. Each decision is grounded in structural biology, published tool compound data, ADMET tractability analysis, and regulatory precedent. (Figure 6)

3.3.6 MedChem Campaign Planning

The Head of Research AI agent created a detailed report providing biological rationale for targeting T1 (CK1δ/ε), elaborating the therapeutic window rationale why partial antagonist is needed, analyzed the insights from structural biology (e.g.: PDB 5W4W), established Structure-Activity Relationships criteria and designed a REINVENT4 campaign and code, designed a lead optimization strategy, drafted the IND-enabling package requirements and finally suggested an intrathecal gapmer ASO targeting TARDBP autoregulation as a backup (Supplementary File 6). The interactive dashboard is illustrated in Supplementary File 7.

3.3.7 in vitro and in vivo Assays and CRO Strategy Planning

The Head of Research developed a detailed plan for how the in vitro and in vivo assays should be sourced covering all preclinical stages: Hit ID, Hit-to-lead (HL), Lead optimization (LO), in vivo efficacy, and IND-enabling GLP toxicology. All the assays support the five targets across three modalities: SM kinase inhibitors (T1/T2/T5), IT ASO (T3), and PROTAC degrader (T4). Detailed rationale for the assay types and formats, CRO’s suggested and cost estimates were provided in the report (Supplementary File 8).

3.4 Head of Clinical Development: Clinical Development Master Plan

The Head of Clinical Development produced the Phase 1-3 strategy, the FALS-PEM patient enrichment model, and the T1+T3 combination trial proposal described below.

3.4.1 Phase 1-3 Clinical Strategy

The Head of Clinical Development produced a comprehensive clinical development master plan and an interactive dashboard (Figure 7) covering Phase 1 (SAD/MAD, n≈96, target FPFV Q1 2028), Phase 2a (open-label biomarker-enriched, n=30-40/target, Q2 2029), and Phase 2b/3 (randomized double-blind placebo-controlled, n=60-150/arm depending on enrichment, Q2 2030). The Neurofilament Light (NfL)-anchored accelerated approval pathway was specified as the primary regulatory strategy - following the tofersen precedent where plasma NfL reduction was accepted as a surrogate endpoint reasonably likely to predict clinical benefit. The HEALEY ALS Platform Trial regimen partnership is identified as the highest-priority operational acceleration opportunity, estimated to save 12-18 months and USD 5-8M in Phase 2b costs.

A seven-member Scientific Advisory Board was proposed and justified, covering ALS clinical trial design (Cudkowicz - SAB Chair), endpoints and patient outcomes (Paganoni), genetics and target validation (Traynor), European clinical development and EMA regulatory strategy (Hardiman), translational medicine and biomarkers (Benatar), Phase 3 experience and PHOENIX trial lessons (van den Berg), and patient advocacy and strategic partnerships (Bruijn). The SAB budget is USD 250,000 per year including equity (Supplementary File 9).

3.4.2 FALS-PEM Patient Enrichment Model

We have previously described a patient enrichment model for Parkinson’s disease (PD) using multi-modal data [42] based on the premise that clinical development in highly heterogenous population would be more successful if patients can be stratified by the progression. Hence selection of faster progressing patients would allow faster proof of concept (PoC) read-out and/or require reduced number of patients for PoC. We provided the reference to this paper (PMC9613892) and prompted the CEO to direct the Head of Clinical Development to evaluate this paper and use that as a guidance to develop a patient enrichment strategy for the planned ALS trials. The AI model three times misidentified the paper to a different one [43] and developed different interpretations of the direction and build patient enrichment models. After the third correction a reasonable interpretation of the paper was made (Figure 8a) and the core five-step framework from the Sadaei et al. paper [42] - train on natural history data, classify trajectories, simulate enrichment, add objective biomarkers, validate externally - was adapted for ALS using ALS-specific primary literature. Applying this architecture to sporadic ALS produces the FALS Patient Enrichment Model (FALS-PEM): a multi-modal XGBoost meta-predictor trained to classify sALS patients as short-term progressors (predicted ALSFRS-R slope ≥1.0 point/month over 12 months) or non-progressors (slope <1.0 point/month) from baseline features alone, before randomization. All four feature modalities from Sadaei et al. are translated to ALS-specific equivalents. Also, the detailed recipe for the PRS modeling step (Layer 3) was developed and presented in an interactive dashboard (Figure 8b).

According to the Head of Clinical Development AI agent’s interpretation, Sadaei et al. did not directly compute sample size savings in their paper - their goal is classifier performance (AUC). Therefore, the Head of Clinical Development built the sample size model that translates the FALS-PEM classifier performance into concrete Phase 2b sample size reduction, using established ALS trial design benchmarks (Supplementary Files 10 and 11).

3.4.3 T1+T3 Combination Trial

The Head of Research AI agent during the development of the T1 targeted medicinal chemistry campaign planning document, made an unprompted proposal to combine T1 and T3 targeted assets in a combination trial. It suggested that the mechanisms of the two targets are orthogonal and additive: T1 reduces the rate of cryptic exon inclusion by restoring TDP-43 nuclear function; T3 blocks cryptic exon inclusion at the splice site even if TDP-43 function is only partially restored. In patients where T1 achieves 50% pTDP-43 reduction and 30% TDP-43 N/C restoration, T3 provides insurance for the remaining 70% of UNC13A whose cryptic exon continues to be de-repressed. The combination allows each individual drug to be used at a lower dose - reducing the T1 circadian safety risk and the T3 IT dosing burden simultaneously. This is the scientific logic that makes T1+T3 the FALS flagship clinical program rather than either agent alone. Accordingly, the Head of Clinical Development AI agent drafted detailed plans for a Phase 2a and Phase 2b trials to test this idea (Supplementary File 12 and 13, respectively).

3.5 Head of CMC: Manufacturing and Quality Control Plans

The Head of CMC produced manufacturing and quality control plans for all five target tracks across four modalities (Figure 9): small molecules (Targets 1, 2, 5), ASO intrathecal (Target 3), and PROTAC degrader (Target 4). A phase-appropriate manufacturing philosophy was adopted - investing the minimum necessary to advance each candidate safely through each gate, deferring expensive validation activities until clinical viability is confirmed.

Key CMC decisions include: (1) WuXi STA as primary small molecule CDMO, Quotient Sciences for formulation and Phase 1 supply; (2) Agilent/BioVectra as primary ASO CDMO (Ionis Pharmaceuticals excluded due to ALS competitive conflict - tofersen/Qalsody competitor); (3) MilliporeSigma CTDMO as primary PROTAC CDMO, with HPOTC OEB4-5 containment confirmed; (4) Benchling QMS (21 CFR Part 11 compliant) as the virtual quality management system; (5) a contract QA consultant for Year 1-2 (~USD 100K/year) transitioning to a hired Head of Quality post-Series A.

The consolidated CMC budget from seed to NDA is ~USD 35.5M: Phase 1 CMC USD 7.2M (all five targets); Phase 2 CMC USD 11.7M (ten candidates, two per target); Phase 3 CMC USD 15.8M (three clinical candidates). The seed-stage CMC allocation from the USD 5M founding capital is USD 0.5-0.8M, covering route scouting for T1 and T2, ASO synthesis initiation for T3, and QA consultant engagement. This 30-35% CMC-to-total-development-cost ratio is consistent with published benchmarks for clinical-stage biotechs (Supplementary File 14).

Supplementary Figure 1

Supplementary Figure 1. Claude Sonnet 4.6 generated interactive illustration of ALS pathophysiology and the suggested targets. The following captions are generated by the Head of Research AI agent.

The motor neuron soma occupies the centre. Inside the nucleus, TDP-43 (purple dots) performs its normal function of repressing cryptic exon inclusion across thousands of RNA targets. CK1δ/ε kinase (amber badge, upper left) phosphorylates TDP-43 at Ser409/410 and NLS-adjacent residues, driving export through the nuclear pore complex — this is the upstream event that initiates all downstream pathology. T1 (CK1δ/ε inhibitor, green callout) intercepts here.

The exported, hyperphosphorylated TDP-43 accumulates as red aggregates in the cytoplasm. This single event branches into four simultaneous downstream cascades, each representing one FALS target:

The cGAS–STING–TBK1 cascade (left side) starts when TDP-43 aggregates damage mitochondrial membranes, releasing mtDNA into the cytoplasm. cGAS senses the mtDNA, synthesizes cGAMP, which binds STING on the ER membrane, which recruits TBK1. TBK1 then bifurcates: one arm phosphorylates IRF3, driving IFN-I transcription and activating the microglia shown on the far right; the other arm phosphorylates OPTN/p62, promoting autophagic aggregate clearance. T2 (TBK1 inhibitor) sits precisely at this bifurcation node, targeting both arms simultaneously — the key feature that makes TBK1 uniquely valuable.

The UNC13A cryptic exon branch (right of nucleus) shows the TDP-43 nuclear depletion opening the pre-mRNA to aberrant splicing. The resulting cryptic exon inclusion triggers nonsense-mediated decay, eliminating UNC13A protein and impairing synaptic vesicle release at the NMJ shown at the bottom. T3 (UNC13A ASO, intrathecal) blocks the cryptic splice acceptor site directly on the pre-mRNA.

The G3BP1/2 stress granule (centre cytoplasm, purple dashed ellipse) forms when TDP-43 is captured in the liquid condensate during cellular stress. TDP-43 trapped in persistent stress granules undergoes liquid-to-solid phase transition, seeding the irreversible solid aggregates. T4 (G3BP1/2 PROTAC) depletes the scaffold protein that nucleates this condensate.

Running down from the soma, the axon shows the microtubule track along which kinesin-1 and dynein transport mitochondria, vesicles, and neurotrophic factors. HDAC6 hyperactivity (driven by TDP-43/FUS dysregulation) deacetylates α-tubulin at K40, reducing motor protein binding affinity and causing the stalled kinesin-1 shown. T5 (HDAC6 inhibitor, non-hydroxamate scaffold) restores AcK40 acetylation, confirmed by the EKZ-438 publication showing 35% plasma NfL reduction in vivo.

4. Discussion

The following subsections situate FALS Therapeutics within the AI-drug-discovery design space, summarize its three unprompted scientific contributions, address the hallucination risk inherent to AI-generated scientific content, and outline the study’s limitations.

4.1 The Prompting-Only Approach in Context

FALS Therapeutics demonstrates that expert-guided prompting of an unmodified commercially available conversational AI can generate the nearly complete blueprint for a rare-disease biotech - commercial intelligence, biology mastery, target ranking, tool stack, molecule generation plan, medicinal chemistry strategy, in vitro and in vivo CRO strategy, clinical development strategy, CMC plans, regulatory strategy, patient enrichment model, and a KOL advisory board - in the equivalent of days of working sessions rather than months. The AI agents created numerous artifacts for each stage of drug discovery and development: planning documents, interactive dashboards, figures, tables and code. Since the drug discovery and development is an iterative process and many activities can only be started after certain decisions and stage gates reached, the blueprint presented here is not complete but there is no indication based on our work that planning documents which were out of scope can be also generated by this type of AI agent at comparable quality and high speed potentially increasing the throughput of resource limited biotechs.

This approach occupies a distinct position in the emerging design space for AI-assisted drug discovery. The Virtual Biotech by Zhang et al. demonstrates what is achievable with a custom-engineered multi-agent platform: 37,075 parallel clinical trialist agents processing 55,984 trials in 6 hours [9]. Biomni demonstrates what is achievable with a general-purpose biomedical AI agent trained across 25 domains, capable of autonomously executing tasks from gene prioritization to drug repurposing to microbiome analysis without predefined templates [10]. The prompt-to-drug vision by Zhavoronkov et al. articulates the ultimate convergence: a plain-language prompt initiating a fully autonomous end-to-end drug development program integrating AI with automated laboratory systems [11]. FALS Therapeutics sits at the accessible entry point of this spectrum - requiring no engineering investment, immediately reproducible by any team with a claude.ai subscription, and yet generating substantively useful governance and scientific planning outputs.

The key insight that distinguishes prompting from mere chatbot use is the structured mandate specification. In each work-stream, the agent received a numbered task list specifying deliverables, decision criteria, and output formats. This mandate structure - combined with the gap-identification loop that systematically compared each output against the mandate - is what produces governance-quality results rather than conversational summaries. The prompting strategy, not the AI architecture, is the primary innovation.

4.2 Scientific Contributions: TBK1, FALS-PEM, and Combo Trial

Three substantive scientific contributions emerged from the agentic workflow that were not in the initial mandate. First, TBK1 was identified as a novel Tier 1 ALS target through systematic multi-platform data mining. The MR evidence from Duan et al. 2024 (OR 1.30 [1.19-1.42], p=5.3×10⁻⁷) [41], its second-highest betweenness centrality in the ALS STRING network, its dual mechanism (simultaneously suppressing neuroinflammation via STING-IRF3 and promoting autophagic TDP-43 clearance via OPTN/p62), and the complete absence of any clinical-stage TBK1 inhibitor program in ALS collectively make a compelling case. As a serine/threonine kinase with published crystal structures (PDB: 4IM0, 4IW0) and multiple known tool compounds, TBK1 is among the most tractable targets on the list.

Second, the FALS-PEM patient enrichment model represents a rigorous methodological contribution - adapting the PD enrichment modeling framework from PMC9613892 to ALS using primary ALS literature, with a concrete implementation plan including a 150-patient natural history study, six identified collaborators, and a USD 2.0-2.5M budget with a projected 4-5× ROI from clinical cost savings. The FALS-PEM is estimated to reduce Phase 2b sample sizes by 40-60%, transforming an underpowered expensive study into a precision-enriched signal-detection study. This is arguably the highest-leverage clinical design decision available to FALS Therapeutics at this stage.

Third, the Head of Research AI agent identified a scientific rationale for combining the molecules developed against T1 and T3 targets in a combination Phase 2a and 2b trial without explicit prompting. Subsequently, the Head of Clinical Development AI agent created a detailed clinical trial plan - this would serve as the flagship clinical trial for FALS Therapeutics.

4.3 Hallucination and the Reference Verification Imperative

This manuscript documents one confirmed hallucination of source attribution - the v1 and v2 manuscript incorrectly describing PMC9613892 as an ALS paper by “van Es et al.” twice and once by “Bhidayasiri et al.” and fabricating ALS-specific data attributed to them. This error was identified by the human operator, not by the AI, and is corrected in v3. It demonstrates that LLMs remain susceptible to confident source attribution errors even on well-studied topics, and that a systematic human verification step - the Reference Audit protocol described in Methods - is not optional but mandatory for any AI-generated scientific content intended for regulatory or peer-review use.

The Reference Audit protocol implemented in this work (independently verifying every factual claim and numerical statistic against primary sources; explicitly flagging unverified claims with ⚠ notation; formally documenting corrections between versions) provides a model for responsible AI-generated scientific reporting. Seven categories of AI-generated claims were flagged across v1-v3. The most important recurring flag is specific numerical statistics without verifiable primary sources (the 78% SOD1 mouse trial figure; the USD 835M market point estimate; the AF3 76.4% benchmark value) - a consistent failure mode that domain-expert review must specifically target (Supplementary File 15). Nevertheless, there are probably still hallucinations, source attributions, misidentifications, erroneous results and inferences which are given the comprehensive scope and large number of documents created is not easy to catch. Human expert review is necessary, and one can consider this approach as replacing numerous interns and graduate students in a biotech with AI agents supervised by domain experts in each area of R&D value chain.

4.4 Limitations and Potential Extensions

All outputs are computational and strategic - no experimental data has been generated. The value of the approach lies entirely in automating some in silico analyses and compressing pre-experimental planning; experimental work remains to be executed. The sequential conversation structure means the approach cannot match the throughput of parallelized systems like the Virtual Biotech or Biomni for large-scale data processing tasks. Domain expertise in the human operator is non-substitutable: the mandate structures, decision criteria, and gap-identification prompts all required pharmaceutical industry experience. Finally, the reference verification audit, while systematic, was conducted by the same human operator who specified the mandates - an independent expert audit would provide higher assurance. Important to mention that commercially available conversational AI systems are continually improving and the amount and type of hallucinations present just a few years ago are now rarer. Therefore, one can reasonably expect that the amount of tedious review and corrections will decrease in the near future, and more human efforts can be devoted to pharmaceutical innovation.

To counterbalance the limited expertise of the author and learning from the limited success controlling of hallucinations and mistakes by the gap-identification pattern employed during this study, a virtual Scientific and Technical Advisory and Review Board can be constituted in a parallel LLM system (such as OpenAI, Google, Meta, etc.). This virtual board would then receive all prompts and generated artifacts for critiquing and correcting, helping the experts overseeing this virtual biotech company to limit mistakes and hallucinations.

It is expected that hallucinations and mistakes will continue to be a feature of these commercial, general purpose AI systems based on their stochastic nature – giving (slightly or substantially) different answers for the same prompt. To address this limitation, one can adopt a strategy to instantiate three-to-five CEOs with the same mandate and use the same prompts to run in parallel these agentic artifact generation workflows resulting in multiple similar full blueprints. At the end, or even at some stage gates, another synthesis agent can be prompted to create consensus reports hence attempting to minimize the chances of hallucinated or mistaken information being reported back to human users and reducing the overhead needed to check and correct all analyses and planning documents.

Another important factor also needs to be noted - many data resources and databases are either behind paywalls or accessible through trusted research environments. Thus, agentic AI currently has limited capability accessing them and human experts still need to use their skills to run analyses - many of them mentioned in the planning documents.

And finally, once data starts generated in such biotech organization, it is imperative that they are stored and appropriately governed so current and future AI agents can find, access, interpret and use them - currently best described as FAIR data practice [44].

5. Conclusions

FALS Therapeutics demonstrates that expert-guided iterative prompting of a commercially available conversational AI can produce substantive, governance-ready drug discovery planning outputs for a rare disease with profound unmet need. The four completed work-streams generate a comprehensive pre-IND and clinical development planning package spanning commercial intelligence, multi-platform target identification, open-source tool stack development, molecular generation strategy, medchem design, in vitro and in vivo assay considerations and CRO strategy, clinical strategy and clinical trials plans with patient enrichment modeling, and multi-modality CMC planning - all without bespoke agent infrastructure.

The work positions itself as the accessible entry point in a design space that now extends from prompting-only approaches (FALS Therapeutics) through custom multi-agent platforms (Virtual Biotech, Zhang et al. 2026) [9], general-purpose biomedical AI agents (Biomni, Huang et al. 2025) [10], and towards fully autonomous prompt-to-drug pipelines (Zhavoronkov et al. 2026) [11]. Each point in this design space has distinct trade-offs in engineering investment, scalability, immediate accessibility, and output type. All are needed. FALS Therapeutics shows what one person with domain expertise and a conversational AI subscription can accomplish in days; the other systems show what coordinated engineering investment can achieve at scale.

In a rare disease like sporadic ALS, where patients await effective treatment with urgency and no approved disease-modifying therapy exists, every acceleration matters. The ability to compress months of pre-IND and clinical strategy planning into days - even with the hallucination risks that require multiple domain expert verification - represents a meaningful contribution to the pace of therapeutic development. The FALS-PEM patient enrichment model and the TBK1 identification are the two outputs most likely to have lasting scientific value beyond the demonstration itself.

The full reference verification audit, with flagged and corrected claims, is provided in Supplementary File 15 to model responsible reporting of AI-generated scientific content.

6. Contributions

Sándor Szalma designed and prompted the study, reviewed, edited and finalized the manuscript. Anthropic Claude Sonnet 4.6 was used to execute the prompts, generate the artifacts and draft the manuscript.

7. Acknowledgment

The author wishes to thank Christopher DeBoever (Maze Therapeutics), Jay Tang (Takeda Pharmaceuticals), John Shon (Calico Labs), and Jawahar Jagamoorthy and Fabio Urbina (Zifo Technologies) for the numerous enlightening discussions about the utility and limitations of AI tools.

Data and Supplementary Material Availability

This manuscript accompanies 15 supplementary files (PowerPoint decks, dashboards, and detailed planning reports referenced throughout the Results as Supplementary Files 1-15, plus Supplementary Figure 1 above) that were part of the original submission package. GenRxiv accepts a single Markdown file per submission with no separate file attachments, so the supplementary files are not reproduced here; they are available from the author upon request and are archived alongside the original version of this preprint.

This work is licensed under CC0 1.0 (Public Domain Dedication).